The client is Overwatch AI, a startup building AI assistants for major airlines. Their pilot project — the largest European airline (NDA).
In flight, the pilot has an iPad with EFB documentation in hand. Often without internet (aircraft, remote locations). The documentation — flight manuals, checklists, operating procedures — runs hundreds and thousands of pages.
On the ground, at the airport, the picture is different: internet access and large compute budgets, but an even larger document volume and more complex ground-staff workflows.
The task — two distinct perimeters of one product.
Pilot copilot. Fully on-device inference and retrieval. No cloud, no network fallback in flight. Distilled-model selection and tuning under the iPad's hardware constraints. The index lives on the device and updates without a network.
System-evolution pipeline. We left the client not just the system, but the process for evolving it: how to add documents, update the model, evaluate quality, iterate.
Ground operations: agentic system. A full agent doing multi-step reasoning over thousand-page manuals — beyond the classical RAG pipeline of «query → embeddings → answer».
Hiring help. We took part in hiring two in-house AI engineers for the client: helped with role design, technical interviews, onboarding. Goal — the company should be able to evolve the product without us.
Send a brief description of what you have. From there — in correspondence: we get the context, scope the boundaries, and agree on a PoC in days–weeks.
