Thinking Out Loud

Notes on building AI systems in production.

Building an AI Agent That Refuses to Guess: SEC 10-K Diligence

Most AI assistants will confidently hand you a wrong revenue figure. I built a financial-diligence agent where hallucinated numbers are structurally impossible — every claim grounded in a real SEC filing, or no answer at all.

13 min readMay 30, 2026

LLMsAI ArchitectureSystem Design

Vectorless RAG: Why Retrieval Is a Reasoning Problem, Not a Geometry One

Vector search treats relevance as similarity. On long, structured documents that assumption breaks — and reasoning-based tree search (PageIndex) is a better fit.

8 min readApr 18, 2026

LLMsAI ArchitectureSystem Design

Inside Claude Code: How Anthropic's AI Coding Agent Actually Works

An accidental source map leak revealed the architecture behind Claude Code — an agent loop with tool middleware, memory compaction, and sub-agent orchestration.

12 min readApr 8, 2026

LLMsAI ArchitectureSystem Design

Building a Production-Grade Agentic RAG System

Architecture deep-dive: structure-aware chunking, 3-way hybrid search with Reciprocal Rank Fusion, and a self-correcting LangGraph agent pipeline — all backed by PostgreSQL + pgvector.

14 min readApr 8, 2026

LLMsAI ArchitectureSystem Design

Building a Production Multi-Agent System with LangGraph

Architecture deep-dive: how I orchestrated 10 specialized domain agents with hybrid routing, cross-worker state, and multi-level caching.

14 min readFeb 28, 2026

MLOpsSystem Design

MLOps on Azure: From Experiment to Production

A practical guide to deploying ML models on Azure Container Apps with Bicep IaC, KEDA autoscaling, and LangSmith evaluation pipelines.

8 min readFeb 10, 2026