skip to content
Far World Labs

November Rollup

/

jxnl.co
Author: Jason Liu
Date: 2024/05/11

Key Improvements for RAG Search Systems:

  • Synthetic Data for Baseline Metrics:

    • Generate synthetic questions from text chunks to establish precision and recall baselines.
    • Enables repeatable, low-cost testing and system evaluation.

  • Tracking Average Cosine Distance and Cohere Reranking Score:

    • Log metrics to analyze query performance and identify areas for improvement. Once you have a table of query and scores, you will be able to do data analysis to figure out areas where you are underperforming.
    • Example schema includes request ID, cosine distance, and reranking scores.

  • Using Full-Text Search:

    • Combine BM25 with semantic search for better results.
    • Tools like LanceDB1 enhance user experience.

  • Making Text Chunks Look Like Questions:

    • Pre-format chunks as questions to align embeddings and reduce runtime latency.

  • Including File and Document Metadata:

    • Append metadata (e.g., title, tags, author, path, dates) to text chunks for richer search results.

Notable Resources:

RAG: 25 Types of RAG (Retrieval-Augmented Generation)

linkedin.com
Author: Bhavishya Pandit

standard and corrective RAG speculative and fusion RAG agentic and self RAG adaptive RAG and REFEED Retrieval Feedback realm and RAPTOR-tree-organized-retrieval REVEAL for visual language and REACT REPLUG retrieval plugin and MEMO RAG Attention-based RAG and RETRO AUTO RAG and CORAG EACO-RAG and RULE RAG Conversational and Iterative RAG Context-driven-Tree-structured Retrieval and CRAT Graph RAG
  • Additional Types: Fast Graph RAG, Light RAG, Nano Graph RAG, Contextual RAG, Multi-Source RAG, Hierarchical RAG, Multi-pass RAG, Feedback-based RAG, Adversarial RAG, Hybrid RAG, RePLUG, ATLAS

Lesser-Known Frameworks to Build Your RAG System

linkedin.com
Author: Shantanu Ladhwe

  1. RAGFlow: Focused on document understanding with grounded citations, explainable chunking, and seamless data integration.2 GitHub

  2. Cognita: Scales from prototype to production, offers modular components, no-code UI, and incremental indexing.3 GitHub

  3. STORM by Stanford: Designed for creating Wikipedia-like articles with humans in the loop, ideal for collaborative knowledge curation.4 GitHub

  4. FlashRAG: Python toolkit with benchmarks and RAG algorithms, supporting preprocessing, indexing, and efficient execution.5 GitHub

  5. LLM-App: Builds scalable RAG applications with live indexing, multimodal pipelines, and SQL Q&A.6 GitHub

  6. R2R: Bridges experimentation and deployment with multimodal ingestion, hybrid search, and observability tools.7 GitHub

Comments:

  • Franck Stéphane NDZOMGA: Suggests optimizing RAG pipelines with PostgreSQL and pgvector; drops unnecessary abstractions.8 Blog

  • Michael Trimmel: Appreciates R2R for enabling GraphRAG prototyping without Neo4J.9

  • Christopher Danz: Recommends Microsoft’s GraphRAG for graph-based RAG. GitHub

  • Chakka Guna Sekhar: Highlights a multi-modal RAG chatbot project that integrates visual information for user proofs.10 App | GitHub

Notable Machine Learning Papers

reddit.com

  • ViT Efficiency Analysis:

    • Paper: Reassesses Vision Transformers (ViT) against newer models using the same training setup.
    • Insight: Original ViT remains Pareto optimal in accuracy vs. cost for some metrics, despite numerous alternatives.

  • Mechanistic Interpretability:

    • Paper: Focuses on mechanistic interpretability, predicted to popularize the next significant research trend.
    • Comment: Anthropic’s work draws attention to the internal behavior of transformers.

  • Taylor Series Approximation for ML:

    • Paper: Links MLPs and self-attention to Taylor Series.
    • Insight: Highlights theoretical relations between function approximation in physics/math and ML architectures.

  • Energy Transformers:

    • Paper: Derives transformer-like architectures from modern Hopfield networks.
    • Practical Use: Simplifies theoretical understanding and implementation of transformers for image generation.

  • The AI Scientist:

    • Paper: Proposes a step toward more autonomous AI agents in scientific discovery.

  • Sparse Autoencoder Geometry:

    • Paper: Explores sparse autoencoders’ ability to create geometric structures in high-dimensional space, aligning AI design with cognitive organization.

Controversial or Bold Predictions

  • Return of RNNs:
    • Debated relevance and potential comeback of RNNs due to theoretical properties like infinite recursion.
    • Suggested fusion with transformers for new architectures.

Additional Mentions

  • QLora (NeurIPS 2023):

    • Paper: Quantization of LoRA, highly cited but debated impact.

  • Diffusion Models Optimization:

    • Paper: Nvidia’s work on enhancing diffusion models with EMA and hypersphere-constrained weights.

Emerging Themes

  • Cross-disciplinary applications (physics/math insights into ML).
  • Importance of mechanistic interpretability and sparse representations.
  • Debate on theoretical frameworks vs. practical innovations.

Resources Captured1112131415161718

Why Agents Are Stupid & What We Can Do About It

YouTube | tactiq.io transcript

  • Challenges of AI Agents:

    • Current agents fail at complex, long-running tasks due to compounding errors and lack of world-model reasoning.
    • Memory systems are inadequate; retrieval and reasoning are more critical than data storage.
    • Lack of common sense and abstract reasoning leads to cascading task failures.

  • Categories of Problems:

    • Big Brain: Issues with long-term planning, strategic thinking, and abstraction.
    • Little Brain: Tactical errors in moment-to-moment decision-making.
    • Tool Brain: Limitations in precision and interface capabilities, such as navigation or physical manipulation.

  • Improving Agents:

    • Use reinforcement learning, scalable data, and generalizable algorithms to build better systems.
    • Develop shared memory platforms to distribute knowledge among agent systems.
    • Incorporate modular skills (e.g., hot-swappable capabilities for specific tasks).

  • Future Directions:

    • Focus on building middleware and improving fine-tuning for task-specific intelligence.
    • Advocate for open-source contributions to compete with large-scale labs.
    • Aim for better integration of agent intelligence into real-world workflows.

Recommended Resources

  • Book: A Brief History of Intelligence19 – Explores neurobiology and AI.
  • Dataset/Model: Wave UI Dataset20 – Data for training UI interaction models.
  • Model: MOMO by AI221 – Fine-tuned for GUI and robotics accuracy.

Atomic Agents: A Modular Framework for AI Agents

linkedin.com GitHub Repo
Author: Pietro Bolcato22

  • Overview:

    • Atomic Agents23 is a framework for building AI agents with a focus on simplicity, modularity, predictability, and control.
    • Designed for reliable AI systems and built on top of Instructor with Pydantic for data validation.

  • Key Features:

    1. Modularity: Combine reusable components to build AI applications.
    2. Predictability: Clear input and output schemas enhance reliability.
    3. Extensibility: Easily integrate new components.
    4. Control: Fine-tune individual system parts.

  • Dynamic Context:

    • Uses Context Providers to enhance agent responses with real-time information.
    • Allows chaining of agents and tools by aligning their input and output schemas for reusability.

