The sciFi blog
Practitioner writing on small language models, regulated ML, and the unglamorous engineering that decides whether a model ever reaches production.
Three 2026 papers converge on the same finding: the hard problem is no longer modelling, it's evidence — reproducibility you can prove, latency and cost you actually measured, and attributions that don't move between runs.
Read postGradient boosting still wins on transaction fraud. The real gains from small language models are in unstructured features, alert triage, and deduplication — everywhere except the authorisation path.
Read postMost credit copilots stall at the compliance review. The fix is architectural: keep the scorecard conventional, ground every generated number in a validated fact, and never let the model recommend.
Read postUnit economics, latency budgets, data residency and reproducibility — the constraints that decide fintech architecture are not the ones frontier benchmarks measure, and that changes which model you should reach for.
Read postWatch the walkthrough, or get a live tour — and our Fintech AI Hackathon is opening for registration soon.