Problem
Paralegals search the trademark register while a client is on the phone. Results must be fast and trustworthy.
What happened
Early prototypes used broad similarity scoring. Queries routinely crossed two seconds on production-scale data. Ambiguous “smart” ranks made lawyers stop trusting the tool mid-call.
Why it happened
I optimized for demo cleverness — fuzzy matching felt like intelligence — instead of desk-call latency and false-conflict risk.
What changed
Exact-first FTS with aggressive normalization. Controlled expansion later, if ever. Search treated as a trust surface.
What I’d do differently
Define latency and trust budgets before choosing a ranking approach. Instrument abandoned queries early.