Optimize AI
Automatically

Deadpipe is laser-focused on optimization: we learn from your successful calls to automatically improve performance, reduce costs, and prevent issues. We build statistical models per prompt that suggest optimizations, validate inputs, and ensure consistent quality—all with one line of code.

Use our universal wrap() function: pip install deadpipe, then wrap(client, app="your_app") and pass prompt_id per call. Auto-detects your provider (OpenAI, Anthropic, Gemini, Mistral, Cohere). Tracks all calls with full context. Baselines build after ~10 calls.

Your LLM calls continue running normally. Deadpipe is fail-safe by design—if our servers are unreachable, your application doesn't fail. Telemetry is fire-and-forget, asynchronous, and never blocks your critical path. We never break your app.

The wrap() function extracts prompts, tools, and system messages from every call. Creates hashes for change detection—know exactly what changed when behavior shifts. Zero manual work.

After approximately 10 calls per prompt_id, we establish statistical baselines for latency (mean, p50, p95, p99), token counts, schema pass rate, empty output rate, and more. Each new call is compared against the baseline, and we alert when metrics exceed thresholds (e.g., latency > p95 × 1.5).

We can't detect hallucinations directly, but we track practical signals that something went wrong: JSON parse failures, schema validation errors, out-of-range enum values, out-of-bounds numeric fields, empty outputs, refusals, and output pattern changes. These proxies catch many real-world failures.

Yes! For streaming, call tracker.mark_first_token() when your first chunk arrives to capture time-to-first-token (TTFT). Then call tracker.record() with the final assembled response. We capture both TTFT and total latency automatically.

The SDKs support 5 major providers with the universal wrap() function: OpenAI (GPT-4o, o1, o1-pro), Anthropic (Claude Opus 4, Sonnet 4, Haiku 4), Google AI (Gemini 2.0, 1.5), Mistral (Large, Medium, Codestral), and Cohere (Command R+, Command R). Each has automatic provider detection, response parsing, and cost estimation. You can also use track() for any other provider.