  • Commentary:

    • Craig Harper: Praises Atomic Agents as a strong alternative to LangChain.

The Impact of Reducing Work in Progress (WIP)

linkedin.com
Author: Allen Holub

  • Core Insight: Limiting WIP to one task maximizes productivity by avoiding multitasking penalties. Each additional task reduces productivity by ~20%.
  • Dependency Reduction: Empower cross-functional teams to handle platform/database changes directly while aligning with others via occasional meetings to ensure consistency.
  • Mathematical Impact: Working on three 5-day tasks simultaneously extends delivery from 5 days per task to 21 days for all tasks.
  • Commentary:
    • Dependency-heavy structures (e.g., platform teams) amplify delays and inefficiencies.
    • Cross-functional teams with clear boundaries prevent entropy (Ben B.).
    • Aligns with “Team Topologies” concepts, focusing on cognitive load and reducing inter-team dependencies (Luiz Armesto).
    • “Slow is smooth. Smooth is fast.” philosophy fosters accuracy and smooth workflows (Craig Imlach).

Additional Resources24

⛅🌱 Cloud Seed: Simplifying Cloud Services with GitLab

gitlab.com

  • Overview:

    • Cloud Seed integrates Google Cloud services directly into GitLab for streamlined cloud adoption and automation.

  • Key Capabilities:

    1. Google Cloud Service Accounts:
      • Generate authentication credentials from the GitLab interface for wide-ranging integrations with Google Cloud.
    2. Deploy to Google Cloud Run:
      • Automate containerized app deployments, including Preview Environments for branch-specific commits.
    3. Provision Google Cloud SQL Databases:
      • One-click provisioning of PostgreSQL, MySQL, and SQL Service instances, with Git environment awareness.

  • Use Cases:

    • Cloud-native app development with seamless deployment.
    • Cloud migration and modernization with automated database provisioning.

  • Positive Business Outcomes by Persona:

    • Developers & Engineers: Continuous deployment and easy database provisioning.
    • Managers: Preview environments for feature validation.
    • Ops Teams: Secure, standardized deployment pipelines.
    • Directors: Improved operational efficiency, reduced compliance risk.

Resources Captured252627

Entrepreneurship Concepts: 10 Concepts Every Entrepreneur Should Learn

linkedin.com
Author: Guillermo Flor

  1. Peter Thiel: Compete by creating something unique. Source
  2. Naval Ravikant: Guide to wealth creation. Source
  3. Asymmetric Risks: Leveraging risks effectively. Source
  4. Startup Cash Management: By Sequoia Capital. Source
  5. Ben Horowitz: Contribution trumps passion. Source
  6. Brian Chesky (Airbnb): Funding requires just one “yes.” Source
  7. Jen-Hsun Huang (Nvidia): Focus on massive action over business plans. Source
  8. Product Market Fit: Understanding its role. Source
  9. Sam Altman: Keys to success. Source
  10. Steve Jobs: Start with customer experience, then innovate backward. Source

Comments

  • Dennis Trautmann: Focus on passion, grit, and customer obsession.
  • Dr. Amr El Fawal: Criticism of Adam Neumann as a poor example.
  • Sharad Agarwal: Differentiates passion (Web2) and obsession (Web3).
  • Georgi Furnadzhiev: Adds Alistair Croll’s “just evil enough” concept.
  • PRADEEP KUMAR GUPTAA: Balancing uniqueness with market validation.

Resources Captured28293031323334353637:

Pdoc: Auto-Generate API Documentation

linkedin.com
Author: Banias Baabe
Date: 1 month ago

  • Introduction to pdoc:

    • A simple library for creating documentation pages aligned with your Python module hierarchy.
    • Minimal configuration; runs without the complexity of setting up tools like Sphinx.

  • Comparison with Alternatives:

    • Positioned as an easier-to-use alternative to Sphinx.
    • Comments mention that mkdocs offers more flexibility in themes and styling.

Comments

  • Muhammad Bilal Khatri: Prefers mkdocs for its flexibility in theming and styling. Questions whether pdoc offers similar features.

Tools and Links Captured3839

Launch of New RAG Course by Elvis S.

linkedin.com

Modules:

  1. RAG Fundamentals:
    • Core components, architecture, and advantages over traditional AI.
  2. Building Naive RAG:
    • Construct a personalized tutor.
  3. RAG Chat Assistant:
    • Build a customer service chatbot with query expansion techniques.
  4. Advanced RAG:
    • Implement advanced techniques like CoT prompting and tool integration.
  5. Agentic RAG:
    • Build an Agent RAG system interacting with tools (calculator, reasoning chain, LLM chain).
  6. Deploy RAG Apps:
    • Create shareable online applications and improve iteratively.
  • Course Link40

Vaughn Vernon Workshop: Ports and Adapters Architecture

YouTube Part 1 | Part 2 | Part 3
Author: Vaughn Vernon

1. Ports and Adapters Overview

  • Simplifies separating business logic (inside) from technical details (outside).
  • Ports define application interactions (e.g., catalog queries), while Adapters handle technological concerns (e.g., HTTP or database queries).
  • Supports both monolithic and microservice architectures, ensuring technology independence for the domain model and a clear separation of concerns.

2. Core Architecture Concepts

  • Driver Side:
    • Handles external requests (e.g., HTTP) via adapters like CatalogController.
    • Translates incoming requests into actions for the domain model or application services.
  • Driven Side:
    • Manages external dependencies like databases.
    • Example: PostgresCatalogRepository implements a CatalogRepository interface.
  • Role-Based Interfaces:
    • Differentiate user types (e.g., customers vs. content managers) to enforce security and role-specific access.
    • Compiler-level enforcement prevents misuse of role-specific methods.

3. Behavior and Enhancements

  • Business Logic:
    • Catalog services manage actions like querying categories/products, adding items to carts, and defining products/categories.
  • Role-Based Behavior:
    • Customers access basic catalog functions.
    • Content managers access advanced features for managing categories and products.
  • Access Control:
    • Implemented through role-based interfaces, adapters, or service layers to handle transactional and security concerns.

4. Dependency Management

  • Lightweight alternative to IoC containers:
    • Constructor injection simplifies dependency handling.
    • Instantiate components in a main method (e.g., PostgresCatalogRepository, CatalogService).
  • Benefits:
    • Removes external dependencies for tests.
    • Simplifies architecture without sacrificing flexibility.

5. Testing and Mocking

  • Custom mocks for port interfaces ensure fast, technology-independent tests.
  • Testing the catalog service and repository independently improves reliability and speed.

6. Naming and Organization

  • Modules and Packages:
    • In Java: com.company.application, com.company.model, com.company.infrastructure.
    • In .NET: Company.Application, Company.Model, Company.Infrastructure.
  • Separation of Concerns:
    • Application layer: Contains core business logic and services.
    • Infrastructure layer: Adapters handling external dependencies like controllers or database access.
    • Domain layer: Represents core entities and aggregates.

7. Benefits

  • Technology independence for the domain model.
  • Reusable patterns for different interaction types.
  • Fast, controlled testing with clear separation between layers.

Additional Resources

A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications

arxiv.org

  • Prompt Engineering Definition: Strategic task-specific instructions guide LLM and VLM outputs without parameter modifications.

