AI, weighed before it's built.
AI engineering studio · Bengaluru. Cost against speed, big against small, cloud against device. We put both sides on the scale, bring them level, and hand you the result. You own what we build.
What we weigh
- A demo against production: evaluation, guardrails (rules the system can't break) and monitoring.
- A big general model against a small model tuned on your data: your examples, fine-tuning, evaluation. You keep the weights, the numbers the model learned.
- Cost against speed: caching (reusing answers), routing (the cheapest model that can do the job) and batching (grouping requests).
- A data-centre GPU against a device in the field: quantising (storing each weight in fewer bits), fusing kernels (merging steps on the chip), a trimmed model.
Four layers. One studio.
- L4 Products — what people use: assistants, knowledge search, document AI, agents.
- L3 Models — what thinks: evaluation, fine-tuning, models from scratch.
- L2 Systems — what serves it: private deployment, cost and latency, serving surgery.
- L1 Silicon — what runs it, the code closest to the chip: edge inference, GPU kernels.
Made here
- Remotly — shipped, open source (Apache-2.0). Watch and steer terminal coding agents from your phone, directly over your own private network. remotly.dev
- Entwyn — in development, waitlist, Bangalore first. Matchmaking built on a compatibility model we wrote in-house. entwyn.ai