Five projects, what each one prevents going wrong.
Five projects, the failure mode each one defends against, the one number a recruiter
needs. The full write-up is one click deeper on every entry.
TermsGuard AI
Reads the small print in a privacy policy and flags the unfair parts.
Failure mode it prevents
A clause being labelled "checked by the model" when no model actually looked at it. The previous build did that, silently.
Cost of being wrong
Telling you your bank's terms are predatory when they are not. A bad warning is worse than no warning, because you stop reading the ones that matter.
Headline number
102 hand-labelled reference clauses (52 fair, 50 unfair). 4,527 chunks of retrieval context. A 12-case regression suite runs against the deployed service and currently passes 12 of 12 live.
Honest limit
The retrieval index is mostly data-governance and GDPR context. About 2% of it cites a GDPR article. The tool surfaces context, not regulation.
Two bots that took over two jobs I used to do by hand every morning at Huawei.
Failure mode it prevents
A complaint being forced through to a team the classifier was unsure about. Below 0.55 confidence the ticket goes to a human instead of being routed.
Cost of being wrong
A misrouted complaint costs the customer days of waiting. A needless escalation costs an agent five minutes. Wildly lopsided, so the gate is set to escalate freely.
Headline number
95.2% accuracy on 42 held-out tickets. Both errors were P_BILL and REFUND confusion, the one boundary the model cannot tell apart.
Honest limit
42 tickets is a thin evaluation. The 0.55 escalation threshold was set by judgement about the cost of a needless escalation, not swept.
A doctor asks in plain English and gets the answer out of the patient's own file.
Failure mode it prevents
A different patient's note being returned because it happens to be worded the same way. The retrieval query carries the patient_id, and the database filters on it before ranking.
Cost of being wrong
A data breach, and possibly a wrong prescription. Too expensive to leave to a threshold, so this system has no dial in it at all. The failure is designed out at the query, not tuned down afterwards.
Headline number
30 adversarial probes designed to leak cross-patient data, zero leaks. Self-hosted on the hospital's own machine. PII stripped pre-embedding.
Honest limit
Designed and tested on synthetic records, not a hospital deployment. A real deployment is integration work, not model work.
One project kept off the public site on purpose.
RL vs Genetic Algorithm for Surge Pricing: tested on 637,976 real Boston ride
records, GA +21.4% over flat pricing, Q-learning +18.2%, hand-written rule
+7.9%. State space kept small enough to read.
Report (PDF) →
What this page is and is not
This is: the 60-second skim. Each entry leads with what the project prevents going wrong, what that mistake costs, the headline number, and the honest limit. The full write-up is one click away.
This is not: the full write-up. The full write-ups have the architecture, the diagrams, the per-section honest limits, and the reproduction instructions. They are linked from each entry.
This is not: an alternative to the homepage. The homepage is the project card grid for someone choosing what to read. This page is for the recruiter who wants to read more than a card but less than a write-up.