I'm Crish.Most of the effort in AI goes into making a model sound right. I work on the other half: finding out whether it actually is, and building the instruments that tell you when it isn't.
For the last two years that has meant , a legal AI that reads UK family-law judgments so solicitors don't have to. When I arrived it was citing the wrong authority with total confidence. The cause turned out to be . Rebuilding retrieval around that took correct-authority answers .
Before that I asked a smaller, more uncomfortable question: when we grade a model's reasoning automatically, does the grade mean anything? And if the judge can be wrong, the natural next question is whether the judge can be fooled — so my newest project : the scheming detector from a published safety eval, scored against mechanical ground truth instead of human labels.
The rest of the toolbox is open source. watches a running retrieval system and says when it has quietly gone bad. asks whether a model can write code a proof checker will actually accept, rather than code that merely reads as correct.
If there's a pattern, it's that I'm interested in the moment a system admits it doesn't know.
Work
Experience
university of leeds · ailes
Owned retrieval and evaluation for a solicitor-facing legal AI: rebuilt hybrid search over 350,000+ chunks of case law, added calibrated abstention and token-level attribution so every conclusion cites its evidence, and distilled 300+ production failures into a fine-tuning dataset locked in by regression tests.
primark · leeds
Python capacity models linking delivery schedules to physical backroom constraints cut intra-day stockouts 15% across a £50k+/week Click & Collect pipeline; a Power BI dashboard let management pre-position staff around predicted bottlenecks.
miraiyantra · pune, india
Early-stage startup. Built the prototype ML pipeline for E-Vaidya, an AI-assisted wheelchair for cardiac patients — ECG/SpO₂ telemetry in, 1D-CNN anomaly classifiers out, Dockerised inference serving the live demo that won the £10,000 Young India prize.
Rooms
Off hours
Away from the terminal: badminton twice a week, a sketchbook that is mostly anime, and a cat who reviews all my work.
Say hi