Founder
Jomoq AI Labs
December 2025 – Present- Type
- founder
- Building agent harnesses to make frontier models up to 40× faster and 20× cheaper, while developing a commercial suite featuring an in-browser site builder, emotionally intelligent chain-of-thought systems, and a unified agent-publishing studio.
- The product is three surfaces on one backend: a chat client, a builder that edits a live HTML document inside a preview frame, and a studio for publishing agents to their own slug or domain, plus the admin console and billing behind them. I designed, built, and operate all of it alone.
- Agent loop Tool calls run through a server-side loop against a virtual, in-memory filesystem rather than the real disk, dispatching across seven tools (read, write, diff, search, reapply, fetch) and streaming tokens back over SSE. The loop is bounded on three axes: a hard iteration cap, a wall-clock runtime budget, and a per-call model timeout. So a model stuck re-reading or re-editing the same file can't hold a request open indefinitely.
- Providers Six chat providers (Groq, Cerebras, DeepSeek, Gemini, OpenRouter, NVIDIA NIM) sit behind one interface, with separate providers for fast-apply, embeddings, rerank, and web search. Requests bill against the caller's own key first; the server only falls back to house credentials once the caller's session is verified, so an unauthenticated request can't spend house credit.
- Billing Usage caps are enforced with a single row-locked database upsert: the cap check and the charge happen in one statement, rather than a separate read, then check, then write, which is the version of this with an actual race in it. Each billable turn is also keyed to its own stable request content, so resubmitting the same turn with a new prompt can't get billed as if it were new work.
- Write safety Before model output can overwrite a file, it's checked against patterns tuned to catch truncation markers inside comments (the ellipses and "rest of the code unchanged" a model uses when it silently drops content), without flagging real code that legitimately contains the words "rest" or "remaining."
- Fast pass Simple edits resolve through one fast-model call with deterministic verification instead of the full agent loop, escalating to the frontier model only for genuinely complex requests, with a pre-built orientation brief attached so the escalation starts warm. An internal 20-task benchmark showed an 80% fast-path rate at a 276ms median, zero incorrect ships, and task success on par with using the frontier model alone.