AI platformPrivate source
ARK AI
A provider-facing AI product with a strict request contract.
AI provider and experimentation platform for model access and developer workflows.
Production service with private source. The public contract is documented; internal provider routing is not.
Checked version 2.5.0
Evidence checked 2026-10-30
The problem
An AI gateway that accepts arbitrary request shapes is difficult to bill, cache and debug. Provider differences leak into every client, and a malformed field can silently change cost or behavior.
What I built
- Built the canonical OpenAI-compatible ingress and the internal provider routing around it.
- Added usage accounting with idempotent reserve, settle, release and refund operations.
- Built an agent loop with bounded continuation turns and tool-call limits.
- Committed an internal evaluation dataset and methodology instead of quoting a vague capability score.
Architecture and trade-offs
One canonical parse
Requests are parsed once into a canonical shape, then plan resolution, billing and provider calls use that object. Unknown top-level fields are rejected, and the provider body is rebuilt field by field.
Streaming without a fake stream
Provider streaming is normalized through one path, including a single terminator and sanitizer. The API does not synthesize a stream after buffering a complete response.
Bounded agent work
Continuation turns and tool calls have explicit caps. Read-only calls can run concurrently; writes and destructive calls stay sequential.
Evaluation with a method
The ARK Capability 100 benchmark is a closed-book, no-tools, no-judge set with a committed dataset hash and published per-profile scores. It is an internal benchmark, not an independent ranking.