  • Core Techniques:

    • Zero-Shot Prompting: Leverages pre-existing knowledge without labeled data.
    • Few-Shot Prompting: Demonstrates tasks with a few examples to enhance understanding.
    • Chain-of-Thought (CoT): Step-by-step reasoning to handle complex logic and problems.
    • Retrieval-Augmented Generation (RAG): Integrates external knowledge for factually enriched prompts.
    • Chain-of-Code (CoC): Uses pseudocode to improve logical and semantic reasoning.
    • Emotion Prompting: Enhances LLMs’ emotional intelligence through targeted stimuli.

  • Advanced Methods:

    • Graph-of-Thought (GoT): Models reasoning processes as directed graphs for complex tasks.
    • System 2 Attention (S2A): Selectively attends to relevant context for enhanced focus.
    • Chain-of-Verification (CoVe): Emulates human-like verification processes to improve factual accuracy.

  • Applications:

    • Reasoning, hallucination reduction, code generation, user interaction, and intent understanding.
    • Notable datasets: GSM8K, TriviaQA, WikiTQ.

Guide to Weisbord’s 6 Box Model

getlucidity.com

  • Purpose:

    • Focuses on business objectives, target markets, and strategic alignment.
    • Questions include: Why does the business exist? What are its vision and mission?

  • Structure:

    • Reviews organizational design, departmental relationships, and operational efficiency.
    • Emphasizes creating structures suited for achieving strategic goals.

  • Relationships:

    • Addresses team dynamics, collaboration, communication, and conflict resolution.
    • Negative relationships can obstruct strategy execution.

  • Rewards:

    • Considers employee incentives, recognition systems, and goal-setting processes.
    • Poor reward systems can demotivate staff.

  • Leadership:

    • Highlights the role of managerial behavior and alignment with company values.
    • Good leadership ensures unified team direction.

  • Helpful Mechanisms:

    • Focuses on tools, systems, and processes that support business execution.
    • Includes regular system reviews and feedback collection.

Advantages

  • Simple and comprehensive framework for internal analysis.
  • Broad application across company or department levels.

Disadvantages

  • Solely inward-looking; lacks external market consideration.
  • Time-consuming to implement thoroughly.

Alternatives

  • VRIO Framework: Focuses on competitive advantage.
  • VMOST Analysis: Links strategy with business activities.

Hierarchy of Traction for VC Pitches

linkedin.com
Author: David Foreman

  • Traction Hierarchy:
    1. Revenue / ARR: Definitive proof of market success.
    2. Contracted ARR: Signed agreements for predictable future revenue.
    3. Paid Proof of Concepts (PoCs): Validates willingness to pay.
    4. Unpaid PoCs: Demonstrates user interest and product validation.
    5. Trials: Early user feedback through active engagement.
    6. Master Service Agreements (MSAs): Framework agreements without guaranteed implementation.
    7. Letters of Intent (LOIs): Soft commitments showing potential interest.
    8. Pipeline Value: Early discussions; lowest credibility for traction.

Insights

  • Focus investor attention at the highest credible traction level.
  • Awards don’t matter unless tied to revenue, market access, or customer acquisition.
  • MSAs are often low-value unless they lead to active adoption.

Comments

  • Scott Newton: Awards can matter if they provide grants, target customers, or unique market access.
  • Jem Stein: Pipeline value often overstated; skepticism warranted.
  • Paul Adams: True traction requires monetary transactions; user metrics without revenue are insufficient.
  • Metin Emenullahi: Questions the practicality of high-ranking unpaid PoCs in early-stage EU startups.
  • Evaldas Girskus: Founders should understand Go-To-Market (GTM) strategies and scale early successes.
  • Pete Stubbs: Consider alternative funding, such as founder-friendly capital online.
  • Chris Parnell: “R for Revenue, not for awaRds. It’s traction, not trickion, that matters.”

How to Miss Your Dates and Fail Your Goals With Finesse

scarletink.com
Author: Dave Anderson
Date: January 20, 2022

  • Key Principles for Managing Goal Failure:

    • Understand the business impact: Assess the consequences of missing a date.
    • Minimize harm: Strategize ways to reduce negative effects.
    • Reinstate predictability: Ensure the team and business can depend on revised deliverables.

  • Avoid Personal Obsession:

    • Focusing on personal reputation may hinder effective decision-making.
    • Example: A manager, panicked about optics, proposed rushing a critical fix, risking major customer impact instead of tolerating a minor delay.

  • Critical Questions When Delays Arise:

    • Why was the original date set? Often arbitrary and outdated.
    • What is the actual business impact of the delay? Typically less severe than perceived.
    • What are the risks of rushing? Hasty fixes often introduce greater long-term harm.

  • Takeaway: Success isn’t about avoiding failure but thoughtfully managing it.

Model2Vec: Efficient Static Embeddings

github.com

  • Overview:

    • Optimizes Sentence Transformers into compact static embeddings, 15x smaller and up to 500x faster.
    • Ideal for NLP tasks like classification, clustering, and Retrieval-Augmented Generation (RAG).

  • Key Features:

    • High Performance: Outperforms static models like GLoVe with minimal trade-offs.
    • Fast Distillation: Distill models in ~30 seconds on a CPU without training data.
    • Lightweight: Minimal dependencies; small models (e.g., potion-base-8M at 30MB).
    • Seamless Integration: Supports HuggingFace, Sentence Transformers, and txtai.

  • Example Code:

    from model2vec import StaticModel
    model = StaticModel.from_pretrained("minishlab/potion-base-8M")
    embeddings = model.encode(["Example sentence 1", "Example sentence 2"])

Kubernetes Troubleshooting Use Cases

linkedin.com Author: Govardhana Miriyala Kannaiah Date: 4 weeks ago

  • Key Troubleshooting Use Cases:

    1. Fixing Kubernetes Node Not Ready: Guide here.
    2. Resolving Kubernetes Node Disk Pressure: Guide here.
    3. Understanding CreateContainerError: Guide here.
    4. Understanding CreateContainerConfigError: Guide here.
    5. Explaining Kubernetes ImagePullBackOff: Guide here.
    6. Explaining Kubernetes RunContainerError: Guide here.
    7. Fixing OOMKilled Issues: Guide here.

  • Newsletter: Bite-sized TechOps examples covering DevSecOps, Cloud, Containerization, IaC, GitOps, and MLOps. Subscribe here.

  • Twitter (X): Follow Govardhana on X for updates here.

Adaptive RAG for Optimized Query Handling

linkedin.com
LangGraph Adaptive RAG Tutorial

  • Adaptive Retrieval: Dynamically selects the optimal strategy for handling queries based on complexity, improving efficiency and cost-effectiveness.
    • Complex Queries: Uses iterative multi-step retrieval for layered, accurate responses.
    • Moderate Queries: Employs single-step retrieval for concise and relevant information.
    • Simple Queries: Skips retrieval entirely, generating direct answers.
  • Efficiency: Minimizes unnecessary compute usage, offering substantial cost savings at scale.
  • Implementation: Built with LangGraph, leveraging LangChain for a modular, easy-to-extend pipeline.

