Archives

What the model may do, and what it may say. Structured tool errors you raise instead of return, at-most-once execution for actions you cannot undo, and why the guardrail that works is demarcation rather than detection.

A typed model call and the flow that routes it, taught together because they compose. Why the agent's output type is also your routing table, how a node declares its own legal moves, and what replaces the if/else on an intent classifier.

Three services in two languages, and nothing hand-written at the seams. How OpenAPI and AsyncAPI become generated clients, why a Python docstring ends up in your TypeScript, and the form nobody wrote that stops a reviewer approving arguments they were never shown.

Why the agent graph and the business logic belong in dedicated services, how to draw the line at state ownership rather than at layers, and the bug that taught me the thread a person reads and the thread a model remembers are not the same object.

The stack for building a typed, observable, controllable agent graph: Pydantic AI, LangGraph, chanx, OpenTelemetry, pytest and evals. What each piece buys you, what it costs, the architecture we build across the series, and the full series map.

Years of shipping LLM features as intent detection and nested conditionals, the state machine idea that would not go away, and the library I started building and abandoned before the practice had a name. The story behind this series, and why graphs outlive the frameworks that implement them.

Ten framework and server configs benchmarked on real PostgreSQL workloads and JWT httpOnly cookie authentication. The 37x JSON gap collapses to 1.8x on authenticated queries, auth itself costs about 10%, and tail latency reorders the leaderboard completely.

Building a production-style AI agent server with Pydantic AI + FastAPI + chanx: streaming structured output and human-in-the-loop tool approvals over a typed WebSocket contract, background runs broadcast to conversation groups (multi-tab, refresh-proof, approve from any device), AsyncAPI-generated TypeScript clients, and offline tests without an LLM.

An in-depth take on DRF API design: when to reach for ModelViewSet, GenericViewSet with mixins, the concrete generic views, or plain APIView, and why the serializer is what makes the choice pay off.

How my team ships DRF + React features without ever hitting a mistyped param or outdated API call: drf-spectacular as the source of truth, generated zod clients, auto-generated forms, and strict typing on both sides.

A maintained, improved successor to nplusone — no more false positives on nullable foreign keys, stack traces in every detection, batch reporting, and full strict typing.

A CLI that copies source code from GitHub repos, URLs, and ZIP archives directly into your project — tracked, updatable, and fully yours. A Rust binary installable via pip.

Database-agnostic localized model fields built on Django's JSONField — with DRF serializers, strict type checking, and a clean admin UI. An alternative to django-localized-fields and django-modeltranslation.

I benchmarked the major Python web frameworks against real PostgreSQL workloads. The 20x performance gap you see in JSON benchmarks collapses to 1.3x once the database gets involved.

Modern DRF authentication with JWT cookies, 50+ social login providers, MFA, full type safety, and OpenAPI schemas that are actually correct — no manual fixes required.

A batteries-included WebSocket framework: automatic message routing with Pydantic, mypy/pyright type safety, auto-generated AsyncAPI docs, and testing utilities — one codebase for Django Channels and FastAPI.

Bringing Django Channels' battle-tested consumer patterns and channel layers to FastAPI, Starlette, and any ASGI framework — group messaging, background workers, and cross-process communication included.