Case Studies
Selected outcomes
A sample of AI and ML systems taken from problem to production — across enterprise platforms, consumer marketplaces and healthcare. Each one is measured by business results, not vanity metrics. Clients anonymised.
−80%
Distributed, event-driven matching at scale
Split a monolithic pipeline into independent I/O and compute stages — cutting mean queue time ~80% and decorrelating it from data size, with zero loss of accuracy.
Enterprise · Architecture
−75%
Scoring 25 KPIs per call for a quarter of the tokens
Grouped correlated LLM calls into batched, structured requests — cutting average token spend by three-quarters with no loss of measurement quality.
Healthcare · GenAI
~50k / wk
Bringing new users into recommendations sooner
Halved the interaction threshold to pull tens of thousands of new users a week into personalised recommendations — without degrading results for existing users.
Marketplace · Cold start
+1.8%
Cleaner negative sampling for retrieval
Measured and removed false negatives from Two-Tower training, lifting Recall@2400 with cleaner gradients and no added training cost.
Marketplace · Recommendations
Decorrelated
Caching the expensive path
Decoupled ML from business logic and cached precomputed 3D assets — breaking the link between dataset size and runtime, and unlocking offline experimentation.
Enterprise · Feature caching