How it works

One pipeline, five stages, zero AI. Usage comes in — pulled daily from your connected account, or from a file you upload — and a priced, evidence-backed findings report comes out. Every stage is deterministic: the same data produces the same report, byte for byte.

flowchart LR
    A[Connect: daily pulls<br/>or upload JSONL / CSV] --> B[Ingest & normalize<br/>token counts only]
    B --> C[Price<br/>four-rate table,<br/>effective-dated]
    C --> D[Detect<br/>six waste classes]
    D --> E[Report<br/>JSON + web + PDF]
    E --> F[Purge<br/>raw upload gone in 7 days]

Stage by stage

Ingest and normalize. The parser detects your format per file (OpenAI JSONL, Anthropic JSONL, or the generic CSV contract) and normalizes every call into one record: timestamp (UTC), provider, model, and the four token counts. Text fields are dropped at the door — they never reach the database (see Data handling). Invalid rows are counted and reported, not silently patched.

Price. Each call is priced at the provider rate in effect at its timestamp, using a versioned, human-verified four-rate table (input, output, cache write, cache read). Models without a verified rate card are listed as unpriced and excluded from totals rather than guessed (see Pricing data).

Detect. Six independent detectors walk the priced call frame, each looking for one waste class. Every finding carries an estimated monthly dollar impact, a severity, a confidence label, and up to 20 evidence rows — token counts and timestamps, never text. Estimates are conservative by construction; each detector's exact formula and haircuts are documented on its own page.

Report. All numbers are assembled once into a single report model, then rendered three ways from that one source: deterministic JSON, the private web report, and the PDF. The renderers never recompute anything — what you read in the PDF is what the engine computed, verbatim.

Purge. Seven days after your report is generated, the raw upload is automatically deleted and the deletion is written to an append-only audit log.

Why determinism is the feature

The analysis engine imports no LLM and makes no network calls — this is enforced by an automated import-guard test, not a code-review convention. A deterministic engine cannot hallucinate a savings number, cannot leak your logs to a model provider, and produces reproducible reports you can hand to finance.