AI agents write the code; kstrl makes sure it actually works. Hand it a feature spec and walk away: it steers Claude Code or Codex with computed codebase context, runs the output through an adversarial gauntlet - mechanical checks, code review, security review - retries with structured feedback, and learns from its mistakes across runs.
An autonomous agentic loop for system-scale Rust to Python migration: file by file, with full type safety and behavioural equivalence testing, learning from its own failures as it goes. Verified on real third-party crates: json-rust (4,710 lines, circular deps, unsafe blocks) translated end to end with 796 tests passing under mypy --strict.
A reference implementation of Recursive Language Models (arXiv:2512.24601): an orchestrator loop that executes model-written Python over corpora far larger than any context window, with Lambda-isolated sandboxing and verifiable span citations.
A pure-Python layout engine for PowerPoint slide composition: flexbox and CSS Grid semantics for decks, resolved down to EMU coordinates. The same input always produces the same slide, which starts to matter a lot once something automated is generating them.
An AI usage coach. Pick one behaviour to improve each week; praxis injects the commitment into every Claude Code and Codex session via hooks, nudges you while you work, and at the end of the week shows you whether you actually did it.
Reusable skill frameworks for AI consulting and GenAI training delivery: creative ideation, consulting playbooks and a five-tier training curriculum, written as modules an agent can load directly.
A Chrome extension that detects and blurs cringe LinkedIn posts as you scroll, using GPT-4.1 analysis with caching and an adjustable sensitivity dial. I built it because my feed had become genuinely hard to read.