Towards Data Science

197 articles in the TruthFoundry index. Each links out to the original.

• Codex Hooks: Inject Custom Logic into the Agentic Loop · Codex hooks allow users to attach custom logic to specific events like tool use or session completion.
• Ten Enterprise RAG Positions That Contradict Standard Tutorials · A manifesto argues enterprise RAG requires structure-first retrieval and expert dictionaries, rejecting standard vector-store tutorials.
• Hand-Written CUDA Runtime Solves LLM Latency on Live Robot Cameras · A new CUDA-based architecture prevents LLMs from failing on live robot cameras by enforcing strict 33ms reasoning deadlines.
• Intel DFlash Speculative Decoding Boosts CPU AI Throughput 4x · Intel's DFlash speculative decoding method increases CPU token generation throughput by nearly 4x using vLLM.
• Developer Builds Python Runtime to Enforce Typed Context for AI Agents · A developer created a Python runtime that enforces strict types on AI context objects to prevent structural errors before prompts are built.
• Survival Analysis Explained: Kaplan-Meier and Cox Models for Recidivism · A beginner's guide to survival analysis using the Kaplan-Meier estimator and Cox proportional hazards model on recidivism data.
• Enterprise RAG Systems Require Relational Tables for Case File Folders · A new RAG architecture uses relational tables to validate case file completeness and detect contradictions across heterogeneous documents.
• AI Coding Agents Fail on Simple Bug Due to Missing Context · AI coding agents consistently failed to fix a trivial HTTP client bug because the fix required undocumented API knowledge.
• Developer Builds Postgres Backend for LangGraph Booking Agent · A developer replaces in-memory storage with PostgreSQL to persist state for a LangGraph AI booking agent.
• Building RAG Systems for Heterogeneous Document Folders · Enterprise RAG systems for folders with unrelated documents use nested outlines instead of traditional indexes.
• Alma Media Fine-Tunes SigLip for Real Estate Image Classification · Alma Media engineers fine-tune Google's SigLip model to automate room-type tagging for real estate listings.
• Running OpenAI Codex as a Headless Agent in Automation Workflows · A Python workflow demonstrates running OpenAI Codex as a non-interactive agent to generate structured research digests.
• Neural Networks Learn Continuous Risk Scores from Categorical Medical Outcomes · Researchers demonstrate a neural network architecture that derives continuous risk scores from discrete medical outcome categories.
• Row-Level Chunking Solves Table Retrieval in Enterprise RAG Systems · New indexing method splits tables into individual rows to improve RAG query accuracy.
• Understanding Star Schema Dimensions in Data Modeling · The article explains regular, junk, and degenerate dimensions within Star Schema data modeling.
• Bayesian Guardrails for AI Decisions: Measuring Uncertainty Before Automating · Experts propose using Bayesian methods to measure AI uncertainty before automating high-risk decisions.
• Benders Decomposition Part II: Handling Infeasible Subproblems with Feasibility Cuts · A tutorial explains using Farkas' lemma to generate feasibility cuts in Benders decomposition for infeasible subproblems.
• Strategies for Aligning Intent with Claude Code to Enhance Programming Efficiency · The article outlines methods to improve alignment between human intent and coding agents like Claude Code, focusing on plan mode, detailed prompts, and iterative verification.
• LLM Judges Exhibit Self-Preference Bias in SQL Query Validation · An LLM judge approved a flawed SQL query due to self-preference bias, highlighting systematic errors in automated validation.
• Three RAG Corpus Shapes and the Cost of Building the Wrong One · Enterprise RAG systems fail at scale when flat vector stores are used on collections with complex document relationships.
• Fine-Tuning Mistral 7B Achieves 98% Accuracy on Medical Reporting · A team fine-tuned a Mistral 7B model using QLoRA to achieve 98% accuracy on complex breast cancer reporting templates.
• Author Retires Query Router for Traversable Knowledge Graph · Developer replaces keyword-based query router with a graph traversal architecture for a synthetic insurance knowledge layer.
• Scaling Enterprise Data Pipelines: Balancing Throughput and Correctness · An engineer details strategies to scale data integration pipelines without compromising data integrity.
• Kimi K3 One Million Token Context Window vs. RAG Experiment · An experiment compares Kimi K3's 1M token context window against RAG for cost, latency, and answer quality.
• Anti-AI Protests Driven by Personal Harm and Billionaire Resentment · Public opposition to AI stems from negative personal experiences, material costs, and resentment toward billionaire investors.
• Building an AI Assistant to Solve Jigsaw Puzzles with Computer Vision · A Python-based AI assistant uses OpenCV to help users locate specific pieces in large jigsaw puzzles.
• Architecture for Secure, Governed AI Agents in Production · A hardened architecture for production AI agents uses pre-filters, ACLs, and HITL workflows to mitigate LLM risks.
• Five Principles for Building Trusted and Verifiable Enterprise Agent Systems · An expert outlines five principles for building enterprise agent systems that earn trust and improve over time.
• Graph Density Experiment Shows Fewer Connections May Yield Better Multi-Agent Performance · A controlled Python experiment reveals that increasing communication pathways in multi-agent systems does not improve performance.
• Current AI Hallucination Detectors Fail to Catch Numeric Errors · Testing reveals standard AI tools miss single-digit numeric errors in RAG systems, making them unreliable for production.