Example Workflow

  1. Query complexity predicted using a lightweight model.
  2. Framework dynamically selects between iterative, single-step, or direct answer retrieval methods.
  3. Adaptive execution ensures cost and compute efficiency.
Adaptive RAG

Resources and Tools

AWS Card Clash: Game-Based Learning for Cloud Architecture

aws.amazon.com

  • Overview:

    • A 3D virtual card game to explore and learn AWS Cloud architecture.
    • Designed for both beginners and experienced users to enhance AWS knowledge.

  • Key Features:

    • Play to Learn: Experiment with AWS Cloud service cards to optimize designs and win through friendly competition.
    • Architectural Insights: Identify and deploy scalable, efficient AWS solution architectures.
    • Game Mechanics: Use strategy cards to defend solutions or disrupt opponents’ deployments.
    • Practical Learning: Explore real-world AWS service applications via interactive 3D gameplay.

  • Benefits:

    • Develop foundational knowledge of AWS Cloud services and interfaces.
    • Enhance understanding of scalable, efficient cloud solution designs.
    • Combine learning with engaging, competitive gameplay.

  • Availability:

    • Accessible through all versions of AWS Skill Builder.

Perform Outlier Detection More Effectively Using Subsets of Features

towardsdatascience.com
Author: W Brett Kennedy
Date: November 24, 2024

  • Challenges of Outlier Detection: Curse of dimensionality affects high-dimensional data; relevant subspaces improve detection accuracy.
  • Subspaces: Identify subsets of features to isolate anomalies; mitigates noise from irrelevant features, enhances interpretability, and reduces computation.
  • Key Algorithms:
    • KNN & LOF: Detect anomalies using local and global neighborhood distances.
    • SOD: Leverages shared nearest neighbors for robust subspace detection.
    • FeatureBagging: Randomly creates feature subsets; ensembles base detectors like LOF.
  • Advantages of Subspaces:
    • Improved interpretability and scalability.
    • Parallel execution and reduced dimensionality effects.
    • Easier model tuning over time.

The Essence of Executive Presence

linkedin.com
Author: Ethan Evans

  • Definition of Executive Presence:

    • 60% gravitas (confidence, decisiveness under pressure).
    • 30% communication (speaking, reading rooms, holding attention).
    • 10% appearance (grooming and professionalism, critical for first impressions).

  • Strategies for Improvement:

    1. Practice public speaking in safe settings (e.g., Toastmasters).
    2. Build emotional intelligence through meditation and reflection.
    3. Seek feedback on room-reading and communication post-meetings.
    4. Use structured influence techniques (e.g., How to Win Friends and Influence People).
    5. Consider professional stylists for appearance refinement.

  • Key Comments:

    • Mala Munisamy: Balancing executive presence with authentic, empathetic leadership is challenging.
    • Maria Shum: Thoughtfulness and insights over loudness; empathy aids in emotional management.
    • Joseph Feduccia: Focus on effective leadership first, leveraging situational awareness and mentoring.
    • Sue Bethanis: Establish common ground to ease conversations and build likability.
    • Sasa Djolic: Highlights risks of “dark patterns” in executive presence conflicting with DEI values.
    • Adrian Neumeyer: Emotional control is central to leadership presence.

  • Resources Mentioned:

    • How to Win Friends and Influence People by Dale Carnegie.
    • Cialdini’s principles of influence.
    • Ethan’s course, Breaking Through To Executive (Maven Platform).

Tools and References414243:

21 Rules for Operating Software at Scale

linkedin.com
Author: Owain Lewis

Owain’s Post:

  1. Know if things break before users do.
  2. When things break, ask: “What changed recently?”
  3. Stale runbooks are dangerous. Update them.
  4. Dashboards should answer: “Is everything OK?”.
  5. Failures are inevitable. Prepare for the worst.
  6. Practice rollbacks before you need to rollback.
  7. Regularly practice disaster recovery.
  8. Backups are non-negotiable. Automate them.
  9. Every page should need some action.
  10. Use error budgets to balance reliability and innovation.
  11. Build a blameless culture that learns from failure.
  12. Make fixing things “one-click” where possible.
  13. Feature flags decouple deployment from release.
  14. Roll out changes progressively.
  15. Testing environments aren’t optional.
  16. Load test systems to uncover weak points.
  17. Simplicity scales. Design for maintainability.
  18. Toil reduction improves morale.
  19. Avoid 24/7 on-call where possible.
  20. Practice breaking things to build resilience.
  21. Have clear SLOs and SLAs.

Comments:

  • Automation and Monitoring: Automate repetitive tasks, establish monitoring systems to detect issues early, and regularly tune performance. (Haroon Mushtaq)
  • Blameless Culture: Continuous learning from failure and avoiding blame is crucial for building trust and process improvement. (Gregor Ojstersek, Elis Wilkins)
  • Rollback Preparedness: Practicing rollbacks has proven invaluable in many cases. (Eitan Yehoshua)
  • Simplicity and Failure Planning: Systems at scale require prioritizing simplicity and preparing for failure to handle complexity. (Mike Leber)
  • Key Takeaways: Automation, resilience, and simplicity stand out as themes among the rules and comments.

AWS Systems Manager Features

aws.amazon.com

  • Secure alternative to bastion servers and direct SSH access

  • Unified Node Management:

    • Manage EC2 instances, hybrid servers, or multicloud nodes via a unified interface.
    • Comprehensive visibility of managed and unmanaged nodes across AWS accounts and Regions.
    • Gain detailed node insights: instance ID, OS details, installed agents, and tags.

  • Patching and Compliance:

    • Automate OS and software patching across cloud and on-premises environments.
    • Use patch baselines to auto-approve updates, override rules, and schedule maintenance windows.
    • Ensure patch compliance and seamless updates.

  • Task Automation:

    • Secure remote management eliminates bastion hosts, SSH, and remote PowerShell.
    • Automate tasks like registry edits, user management, and software installations.

  • Simplified Node Diagnostics:

    • Detect and remediate SSM Agent issues, including misconfigurations and outdated software, using predefined runbooks.

  • Amazon Q Developer:

    • Extend visibility using natural language querying for node data across AWS accounts and Regions.
    • Quickly identify and act on issues with AI-enhanced insights.

Tools and Resources4445

Frustrations with Local AWS Lambda Development

reddit.com

  • Main Issues with Local Lambda Development:

    • hyun88: AWS lacks user-friendly local development tools for Lambda.
    • unpluggedcord: Advocates using make commands to invoke Lambdas directly, simplifying the setup.

  • Alternative Solutions:

    • Redis vs. DDB:
      • Naher93: Suggests switching from Redis to DynamoDB (DDB) to reduce costs (~$20–$30/month as a dealbreaker).
    • Lambda Live Debugger:
      • Background_Bag2770: Promotes Lambda Live Debugger for cost-free, serverless infrastructure debugging.
      • Features:
        • Compatible with various frameworks (CDK, SLS, SAM, Terraform).
        • Minimal setup required (e.g., profile and region configuration).
        • Supports production debugging with Observability mode.

  • Recommendations for Common Triggers:

    • Use AWS’s built-in invoke functionality for triggers like SQS.

Relevant Tool:

Describe Your Ticket Workflow

reddit.com

  • Flexible Approaches:

    • Retrospectives are essential for optimizing workflows (multiple commenters).
    • Experimentation and iteration are key; no single “standard” applies universally.

