INTELLIGENCEWITH CONTEXT.
Every model here answers to a product: a tray, a lyric, a voice command, a child's first loop. The lab is judged by outcomes, not benchmarks.
Two-stream jewellery vision: people and hands at 10 Hz, native-resolution tray census gated by hand presence. Measured on real 4K shop footage.
CASE STUDY →PyTorch training, evaluation and ONNX export with provenance enforcement in the runtime.
CASE STUDY →faster-whisper word timestamps aligned to Arabic lyrics with normalisation, fuzzy matching and Needleman-Wunsch.
CASE STUDY →A planner-router agent loop with a tool registry, policy gate and approvals; browser, desktop, mail and messaging tools.
CASE STUDY →An assistant that asks rather than answers; the hint ladder is enforced by the service, and the product works with no key at all.
CASE STUDY →A pet-care chatbot on Cloud Functions with OpenAI or Gemini and a PHP fallback.
CASE STUDY →
