# Zelto > Zelto builds custom models for voice AI agents, trained on your calls and held to your rules, and runs them for you behind one LLM endpoint. Zelto Observability, its second product, traces and evaluates every production call. Custom models install by swapping only the LLM endpoint behind your agent; your voice platform, speech-to-text, voices, and tools stay as they are. Works with Vapi, Retell, Bland, LiveKit, Pipecat, Telnyx, ElevenLabs, and custom stacks. Trusted by voice AI engineering teams including Domu (YC S24), Altur, and Colektia. Observability handles 1M calls every month, 11k hours of conversations, 400+ agents monitored. ## Core - [Homepage](https://zelto.ai/): Custom models for voice AI agents: benchmark, dataset, training, inference, results, and the one-endpoint install. - [Observability](https://zelto.ai/observability): Voice agent observability, QA and evals: findings, OpenTelemetry traces, monitors with alerts, experiments, simulations, and MCP triage from Claude Code, Cursor or Codex. Self-serve, with a one-line install prompt for coding agents. - [Integrations](https://zelto.ai/integrations): Per-platform pages for Vapi, Retell, Bland, LiveKit, Pipecat, ElevenLabs, Telnyx, and Speechify. - [For voice AI startups](https://zelto.ai/for/voice-ai-startups): QA without a QA team — every call graded, fixes drafted for coding agents. - [For voice AI agencies](https://zelto.ai/for/voice-ai-agencies): One QA layer across every client agent, with per-customer Slack alerts. - [Vapi vs Retell](https://zelto.ai/compare/vapi-vs-retell): Neutral comparison of the two platforms — building model, model stack, pricing shape, testing — and what breaks on each in production. - [Vapi alternatives](https://zelto.ai/compare/vapi-alternatives): Retell, Bland, ElevenLabs Agents, LiveKit, Pipecat, and Telnyx compared — who each fits and which failures follow you after you switch. - [Retell alternatives](https://zelto.ai/compare/retell-alternatives): The same comparison for teams leaving Retell. - [Bland alternatives](https://zelto.ai/compare/bland-alternatives): The same comparison for teams leaving Bland. - [Pipecat vs LiveKit Agents](https://zelto.ai/compare/pipecat-vs-livekit): Two open-source voice agent frameworks compared — architecture, hosting, telephony, built-in testing — and what breaks on each in production. - [How to test voice AI agents](https://zelto.ai/guides/how-to-test-voice-ai-agents): The four layers of voice agent testing — text-mode tests, simulated callers, real calls, production QA — what each catches and misses. - [Voice agent evaluation metrics](https://zelto.ai/guides/voice-agent-evaluation-metrics): Outcome, conversation-quality and pipeline metrics for voice agents, and how to turn them into a rubric. - [Docs](https://docs.zelto.ai/): Product documentation, integration guides, and the `calls.ingest` API reference. - [Privacy Policy](https://zelto.ai/privacy): How Zelto collects, uses, shares, secures, and lets you control personal information and customer content. - [Terms of Service](https://zelto.ai/terms): Terms that govern access to and use of Zelto's website, application, and APIs. - [llms-full.txt](https://zelto.ai/llms-full.txt): Full long-form content for LLM ingestion (product, features, integrations, FAQ). - [Sitemap](https://zelto.ai/sitemap.xml): Canonical URL index. ## Custom models Four steps: 1. **Benchmark.** Your rules become a benchmark, built from your input, cases cut from production calls, and red teaming. Every version is scored on it. 2. **Dataset.** Your calls become the dataset, pseudonymized on import: good real calls kept as they are (~55%), calls that broke a rule repaired (~6%), and calls written for rare cases (~39%), evened out so every rule gets enough practice. 3. **Training.** Zelto fine-tunes a small model that writes itself a one-line note before each decision. A version ships only when it holds every critical rule. 4. **Inference.** Zelto runs it: shared capacity (variable volume, lower cost, reserved concurrency), dedicated capacity (better price at huge volumes), or own the model. Results on one collections agent's 280-case benchmark, against the model it ran in production with its full prompt: critical-rule cases failed 11 → 0 (of 104), benchmark cases failed 35 → 15, prompt per turn ~22k → ~2.9k tokens. Median time to first token: 56 ms (measured on Zelto's GPU). ## Observability Three capabilities: - **Observability** — every production call across every agent and provider, streamed in with LLM traces. - **Evals** — every call scored against a rubric Zelto auto-learns from your calls. - **Findings and fixes** — calls that fail for the same reason grouped into one Finding ranked by call volume, each with a drafted prompt + Linear ticket you can hand to Claude, Codex, or Cursor. ## Key features - Findings: recurring failures grouped by the calls that share them, ranked by traffic. - Fix: every finding ships with a prompt and a Linear ticket for your coding agent. - Alerts: per-customer Slack channels so the account owner hears regressions instantly. - Ask Zelto: in-product copilot that answers plain-English questions about your voice ops. - Listen to the moment: jump to the failing seconds of a call, not the full recording. - Learning: corrections you make propagate across every call that shares the pattern. ## Pricing Custom models: pricing depends on capacity (shared, dedicated, or owning the model); schedule a demo at https://cal.com/zelto/demo. Observability: you can try it for free: create an account at https://dashboard.zelto.ai/signup. Paid pricing is not published. Schedule a demo at https://cal.com/zelto/demo and we'll scope a plan to your call volume. Usage is metered per minute of connected calls; you can set a monthly limit to control spend. ## Integrations Native: Vapi, Retell, Bland, LiveKit, Pipecat, Telnyx, ElevenLabs. Custom models replace the LLM endpoint behind your agent. Observability for custom stacks via REST (`calls.ingest`). ## Contact - Schedule a demo: https://cal.com/zelto/demo - Try it free: https://dashboard.zelto.ai/signup - Email: hello@zelto.ai