  • Ownership and Accountability:

    • Epics often created by leadership or Product Owners; tasks by developers for technical details (Charming_Complex_538, iwek7).
    • Developers involved in ticket creation and estimation to improve accountability (Charming_Complex_538).

  • Scope Management:

    • Scope defined through negotiation between PMs (business value), EMs (time-to-market), and ICs (feasibility) (Charming_Complex_538).
    • Adjustments handled collaboratively; PMs usually manage rescoping with input from developers (iwek7).

  • Separate Bug Tracking:

    • Some teams use separate ticket systems for bugs to streamline workflows (Informal-Dot804).

  • Challenges and Anecdotes:

    • Misalignment between business and dev teams leads to last-minute prioritization issues (HowTheStoryEnds).
    • “FrAGILE” methodologies can result in poor developer outcomes when misapplied (Fluid_Frosting_8950).

  • Document retro action items and track progress (Charming_Complex_538).

  • Keep workflows elastic; adapt based on team needs (iwek7).

  • Avoid rigid processes; focus on solving problems practically (discondition).

Saplings: A Framework for Search-Enabled AI Agents

GitHub
Author: Jonathan Shobrook

  • Plug-and-play Search Algorithms: Implements Monte Carlo Tree Search (MCTS), A*, and Greedy Best-First Search for AI agents.
  • Integration with OpenAI APIs: Leverages OpenAI function calling for reasoning and task execution.
  • Modular Design:
    • Agents: MonteCarloAgent, AStarAgent, GreedyAgent, and baseline COTAgent.
    • Tools: Customizable utilities for specific tasks (e.g., arithmetic, web navigation).
    • Evaluator: Guides search with LLM-scored trajectories.
  • Advanced Tooling:
    • Tools access agent memory for dynamic task handling.
    • Output reformatting for model presentation.
    • Terminal tools enforce trajectory finalization.
saplings system diagram

Serverless.tf: Serverless Application Management Using Terraform

serverless.tf

Overview

  • What is it: Opinionated open-source framework for serverless application and infrastructure lifecycle management on AWS using Terraform.
  • Key Value: Simplifies the complexity of serverless workflows by consolidating tools into a Terraform-based ecosystem.
  • Core Benefits:
    1. Unified infrastructure management for all resources, reducing fragmentation.
    2. Reusable, high-quality components using Terraform AWS modules.
    3. Supports faster lifecycle management and accurate AWS service interactions.

How it Differs

  • Single-Tool Approach: Replaces multiple tools (e.g., Serverless Framework, AWS CLI, Makefiles) with Terraform.
  • Reusable Components: Built on years of development with Terraform AWS modules, avoiding custom plugin code in JavaScript.
  • Environment Management: Enables spinning up complete environments for testing/development with real AWS services.

Supported AWS Services

  • Lambda, AppSync, EventBridge, Step Functions, API Gateway, DynamoDB, Aurora Serverless, RDS Proxy, S3, SNS, SQS, CloudFront, CloudWatch, and more.
  • Latest additions: AWS SSM Parameter Store, AWS Secrets Manager.

FAQ Highlights

  • Platform: AWS-only.
  • Runtimes: All supported without restriction.
  • Workflow: CI/CD customizable, managed as code.
  • Support: Commercial support available from Betajob.

Resources

Free Submission to 30 Directories for SaaS or AI Tools

reddit.com
Author: Revolutionary_Hair73

  • Service Overview:

    • Revolutionary_Hair73 offers free manual submission of SaaS/AI tools to 30 directories with at least 50k monthly traffic.
    • The aim is to provide SEO backlinks to boost Domain Rating (DR).
    • Service offered in exchange for testimonials during its early phase.

  • Key Features:

    • 30 high-quality backlinks to improve SEO.
    • Weekly progress reports shared with participants.
    • Designed to save founders time and effort in early-stage marketing.

  • Planned Development:

    • The service is currently manual but may evolve into an automated solution with paid tiers.

  • Critiques and Feedback:

    • fanandrew: Warns that many directories have minimal SEO impact due to high outgoing link counts.
    • inglandation: Questions potential for paid tiers and service scalability.
    • JoaoRochaOnReddit: Inquires about niche-specific directories for unique tools like a real estate buying assistant.

  • Links and Resources:

Comments highlight varying opinions on directory submission effectiveness, with some skepticism about its long-term SEO impact.

The Accidental CIO: A Lean and Agile Playbook for IT Leaders

amazon.com
Author: Scott Millett

  • Core Focus: Balancing innovation and operational stability using lean, agile, and design thinking frameworks.
  • Key Frameworks:
    • Cynefin for decision-making.
    • Wardley Mapping for strategy alignment.
    • Team Topologies for organizational design.
    • Holacratic and hierarchical structures combined for adaptability.
  • Content Breakdown:
    1. New Systems of Work: Philosophies like agile and systems thinking to inspire teams.
    2. Adaptive Operating Models: Governance, talent management, and structural components for dynamic contexts.
    3. Strategy to Execution: From business architecture to IT strategy deployment.
  • Practical Tools: Emphasis on actionable advice, case studies, and clear diagrams for real-world IT leadership.
  • Insights: Explores the CIO’s paradox of fostering innovation while ensuring operational reliability.

Tools and Frameworks4624

The Career-Changing Art of Reading the Docs

pluralsight.com
Author: Forrest Brazeal
Date: June 8, 2023

  • Key Strategy: Read documentation for one job-relevant technology, cover-to-cover, weekly. Focus on technologies you actively use.
  • Approach:
    • Read docs strategically to build a holistic mental model, not tactically for specific problems.
    • Actively question and review understanding; reread regularly to reinforce knowledge.
    • When docs are inadequate, read source code for deeper insights.
  • Common Objections Addressed:
    • Lack of memory: Focus on technologies with immediate relevance and context.
    • Lack of time: Dedicate daily blocks, even 30 minutes, for consistent progress.
    • Poor documentation: Apply common sense, or rely on code for OSS projects.
  • Outcomes:
    • Build rare, authoritative expertise in your technical niche.
    • Enhance troubleshooting by connecting disparate knowledge.
    • Enable career growth through demonstrated mastery.
  • Example: Jared Short reads AWS service docs weekly, gaining insights that resolve complex issues and elevate his reputation.

Critique of Timeboxing in Scrum and Alternatives

linkedin.com
Author: Allen Holub

  • Holub’s Perspective:

    • Timeboxing (e.g., Sprints) creates unnecessary overhead, detracting from actual work.
    • Advocates for small, focused stories without artificial deadlines, emphasizing collaboration and workplace pride.
    • Suggests adding slack time for learning and reflection instead of rigid timelines.
    • Don’t subdivide into tasks. Instead, narrow scope to the point where there are no variables or decisions. Every one of those narrowings is a separate story.

  • Proponents of Timeboxing:

    • Gene G.: Timeboxing fosters discipline, aligns cross-team collaboration, and mitigates estimation risks.
    • Ludovic Urbain: Properly implemented ceremonies like sprint planning/review optimize workflow and reflection without becoming bottlenecks.

