MKX Studio is an independent engineering practice building production systems end to end - firmware, full-stack product, and applied AI. The whole stack stays under one roof, instead of split across separate specialists.
Take one on its own or as a stack. The advantage is the overlap - whoever writes the firmware also knows the cloud schema, and the model is built on data that's already understood.
Web apps from schema to ship. Typed end to end - Next.js and React on the front, Hono and Fastify on the back, Postgres with Drizzle or Prisma underneath. Auth, billing, and multi-tenant SSO built in from the start.
LLM apps and the models behind them. Orchestration over MCP, RAG, and tool use - plus deep-learning work built from first principles when an off-the-shelf model isn't enough. Vision, diffusion, and classical ML too.
Bare-metal firmware where microseconds matter. ESP32 and CAN-bus telemetry, custom wire protocols, and real-time vehicle data - from reverse-engineering proprietary buses to shipping the motorsport gateway that streams them trackside.
The platform under the product. Containerized, reproducible, multi-cloud - Terraform and Docker across AWS, Azure, and Vercel, with CI/CD, background workers, and observability that hold up in production.
One area that keeps coming up: motorsport telemetry and Porsche data systems. The whole chain gets built in-house - the sensor firmware, the ingest pipeline, and the models that turn a session into something you can act on. Grounded in real hardware and real cars, from CAN reverse-engineering to trackside acquisition.
ESP32 gateways tap the CAN bus straight off the ECU - decoding DBC signals, timestamping at the source, and surviving the vibration and heat of a live session.
Custom COBS + msgpack wire protocol over BLE and serial into Parquet/Polars - lap, stint, and channel data captured at rate, queryable the moment the car pits.
PyTorch models built from first principles turn raw channels into insight - driver traces, lap deltas, and dashboards that make the numbers legible.
A small studio earns its keep on judgment, not headcount. Three commitments hold across every project.
Reverse-engineered vehicle protocols, hand-rolled transformers, custom wire formats. When the abstraction leaks, the work goes a layer down - not around.
A single thread of ownership from the firmware register to the billing webhook. Infra, auth, data, and interface designed together, so the seams don't become someone else's problem.
Modern LLM tooling is part of how the work gets done and part of what it produces - used daily in the build, and shipped inside the product.
Depth in each column, and the connections between them. A CAN-bus signal on an ESP32 can reach a PyTorch model and a live dashboard without leaving the building.
That's the kind of work MKX Studio likes best. Say where it starts and where it needs to reach, and you'll get a straight answer on how it'd get built.