Govern every dollar of AI spend. Prove every saved dollar, net of quality.
The whole product is one idea: control and visibility are live today (one base_url, caps, kill-switch, attribution, an auditable cost ledger); proving a saving without breaking the product is the hard part, and that proof gate is what would let us turn on a savings share. We run request-side levers at the gateway (a light source-side slim is planned, an optional Intake SDK is coming), all through one eval gate and one ledger — verified savings off · proof pending.
Change one line. Recovea does the rest.
Swap your base URL and key — your existing SDK works unchanged (Python, JS, cURL, Anthropic-shaped). We wrote the gateway ourselves; no third-party proxy sits in your request path. Your provider keys pass through; we’re fail-open to baseline — before the first token we fall back cleanly to your provider, mid-stream we surface a clean error so you can retry — and rollback is the same line you changed.
Recovea Intake (Tier 2) is an optional thin SDK: install, wrap your retrieval and tool-output steps with a connector for your framework (LangChain / LlamaIndex / raw / MCP). Tier 1 is one line and zero app change, available today; Tier 2 is coming in v1.1.
Read the docsfrom openai import OpenAIclient = OpenAI(- base_url="https://api.openai.com/v1",+ base_url="https://api.recovea.ai/v1", # your keys pass through · fail-open api_key=OPENAI_API_KEY,)We record what every call costs, and what it would have cost untouched.
Per request we log input/output/cache/reasoning tokens, model, latency, and your original baseline model, so the counterfactual is computable from day one. The locked baseline period and normalization rules go into the contract. No prompt bodies logged by default.
Zero-risk first. Eval-gated next. Nothing risky without a passing test.
| Lever | Typical net band | Risk tier | Status |
|---|---|---|---|
| Prompt-cache prefix hygiene | do-first · eroding without upkeep | do-first | Planned |
| Batch API migration | do-first · flat provider Batch discount on the slice | do-first | Planned |
| Reasoning / effort trimming | do-first on simple routes | do-first | Planned |
| Dedup / single-flight | prevents stampede spikes · concierge-verified | do-first | Concierge · not self-serve billed |
| Exact cache | varies by repeat rate · concierge-verified | do-first | Concierge · not self-serve billed |
| Model routing / cascading | on the routable fraction, eval-gated | eval-gated | Per route, after it passes (shadow today) |
| Light source-side slim (gateway) | obvious RAG / context trim | eval-gated | Planned |
Cache/dedup savings (byte-identical, zero quality risk) are concierge-verified hands-on, never billed as a live self-serve dollar; model routing runs in shadow until its gate passes. Levers marked Planned are not wired yet and bill nothing. Verified savings and the 25% share are off · proof pending until the eval gate ships at volume. Other tools’ routing headlines are conversational benchmarks on the routable fraction, not a blended number. Semantic cache, heavy LLMLingua-2 compression, distillation, and the Tier-2 Intake SDK are designed but not in v1; they arrive in v1.1+.
Shadow mode on your real traffic. No user ever sees a test.
Traffic mirroring shadows the cheaper candidate on a small sampled slice; a layered oracle scores it; a per-route non-inferiority test decides; a promotion ladder ramps it. Per-route gating is non-negotiable.
- Deterministic checks: schema, required fields, tool-call well-formedness, refusal / truncation detection.
- Golden-dataset gate: the route’s curated regression suite.
- Cross-family calibrated judge: a judge from a different model family, calibrated against human labels where a labeled set exists for the route.
gate must hold at each rung · one-config rollback any time
An auditable, basis-labeled cost ledger. Savings, once proven, net of quality.
Live today: spend attributed per key, team, feature, and lever, with the basis labeled. The savings method is IPMVP: Savings = Baseline − Reporting Period ± Adjustments, adjustments we didn’t cause stripped (traffic growth, provider price cuts, your prompt edits), on cost-per-successful-output. Verified savings and the 25% share are off · proof pending until the eval gate ships at volume.
We wrap the commodity. We build only the part that’s defensible.
- Provider Batch APIs (the 50%-off lane) + native prompt caching
- OpenRouter (long-tail model access)
- A cross-encoder reranker for the light trim
- The eval models themselves (rented, cross-family, calibrated)
- The gateway hot path: our own code end to end — no third-party proxy in your request path.
- The cost ledger: per-lever cost attribution, basis-labeled — the verified counterfactual is off · proof pending.
- The eval / quality gate: cross-family non-inferiority, the layered oracle, the promotion ladder.
- We don’t train a black-box router — per-lever attribution beats squeezing the last routing point.
Blunt answers to the real questions.
Want the number for your traffic?
Free savings scan on a sample of your logs: dollar figure back in 48 hours, no proxy.