  • Criticism of Timeboxing:

    • Igor Zhavrid: Argues for Kanban over Scrum for flexibility, focusing on high-level initiative planning and reducing waste in sprint planning.
    • James Ritzman: Compares software estimates to construction projects—uncertainty and dynamic changes make rigid commitments unrealistic.

Additional Insights

  • Scrum’s Practicality: Debates highlight inefficiencies in large organizations and challenges with scaling across teams.
  • Kanban as an Alternative: Promoted for continuous flow, adaptability, and reduced meeting overhead.
  • Cultural Aspect: Creating a disciplined yet flexible environment is seen as more effective than adhering to rigid frameworks.

Mentioned4748

Dexie.js: A Minimalistic Wrapper for IndexedDB

dexie.org

  • Overview:

    • Lightweight wrapper for IndexedDB (~29k minified and gzipped).
    • Focused on simplicity, reactive capabilities, and ease of learning.

  • Key Features:

    • Reactive Queries: Real-time database mirroring with integration in frameworks like React, Svelte, Vue, and Angular.
    • Concise API: Simplifies IndexedDB operations with straightforward syntax.
    • Syncing Made Simple: Dexie Cloud enables local-first apps with authentication and access control.

  • Framework Examples:

    • React: useLiveQuery for real-time DB queries in components.
    • Svelte: liveQuery for observable DB queries in templates.
    • Vue: Uses liveQuery with useObservable from @vueuse/rxjs.
    • Angular: liveQuery with async pipes for reactive data.

  • Sync Implementation:

    1. Create a cloud database: npx dexie-cloud create
    2. Whitelist app origins: npx dexie-cloud whitelist http://localhost:3000
    3. Install dependencies: npm install dexie@latest dexie-cloud-addon
    4. Configure DB with Dexie Cloud.

The Fantastic Four of System Design (Expanded)

linkedin.com
Author: Alex Xu

Original Fantastic Four

  1. Scalability: Handles increased load without compromising performance.
  2. Availability: Ensures minimal downtime and continuous user access.
  3. Reliability: Delivers consistent and correct results over time.
  4. Performance: Operates efficiently under peak load with available resources.

Additional Pillars Proposed by Commenters

  • Security: Protects against unauthorized access with encryption, authentication (e.g., OAuth), and audits.
  • Maintainability: Eases updates/debugging through modularity, clean code, and documentation.
  • Cost Efficiency: Minimizes costs using strategies like auto-scaling and serverless architecture.
  • Fault Tolerance: Maintains functionality during failures using redundancy, retries, and circuit breakers.
  • Observability: Tracks system health via logging, metrics, and tools like Prometheus or Grafana.
  • Extensibility: Facilitates future features with microservices and API-first designs.

Key Insights

  • Balancing these pillars often requires trade-offs tailored to project needs.
  • Tools like Prometheus, Grafana, and structured logging aid observability and proactive issue resolution.

Understanding RAG Architectures: From Naive to Graph Approaches

linkedin.com
Author: Pavan Belagatti

  • Naive RAG:

    • Basic implementation with document collection, chunking, dense embedding, and vector search.
    • Focus: Simplicity and basic vector similarity.

  • Advanced RAG:

    • Enhances Naive RAG with smart chunking, hybrid embedding, re-ranking, and query expansion.
    • Focus: Semantic chunking and improved relevance.

  • Multi-modal RAG:

    • Handles mixed media types with cross-modal embedding, multi-modal fusion, and modal-specific ranking.
    • Focus: Comprehensive understanding of diverse content formats.

  • Graph RAG:

    • Uses entity extraction, graph construction, embedding, and traversal for relationship-centric knowledge structures.
    • Focus: Complex query reasoning via knowledge graphs.

Resources

RAG Approaches

PGlite: Postgres Anywhere

pglite.dev

  • What is PGlite:

    • Lightweight Postgres in WASM, packaged as a TypeScript/JavaScript client library.
    • Runs in the browser, Node.js, or Bun without dependencies.
    • Supports persistence via IndexedDB (browser) or filesystem (Node/Bun).
    • Compact Size: Under 3MB gzipped.
    • Includes support for Postgres extensions like pgvector.

  • Key Features:

    • Embedded Database: Works as in-memory or persistent DB.
    • Edge-Ready: Sync subsets of server-side Postgres, ideal for AI and RAG workflows.
    • Lightweight & Fast: Perfect for unit/CI testing or local/remote development.
    • Multi-tab Support: Share a PGlite instance across browser tabs.
    • Live Queries: Reactivity for UI updates on DB changes.

The 30 Life-Changing Books Everyone Must Read (According to Naval Ravikant)

productmarketfit.tech
Author: Guillermo Flor
Date: November 26, 2024

  • Naval’s Rule for Reading: Invest freely in books; abandon guilt-free if it doesn’t resonate.

  • Where to Start: “Sapiens” (context on humanity) or “Meditations” (personal guidance).

  • Non-Fiction Recommendations:

    • Sapiens by Yuval Noah Harari: Frameworks changing worldview.
    • Skin in the Game by Nassim Taleb: Real-world risk-taking insights.
    • Poor Charlie’s Almanack by Charlie Munger: Wisdom on rational thinking.

  • Philosophy & Spirituality:

    • Meditations by Marcus Aurelius: Naval’s favorite for stoic living.
    • The Book of Life by Jiddu Krishnamurti: Profound guide to self-awareness.
    • The Tao of Seneca by Seneca: Naval’s go-to audiobook.

  • Science Fiction:

    • Ficciones by Jorge Luis Borges: Stories challenging reality perception.
    • Snow Crash by Neal Stephenson: Visionary look at tech and societal change.

  • Top Blogs:

    • Melting Asphalt (Kevin Simler): Philosophy and psychology.
    • Farnam Street (Shane Parrish): Mental models and decision-making.
    • Stratechery (Ben Thompson): Tech and business strategy.

  • Additional Resources:

    • “How to Get to Product Market Fit” guide.
    • 50+ Pitch Decks from successful startups.
    • Data Room Template VCs Love49.

Strategies for Managing AWS Costs

linkedin.com
Author: Sandip Das

Key Discussion Points

  • Identify High Costs: Use AWS Cost Explorer to pinpoint services driving costs—e.g., Compute, Database, and Data Transfer, especially inter-AZ transfer fees.
  • Cost Optimization Suggestions:
    • Adopt Savings Plans/Reserved Instances for consistent workloads.
    • Replace old instance types with next-gen, cost-efficient alternatives (e.g., t2 → t3).
    • Use Spot Instances for non-critical workloads to reduce costs.
    • Implement Private Endpoints to minimize data transfer fees.
  • Data Transfer Costs: Consolidate services in a single AZ, leverage VPC endpoints, or explore external storage services for cost efficiency.
  • Backup Management: Optimize storage policies (e.g., S3 Intelligent-Tiering) to reduce long-term costs.
  • Additional Tools:
    • AWS Auto Scaling: Adjust instance counts dynamically based on demand.
    • Cost Allocation Tags: Attribute costs by team or project for targeted reductions.
    • Cost Anomaly Detection: Monitor and address budget overruns.

