AI workflow automation proof
Every LLM call in the system is traced — provider, model, input/output tokens, latency, and cost in USD — and the traces feed a live cost dashboard in the admin. Snapshot from the build log (July 2026): 407 traced calls, ~300K tokens, $1.24 total spend across Claude Sonnet, Claude Haiku and OpenAI embedding models.
100% (407 of 407)
- Proof type
- metric
- Confidence
- verified
AI generation ran untracked: no record of which model was called, what it cost, or how long it took.
Each call writes a trace (timestamp, task, provider, model, tokens, cache hits, duration, cost) to an append-only log surfaced on the analytics dashboard — e.g. 407 calls / 299,742 tokens / $1.24 in one development window.
How it was measured
Every LLM call routes through one traced client that appends a JSON line — timestamp, task, provider, model, input/output tokens, duration, cost — to data/llm_traces.jsonl. The 407 calls, ~300K tokens and $1.24 are a straight count and sum over that file for one development window in July 2026. Cost is computed from a per-model price table in code at the time of the call.
Honest limitations
This is development traffic, not production load — it says the tracing works, not what the system costs at scale. Cost is derived from a hardcoded price table rather than reconciled against provider invoices, so it drifts when prices change. The "100%" holds only for calls made through the traced client; anything calling a provider SDK directly would not appear.