Human simulation for voice agents. Match experienced test agents to the conversation, then examine what actually happens.
Behavioural testing for voice agents.
You don't script the tests. nForge profiles the target, writes the suite, runs the calls, and grades them, then tells you what it still needs to know.
Describe your bot in plain English. nForge classifies it into one of 8 conversation patterns and profiles it with live discovery probes.
An LLM drafts a full test suite: happy paths, edge cases, off-topic probes, and adversarial prompt-injection / data-exfiltration attempts.
Persona-driven callers connect to your bot over WebSocket and hold natural voice conversations, in parallel, recorded per track.
Every run is scored across three pillars, checked against hard/soft assertions, and compared to a baseline for regressions.
Gap analysis surfaces the missing context behind failures and asks clarifying questions. Answers accumulate as project knowledge.
Intelligence to author the tests, a production-grade voice pipeline to run them, and a rigorous evaluator to grade them.
Describe the bot once; nForge figures out what to test and writes the cases.
Mix and match reusable pieces instead of rewriting prompts for every test.
Swap models and transports per call; connect to virtually any voice bot.
Turn subjective “sounds fine” into deterministic pass/fail your pipeline can enforce.
nForge grades each call across three independent pillars: a deterministic performance layer plus two LLM-as-judge layers, so a score always traces back to the transcript.
Snapshot a good run and nForge auto-generates assertions with safety margins (p95 latency +20%, turn count +50%, duration +30%) so the next release fails loudly when it drifts, not silently.
The catalog defines who calls and why; the pipeline defines how nForge reaches your bot and speaks to it.
A reusable caller persona: voice, mood, quirks and the objections they raise.
The bot stage and the state it must hold: the request, the change of plan, the confirmation.
Persona × scenario × goals, with the pass condition written as expected versus observed.
Create a project, connect over WebSocket or Twilio, and generate your first suite. No callers to schedule, no scripts to write.
Start a projectOpen the dashboard