Machine Learning Engineer (Evals and Voice Models)

Aircall's AI suite includes an AI Voice Agent and AI Messaging Agent that autonomously handle calls, WhatsApp, and SMS, plus AI Assist, which delivers real-time coaching, call summaries, and CRM automation for sales and support teams. We are looking for someone that can build out the evaluation foundation across all of these products and other agentic products. You'll work on voice models, agent capability evals, benchmark design, LLM-as-judge systems, failure analysis, and the infrastructure that ties it together by establishing shared metrics, test sets, and tooling to measure accuracy, resolution quality, and safety consistently across products. You will set up repeatable pipelines for regression testing and benchmarking as models and features evolve so teams can ship confidently without re-inventing evaluation methodology for each product.

Key Responsibilities

• Design and document comprehensive evaluation frameworks for Aircall’s AI agents across voice, chat and messaging.
• Train and fine-tune voice models (TTS, ASR, speech-to-speech) using production and synthetic data, iterating on architecture, data mix, and training strategy to improve accuracy, naturalness, and latency.
• Assess AI generated solutions across training pipelines, experimentation setups, debugging processes, and optimization strategies.
• Analyze system design decisions and identify strengths, weaknesses, and potential failure points.
• Design annotation guidelines and workflows for human-labeled evaluation data, and calibrate LLM-as-judge systems against human raters to ensure automated evals stay trustworthy over time.
• Build and maintain live quality monitoring for deployed AI agents, tracking accuracy, resolution rate, and safety signals in production, and flagging model or data drift before it impacts customers.
• Own the metric contract for every published AI metrics, including definition, population, grain, rollup, validity window.
• Build release gates, the offline regression suite each AI surface must pass before a prompt, model, or config change ships, measuring reliability across repeated trials, not just average pass rates.
• Build voice-specific evaluation: simulated callers across accents, languages, background noise, barge-in, DTMF, and tool failures, with latency and ASR accuracy as first-class quality metrics.

…