Services
Four things I do, and who each one is for.
All four share one constraint: the system must be able to say it doesn't know, and the evidence for everything it does say must be inspectable.
Document-grounded retrieval
For CTOs and engineering managers who have tried a RAG prototype and can't put it in front of customers.
Ingest, chunking and retrieval over your own corpus, with citation grounding enforced as a separate verification pass rather than a prompt instruction. Answers that can't be traced to retrieved evidence are forced back to "not found" before anyone sees them.
Delivered with an eval harness so you can measure regressions rather than argue about them.
Compliance drafting
For compliance and legal leads at mid-size SaaS who lose days per questionnaire.
CAIQ, SIG and RFP responses drafted from your policy base, each answer carrying the file and section it came from. Anything unsupported arrives flagged and blank, in a queue built for a reviewer rather than a spreadsheet.
The review step is part of the design, not a disclaimer added at the end.
Caption and transcript pipelines
For publishers, e-learning teams and localization vendors under the European Accessibility Act.
Since June 2025 the EAA requires captions on video and e-learning sold into the EU. Human-edited transcripts have no timing; the speech-to-text run that has timing no longer matches the text. I close that gap and emit WebVTT and SRT cues that obey the standard constraints, two lines, 42 characters, one to seven seconds, no overlap.
Runs on CPU at roughly ten times realtime, so the cost model doesn't need a GPU line item.
Monitoring pipelines
For agencies subcontracting AI work, and teams tracking a regulatory beat.
Ingest, deduplication across outlets covering the same event, relevance scoring, summarisation and assembly, on n8n with a FastAPI sidecar for the parts that don't belong in a workflow tool.
White-label friendly. I work directly with your team and stay off the client call if you'd rather.