Writing

Clear notes about software that works.

Practical writing about AI tools, security, search, deployment, and the decisions behind reliable products.

Start here

April 2026 · 18 min read

RAG in Postgres: Chunking, Embeddings and Hybrid Search with pgvector

A tested reference build of RAG in Postgres: token-aware chunking, HNSW, hybrid search fused with RRF in one query, the filtered-search recall trap, and answers that cite or abstain.

Read note March 2026 · 16 min read

A LangGraph Support Agent with Tool Permissions and Human Handoff

A tested LangGraph support agent: a charter loaded as data, tools that take identity from the server, interrupt() refund approvals, structured handoffs and scripted evals.

Read note March 2026 · 17 min read

Red-Teaming LangGraph Agents: Prompt Injection, Tool Abuse and Regression Tests

A threat model and code-level defences for a LangGraph agent, plus an attack corpus that runs in CI against a scripted model that falls for every attack.

Read note

All writing

April 202616 min read

LangChain vs LangGraph: When Your LLM Feature Needs a State Machine

One refund feature built twice: as an LCEL chain, then as a LangGraph StateGraph with reducers, retries, a Postgres checkpointer and human approval. Tested code and a decision table.

LangGraphLangChainPython

February 202617 min read

RAG, Long Context or Tools? Deciding with Retrieval Evals

How I choose between long context, RAG and SQL/API tools from token budgets, permissions and freshness, then prove the retriever with recall@k, MRR, nDCG and a CI gate.

RAGLLM EvaluationAI Engineering

January 202617 min read

Shipping LLM Features Safely: Shadow Mode, Feature Flags and Eval Gates in CI

A tested FastAPI reference for replacing a working code path with an LLM: shadow mode that cannot touch the response, sticky percentage rollout, a Redis kill switch and an offline eval gate in CI.

LLMOpsFastAPIEvaluation

January 202617 min read

LLM Ticket Triage with Structured Output, Pydantic and FastAPI

A tested reference design for LLM ticket triage: strict Pydantic output, one bounded repair, rules and PII redaction before the model, Redis caching, and an eval gate in pytest.

Structured OutputFastAPILLM Evaluation

December 202517 min read

Reliable LLM Workflows: Idempotency Keys, Retries and Queues with FastAPI and Redis

How I make LLM workflows safe to retry: Idempotency-Key with Redis SET NX, Streams consumer groups, classified tenacity retries, deadlines and a Postgres outbox, all tested.

FastAPIRedisLLM Reliability

November 202517 min read

Keeping a RAG Knowledge Base Fresh: Incremental Ingestion, Permissions and Citations

How I keep a Postgres + pgvector knowledge base correct over time: hash-based re-embedding, tombstones, advisory locks, Row-Level Security, stable citations and model cut-overs.

RAGPostgreSQLpgvectorSecurity

Hiring?

Want this thinking on your team?

I'm open to senior applied AI and Python backend roles, full-time remote or contract.

See hiring facts