Skrrt LLC

We build for the internet we wanted: the one where things felt alive, had rough edges, and were made because they were interesting — not because a committee signed off.

These days the line between human and machine has mostly dissolved. A lot of what talks to your software now isn't a person — it's a model. So we build for that reality instead of pretending it isn't here: formats, tools, and small weird machines meant for a place where humans and agents share the wire.

We ship, watch what resonates, and keep moving. One weird project at a time.

What's here

• raif-llama-3.2-3b-lora — a LoRA that teaches Llama 3.2 3B to emit RAIF, a wire format for the JSON an LLM produces on a tool call. It repairs its own syntax errors, round-trips losslessly to JSON, and costs ~14% fewer tokens. JSON assumes a careful writer; a model isn't one. RAIF designs for how generated output actually fails.

Elsewhere

skrrt.sh · github.com/skrrt-sh · @skrrt_sh