Two hard problems. Two measured outcomes.
What we were asked to solve, what we built, and the number that came out the other side.
Replacing three unsynchronized data stores with one engine
Challenge
A system needed fast exact lookups, priority-ranked retrieval, and ordered range queries over the same data, under a hard resource budget, without running three separate stores that drift out of sync.
Outcome
The engine built from this research covers all three access patterns from one engine with a fixed, predictable resource ceiling, validated on production-grade cloud hardware.
Results
- 1,407 of 1,414 real benchmark runs completed exactly as intended
- Throughput scaled from 54.8k to 1.83M ops/s across 1–64 concurrent threads
- Overtook common alternative approaches once real contention kicked in
Triaging 47,161 capsule-endoscopy frames under a fixed budget
Challenge
A wireless capsule endoscopy study produces 50,000–80,000 images per patient. The bottleneck wasn't classifier accuracy, it was retaining, ranking, and triaging that volume under real memory and storage limits with a defensible decision record.
Outcome
The triage layer built from this research handles that decision, evaluated on the public Kvasir-Capsule dataset against six alternative approaches.
Results
- Scoring-model choice alone swung the flagged-for-review rate by more than 40 percentage points
- Full durability guarantees measured at 45.8–48.4% latency overhead, reduced substantially once batched
- A common alternative approach showed a 200–500× tail-latency spike invisible at the average