I keep AI agent systems reliable after they ship. Public proof: Blackreach — browser agent with resumable state and outcome verification independent of the agent's self-report, validated at an exact public commit with 3,055 passing tests and a controlled real-browser demo (https://phnix.dev/projects/blackreach). Also Lethe (benchmark for agent drift over extended runs) and Huginn (self-hosted data API, 800+ tests).
Open to full-time remote (or US with relocation) and contract work — I run scoped reliability diagnostics and retainers for teams with browser agents or tool-using workflows in production: https://phnix.dev/reliability
Willing to relocate: Yes (Toronto-based, remote and open to relocate)
Technologies: Python, LangGraph, VLM/LLM integration, autonomous agents, browser automation, multi-agent orchestration, RAG pipelines, edge inference, scraping/crawling
Résumé/CV: phnix.dev
Email: [email protected]
Having Vulkan as the single surface for both the compute and the rendering side means one memory model, one synchronization story, one set of bugs to chase. That alone is worth the effort even before you get to the performance wins.
The dominant frame is "Anthropic hero, Pentagon villain." I think the situation is more complicated, and three specific things are being underreported. 1. The statute has "adversary" in the legal definition. 10 U.S.C. § 3252 defines "supply chain risk" as the risk that "an adversary may sabotage, maliciously introduce unwanted function, or otherwise subvert" a national security system. That...
I build a lot of small AI tools locally, mostly on top of Ollama, and the thing I keep running into is that every session starts from zero. Whatever context I built up the night before, whatever the model learned about how I like things structured, whatever half-finished reasoning we were working through together, it is just gone when I open a new terminal.For a while I was just manually...