What We Build
Physical AI breaks after deployment. Models drift, conditions change, and on the edge there is no cloud to catch it. Edgeble self-correcting runtime keeps deployed AI accurate on-device, without stopping inference. It is in production at Tier-1 manufacturers and hardware-agnostic across edge NPUs.
The Role
Own the model layer of the runtime. How a deployed model degradation is detected, how it is adapted on-device under tight compute and memory budgets, and how you prove an adapted model is better before it ever serves. This is the hard, unsolved part of edge AI: adaptation that is safe enough for production lines. You will take a working system further toward more model families, tighter resource envelopes, and field-grade robustness.
How We Work
Edgeble is agent-native. We build with agentic coding workflows daily on internal platform tooling already set up for it. You direct the agents; your judgment goes on what agents cannot do: correction logic, validation design, and what better means. If you would rather type every line yourself, this role will frustrate you, self-select accordingly.
You
4-6 years of genuine ML engineering depth, training and fine-tuning beyond inference integration, quantization for edge targets, and comfort with resource-constrained deployment. Embedded exposure is a plus. This role should span vision, LLM, foundation, and world models. Concrete evidence you work well agent-augmented matters; we will ask how, specifically.
Why Here
Own a defined layer of a production, patent-pending runtime at the moment it scales. Early-team equity, direct work with a founder with 19 years of experience across silicon, embedded systems, Edge AI, Linux kernel/U-Boot maintenance, and physical systems, in the working style most teams are still debating.
Deep-tech Work India