Patric Fitigau · Romania · Available for contract work
Retrieval systems, caption pipelines and compliance tooling in Python. The design constraint on everything I build is the same: an honest gap is cheap, a confident fabrication is expensive. So the tools abstain when the evidence isn't there.
DSP-05.1 Is data at rest encrypted?
Cited
Yes. All customer data at rest is encrypted using AES-256, with keys managed in a dedicated KMS and rotated annually.
↳ Information Security Policy.md — §4.2 Encryption Standards
TVM-07.3 Are third-party penetration tests performed quarterly?
Flagged for review
NOT FOUND IN PROVIDED DOCUMENTS
↳ no supporting evidence retrieved — left blank for a human
The second row is the point. Guessing here would put a false statement in front of a prospective customer.
Measured on 20 CAIQ v4.0.2 questions: 13 usable as drafted, 7 correctly abstained, 0 fabricated.
Turns a vendor security questionnaire into a reviewable first draft from a company's own policy documents. Citation grounding is enforced twice, independently — the second check verifies that every sentence the model cites actually appears in the retrieved evidence, and forces the answer back to "not found" when it doesn't.
Boundary error on untouched words: 4.0 ms, against 370.9 ms for the difflib baseline. Synthetic corpus; real-transcript validation still open.
Recovers caption timing after a human has edited the transcript — surviving both small corrections and whole sentences moved elsewhere — then segments the result into WebVTT cues that obey the standard constraints and never overlap.
Built and verified end-to-end; publication scheduled.
Weekly EU AI Act tracking for compliance and legal teams: RSS ingest, deduplication against the five outlets covering the same event, relevance scoring, summarisation, and HTML issue assembly on n8n with a FastAPI sidecar.
I benchmark before I claim. Every number on this page comes from a harness in the repository that you can run yourself, and the caveats are published next to the results — including the ones that aren't flattering.
When a measurement says a feature makes things worse, it ships disabled. Anchor-align's phonetic matching is off by default for exactly that reason.