Research
Applied research, on-device.
Speech is our product. Vision and ranking are available as applied research engagements.
Vision
Detect + classify on edge
Multi-object tracking
Pixel-level segmentation
Vision-language models
Ranking
Personalized ranking on-device
Vector search in tens of MB
No data leaves the device
Sub-10ms inference
Model optimization is how any of this ships.
Quantization
Distillation
Pruning
Compilation
Technologies we work with 
How we work
1
Scope Hardware target, latency and accuracy targets, data available. 2
Prototype on target A model running on the actual device, not a cloud demo. 3
Optimize and harden Quantize, benchmark, iterate against the target metrics. 4
Handoff Production weights, integration code, benchmarks, monitoring hooks. No lock-in. Engagement model
Tier 1 Diagnostic / Eval 4 to 8 weeks. Fixed fee.
Benchmark report, feasibility verdict, optimization plan, or reference artifact on a single hardware target.
Tier 2 Research Partner Multi-month retainer.
Embedded co-development. Ongoing optimization, new model exploration, co-authored IP.
Every engagement starts with a scoping call. Tier 1 graduates to Tier 2 when the diagnostic uncovers depth that warrants a retainer.
Latest
Recent work
AuraFits: Multimodal AI Advisory Platform for Specialty Retail
Vision-driven product recommendations from 52 live Shopify inventories. One photo, a short conversation, ranked picks from real shelf stock.
Read → Semantic-Aware ASR Evaluation System for Edge DevicesOn-device ASR evaluation at Fortune Global 50 scale. 90% faster evaluation, 4,000× throughput, deployed to smartphones.
Read →