Notes from building AI
Engineering notes grounded in the systems we actually ship: private and on-prem AI, RAG, voice agents, OCR, and the trade-offs in between.
Private by design: shipping production AI that never leaves your building
On-premise AI has a reputation for being slow, clunky and second-best. It isn't anymore. Here is how we run RAG, voice agents and document AI fully on-premise and air-gapped, with the patterns and trade-offs that actually matter.
The real latency of a production voice AI agent
Everyone quotes model benchmarks. Callers experience wall-clock time. Here is the actual per-stage voice AI latency of a live phone agent, and the one trick that makes it feel instant.
Custom OCR for hard scripts: reading Urdu-Nastaliq when off-the-shelf models fail
Some scripts break standard OCR completely. Here is what it took to train a custom Urdu-Nastaliq model that actually reads, and why we open-sourced it.
Let's build your AI advantage.
Book a strategy call and walk away with a clear, technical plan, whether you build custom or start from an accelerator.