Comments and Suggestions from Experts

  • Nick Jordhan: Focus on instance rightsizing, data transfer optimization (e.g., AWS Direct Connect), and automated scaling.
  • Vishal Bulbule: Break down costs by region and tag; enable Savings Plan recommendations via AWS Cost Explorer.
  • Manoj K: Optimize AZ synchronization frequency for DB clusters to reduce inter-AZ transfer fees.
  • Soumyadip Chatterjee: Replace batch instances with Lambda, use Snowball Edge for large data transfers, and enable RDS performance insights.

Notable Quotes

  • “Treat questions as the real answers.” – Sandip Das
  • “Outbound data transfers are often the most expensive yet hardest to optimize.” – Manoj K
  • “Cost optimization is complicated but achievable with expertise.” – Mohammed Aseem Akram

Automated-AI-Web-Researcher-Ollama: Transforming LLMs into Automated Researchers

github.com

Overview

  • Purpose: Automates research by breaking queries into focus areas, conducting web searches, scraping relevant content, and compiling findings into a structured document.
  • Core Workflow:
    1. Query broken into prioritized focus areas.
    2. Web searches and scrapes content.
    3. Documents results, including source URLs.
    4. Self-improving search cycles for deeper insights.
    5. Generates comprehensive summaries and enables Q&A.

Key Features

  • Automated research planning with prioritized focus areas.
  • Systematic web search, scraping, and documentation.
  • Research summaries and post-research Q&A.
  • Self-improving search mechanisms.
  • Configurable LLM settings (Ollama-based).

ZenML LLMOps Database Overview

zenml.io

  • Purpose: A curated database of 372 real-world LLMOps implementations.

    • Provides detailed summaries and technical notes for various use cases.
    • Filters available by tags like amazon_aws, prompt_engineering, rag, scalability, and more.

  • Highlighted Use Cases:

    • Multilingual Document Processing: Human-in-the-loop pipelines for structured outputs, regulatory compliance, and translation workflows.
    • HR/Payroll Generative AI Platform (ADP): Focus on fine-tuning, scalability, and compliance in high-stakes environments.
    • RAG Systems Optimization (AWS GenAIIC): Lessons on chunking, semantic search, and scaling production-grade RAG architectures.
    • Manufacturing AI Implementation (Accenture): Multi-use case study covering IoT, orchestration, and fallback strategies in compliance-heavy industries.

  • Technical Tags:

    • Covers tools, techniques, and concerns such as vector_search, fine_tuning, multi_modality, monitoring, and legacy_system_integration.

Resources Captured50:

Node.js v22.12.0 LTS Release

nodejs.org
Author: Ruy Adorno
Date: December 3, 2024

  • Key Feature: Enables require(esm) by default.
    • Simplifies compatibility between CommonJS (CJS) and ECMAScript Modules (ESM).
  • Impact on Ecosystem:
    • KaBanks: Concerned this may slow the transition to ESM as CJS now “just works” with ESM.
    • Ryan Shaul: Views this as a major step toward ESM-only libraries, reducing the need for dual publishing.
    • Vincent: Highlights reduced overhead for library maintainers, easing the shift toward ESM.

Relevant Links:

Incident.io: Unified Incident Management

incident.io

  • Integrated Incident Management:
    • Combine on-call scheduling, incident response, and status pages in one platform.
    • Automated workflows in Slack and Teams ensure consistent responses.
  • Alert Integration:
    • Centralize alerts from multiple sources for seamless incident creation and escalation.
  • Customer Communication:
    • Real-time updates through public, private, and internal status pages.
  • Team Coordination:
    • Scheduling, escalation paths, and mobile app to ensure 24/7 readiness.
  • Post-Incident Insights:
    • AI-powered insights and analytics to prevent incident recurrence.

Predicting the Future of Distributed Systems

blog.colinbreck.com
Author: Colin Breck
Date: August 25, 2024

  • Object Storage Trends:

    • Becoming the backbone for transactional and analytical systems due to reliability, flexibility, and two-way-door decisions.
    • Expanding features: cross-region replication, immutability, tiered storage, and interoperability (e.g., S3-compatible APIs).
    • Tools like Parquet, Delta Lake, and DuckDB facilitate seamless data integration and processing.

  • Programming Models:

    • Emerging platforms (Kalix, Dapr, Temporal, wasmCloud, etc.) abstract distributed system complexities (state, workflows, and failures).
    • WebAssembly offers portability and security, potentially enabling migration paths and multi-environment workloads.
    • Long-term trends: infrastructure handles more auxiliary logic (e.g., HTTP servers, logging), leaving “business logic” as portable and secure.

  • Decision Framework:

    • One-Way-Door Decisions: Final, high-risk technology choices (e.g., programming models).
    • Two-Way-Door Decisions: Flexible investments (e.g., object storage) promote iterative innovation.

  • Macro Observations:

    • Distributed systems will prioritize modularity, separation of concerns, and ease of scaling.
    • Early adopters may gain competitive advantage; broader industry adoption depends on clarity and maturity of new paradigms.

Resources515253545556:

Kubernetes Interview Questions and Resources

linkedin.com
Author: Ranjan Yadav

  1. What is Kubernetes, and why is it used?
  2. Explain the components of the Kubernetes master and worker nodes.
  3. What are Pods in Kubernetes?
  4. What is the difference between a Pod and a Node?
  5. What is a Namespace, and why is it used?
  6. How does Kubernetes handle Pod failures?
  7. What are the different types of workloads in Kubernetes?
  8. How does a Deployment work?
  9. What is a StatefulSet, and when would you use it?
  10. Explain DaemonSets and provide a use case for them.
  11. What is the difference between ReplicationController, ReplicaSet, and Deployment?
  12. What is a Service in Kubernetes, and what are the different types?
  13. How does Kubernetes manage internal and external networking?
  14. What is a ClusterIP, NodePort, and LoadBalancer Service?
  15. Explain Ingress in Kubernetes and its use case.
  16. How does DNS work within a Kubernetes cluster?
  17. What are Persistent Volumes (PVs) and Persistent Volume Claims (PVCs)?
  18. Explain the process of binding a PV to a PVC.
  19. What is a StorageClass, and how does it relate to dynamic provisioning?
  20. How are ConfigMaps used in Kubernetes?
  21. What is the difference between ConfigMaps and Secrets?
  22. How can you securely manage secrets in Kubernetes?
  23. How do you scale Pods in Kubernetes?
  24. What is the Horizontal Pod Autoscaler (HPA), and how does it work?
  25. How does Cluster Autoscaler differ from HPA?
  26. What is Role-Based Access Control (RBAC) in Kubernetes?
  27. Explain the difference between Role and ClusterRole.
  28. What is a ServiceAccount, and why would you use it?
  29. How can you implement network policies for Pod communication?
  30. How do you monitor Kubernetes clusters?
  31. What tools would you use to monitor and log Kubernetes applications?
  32. What is node affinity?
  33. What is StaticPods?
  34. What is Taint and Toleration?
  35. What is Sidecar Containers and their purpose?
  36. How to debug specific container logs? Consider there are 2 containers running inside a single pod?
  37. What are the different deployment strategies in Kubernetes?
  38. Describe the process of rolling update in Kubernetes?

