Writing

Notes from the build

Practical write-ups on backend platforms, AI in production, and scaling real systems.

June 19, 2026

Reliable Tool-Calling Agents with LangGraph and MCP

An agent that calls tools in a while-loop is a demo. Making one reliable takes explicit state, real guardrails, and a clean boundary between the agent and its tools.

Treating LLM Evals Like Tests: RAGAS and Langfuse as a CI Gate

A prompt tweak can silently break an AI feature with no stack trace and no failing test. The fix is to treat evals like tests and run them in CI.

Self-Hosting LLM Inference with vLLM (and When Not To)

Self-hosted inference can cut cost and latency at volume, but it is real infrastructure. Here is how we run vLLM, and when a hosted model is still the right call.

RAG in Production: Chunking, Hybrid Search, and Reranking

A demo RAG stack is a weekend project. A production one earns its keep through chunking, hybrid retrieval, and reranking. Here is how we build it.

Building ViHi: A Video-First Gig Platform

Why we're betting on video and AI to reimagine how people get help and find work, and how we're building it with Go, Flutter, and ML at Neosenth.