Trustworthy output from untrustworthy data. I build systems that catch what's silently wrong — a dead sensor nobody noticed, contradictory records, a garbled scan, a confidently-wrong model — so it surfaces as a flag today, not a six-month catastrophe.
I use LLMs where they earn their place (development, meaning, embeddings, plain-language access to jargon-locked data) and keep them out of decisions and the data flow, which stay deterministic and verifiable. An LLM can help you find the answer; it can't be the answer.
Senior systems engineer, multiple AI cycles of judgment on what survives the hype and where it breaks. Currently turning messy Polish legal PDFs into a correctness-checked graph — the hard part is the verification, not the AI. Previously I built a legislative-notification service that ran autonomously in production for over three years before I retired it. I've written about where AI coding agents actually fail and why verification is the real work: <a href="https://walat.eu" rel="nofollow">https://walat.eu
Best fit when: correctness matters, the data is real-world-messy, and you want someone who names what doesn't work and why.
stakent · · focus · HN ↗
Remote: Yes — remote only (~EU hours or async)
Willing to relocate: No
Technologies: Python, LLMs & embeddings, semantic search, deterministic data pipelines, PDF/document extraction, OCR
Résumé/CV: <a href="https://walat.eu/cv/" rel="nofollow">https://walat.eu/cv/
Email: dariusz@walat.eu
Trustworthy output from untrustworthy data. I build systems that catch what's silently wrong — a dead sensor nobody noticed, contradictory records, a garbled scan, a confidently-wrong model — so it surfaces as a flag today, not a six-month catastrophe.
I use LLMs where they earn their place (development, meaning, embeddings, plain-language access to jargon-locked data) and keep them out of decisions and the data flow, which stay deterministic and verifiable. An LLM can help you find the answer; it can't be the answer.
Senior systems engineer, multiple AI cycles of judgment on what survives the hype and where it breaks. Currently turning messy Polish legal PDFs into a correctness-checked graph — the hard part is the verification, not the AI. Previously I built a legislative-notification service that ran autonomously in production for over three years before I retired it. I've written about where AI coding agents actually fail and why verification is the real work: <a href="https://walat.eu" rel="nofollow">https://walat.eu
Best fit when: correctness matters, the data is real-world-messy, and you want someone who names what doesn't work and why.