Building Bluesky: Real-World Engineering Challenges

pragmaticengineer.com
Author: Gergely Orosz and Elin Nilsson
Date: April 23, 2024

  • Decentralized Social Network: Bluesky enables user-run servers and full data ownership, distinguishing it from centralized platforms.
  • Development Milestones:
    • Built by 2 engineers initially; expanded to 15, supporting 25M users.
    • Transitioned from experimentation (Postgres, AWS) to scalable federation (ScyllaDB, SQLite, on-prem infra).
  • Architectural Evolution:
    • v1: Monolithic Postgres + AWS; v2: Modularized, federated architecture with services like Appview and Ozone.
    • SQLite for PDS; ScyllaDB for read-heavy services. Federation tested via sandboxes and internal servers.
  • Challenges and Lessons:
    • Frequent scaling bottlenecks (e.g., Postgres locks); resolved with ScyllaDB.
    • Rapid feature builds driven by user feedback (e.g., blocking in one day).
    • Transitioned to cost-efficient on-prem hardware for scalability.
  • Cultural Insights:
    • Open-source ethos; active GitHub and real-time user support via the app.
    • Federation shifts moderation and algorithmic control to users.

Resources and Tools575859:

27 People to Follow for MLOps Insights

linkedin.com
Author: Raphaël Hoogvliets

MLOps architects & executives

  • Alejandro Saucedo - OG in MLOps content, amazing newsletter
  • Andy McMahon - book author, community organiser
  • Aurimas Griciūnas - top notch architectures, system design, posts, blogs

MLOps tech leads

  • Eric Riddoch - one of the best I know, shares thought processes
  • Maria Vechtomova - true expert, Databricks Beacon
  • Matthew Sharp - always looking to question important stuff
  • Maxime Labonne - really knows his stuff, wrote multiple books
  • Médéric HURIER - insanely great code + docs
  • Nemanja Radojkovic - very knowledgeable, has dad jokes

ML engineers

  • Başak Tuğçe Eskili - amazing skills, listen up when she speaks!
  • Chris B. - doing many interesting projects
  • Lisa Becker - multi talented course maker, teacher, and specialist
  • Luca Baggi - has a lot of great takes
  • Pau Labarta Bajo - MLOps posts, videos, and blogs for learners
  • Tales M. - MLOps posts in English and Portugese

LLMOps specialists

  • Abi Aryan ☯︎𓁿 - pioneering this subject, book author
  • Paul Iusztin - GOAT of LLMOps & MLOps content
  • Rafael V. Pierre, MSc. - working with GenAI, LLM, and MLOps

Model monitoring specialists

  • Hakim Elakhrass - hilarious and insightful content
  • Wojtek Kuberski - CTO at NannyML
  • Wiljan Cools - post-deployment data science specialist
  • Santiago Viquez - wrote The Little Book of ML Metrics

Community gurus

  • Alexey Grigorev - Founder and organiser DataTalksClub
  • David Scharbach - Founder and organiser MLOps World conference
  • Demetrios Brinkmann - Chief Happiness Engineer MLOps Community
  • Meri Nova - Founder and organiser Break Into Data

Product people

  • Eduardo Bonet - Product Manager MLOps @ GitLab
  • Chris “The Wiz” Alexiuk - Developer Advocate @ NVIDIA

Footnotes

  1. LanceDB ↩

  2. RAGFlow GitHub Repository – Workflow and citation-focused RAG framework. ↩

  3. Cognita GitHub Repository – Modular RAG with no-code UI and incremental indexing. ↩

  4. STORM by Stanford GitHub Repository – Framework for creating Wikipedia-like articles. ↩

  5. FlashRAG GitHub Repository – Python toolkit for RAG benchmarking. ↩

  6. LLM-App GitHub Repository – Scalable RAG application framework. ↩

  7. R2R GitHub Repository – Experimentation-to-deployment RAG framework. ↩

  8. PostgreSQL and pgvector for RAG – Simplified RAG pipeline guide. ↩

  9. GraphRAG by Microsoft GitHub Repository – Graph-based RAG system. ↩

  10. Multi-Modal RAG Chatbot GitHub Repository – Multi-modal RAG architecture for visual Q&A. ↩

  11. ViT Efficiency Analysis Paper ↩

  12. Taylor Series Approximation Paper ↩

  13. Scaling Monosemanticity by Anthropic ↩

  14. Energy Transformers ↩

  15. The AI Scientist ↩

  16. Sparse Autoencoder Geometry ↩

  17. QLora (NeurIPS 2023) ↩

  18. Diffusion Models Optimization ↩

  19. A Brief History of Intelligence – Neurobiology and AI. ↩

  20. Wave UI Dataset – Data for GUI interactions. ↩

  21. MOMO by AI2 – Advanced fine-tuned model for GUIs and robotics. ↩

  22. AI Pulse: Curated daily updates on AI by Pietro Bolcato ↩

  23. https://github.com/BrainBlend-AI/atomic-agents ↩

  24. Team Topologies – Framework for reducing dependencies and managing team cognitive load. ↩ ↩2

  25. Cloud Seed GitLab Page – GitLab’s handbook entry for Cloud Seed. ↩

  26. Cloud Seed Documentation – Official Cloud Seed documentation. ↩

  27. Trusted Testers Signup – Join the beta program and provide feedback. ↩

  28. Peter Thiel’s Zero to One ↩

  29. How to Get Rich by Naval Ravikant ↩

  30. Asymmetric Risks Framework ↩

  31. Sequoia Capital on Startup Cash Management ↩

  32. Ben Horowitz on Contribution ↩

  33. Brian Chesky on Funding ↩

  34. Jen-Hsun Huang’s Advice ↩

  35. Finding Product Market Fit ↩

  36. Sam Altman on Success ↩

  37. Steve Jobs on Customer Experience ↩

  38. pdoc – A lightweight Python documentation generator. ↩

  39. mkdocs – A static site generator for project documentation, known for flexible styling and themes. ↩

  40. RAG Course by Elvis S. ↩

  41. Toastmasters International – Public speaking and leadership training. ↩

  42. How to Win Friends and Influence People – Dale Carnegie’s guide to influence and communication. ↩

  43. Breaking Through To Executive Course – Ethan Evans’ executive presence program. ↩

  44. AWS Systems Manager Features – Official AWS documentation for Systems Manager. ↩

  45. Amazon Q Developer – Natural language querying tool for AWS Systems Manager. ↩

  46. Wardley Mapping – Strategic mapping tool. ↩

  47. Should Teams Use Kanban Instead of Scrum? – Exploration of Kanban vs. Scrum benefits. ↩

  48. Can Kanban Be Used with LeSS or Scrum? – Discusses integration of Kanban with large-scale Scrum. ↩

  49. Premium PMF Resources – Access guides, templates, and pitch decks. ↩

  50. ZenML LLMOps Database – Curated knowledge base for LLM operations. ↩

  51. Kalix – Platform for distributed, stateful application development. ↩

  52. DuckDB – Embedded analytical database. ↩

  53. WebAssembly – Secure, portable execution environment. ↩

  54. MinIO – S3-compatible object storage. ↩

  55. Apache DataFusion – Query engine leveraging Arrow for fast, distributed processing. ↩

  56. Delta Lake – Open table format for analytics and data lakes. ↩

  57. Full Article on Pragmatic Engineer ↩

  58. Authenticated Transfer Protocol Overview ↩

  59. Bluesky Architecture Paper ↩