Research Engineer - Agent Intelligence & Evaluation

Company Description

We’re building self-healing voice agents for enterprise customer support within ixigo. The system has to know when it’s failing, why it’s failing, and how to fix itself before a human notices. This fellowship sits at the intelligence layer behind that work.

Job Description

Voice agents fail in ways traditional software doesn't. ASR confidence drops on an accent and a tool call misfires. Latency breaks turn-taking and the LLM hallucinates a policy. A model swap silently regresses production and nobody catches it for a week.

We're building self-healing voice agents for enterprise customer support. This role owns the intelligence layer: the evals that catch failures before shipping, the observability that traces them across the pipeline, and the feedback loops that let agents fix themselves

What you'll own

• Evaluation infrastructure. Audio-native metrics for barge-in, prosody, and turn-taking. Adversarial datasets across accents and edge cases. LLM-as-judge rubrics for task success, tool-use correctness, and recovery.

• Observability across the pipeline. Tracing that correlates audio, STT, LLM reasoning, tool calls, and TTS to a single conversation. Analysis and alerting that surfaces cascade failures instead of hiding them.

• Self-improvement systems. Mine production traces for failure patterns, generate targeted training or prompt data, validate fixes with adversarial replay, and guardrail against regressions.

Qualifications

Who we're looking for

• 3 to 5 years in ML engineering, research engineering, or applied research. Strong Python and modern ML tooling. Depth in at least two of: speech and audio models, LLM agent systems, and eval or observability infrastructure.

• You've shipped something non-trivial where research met production. You read papers, spot when a benchmark measures the wrong thing, and translate ideas from Interspeech, ACL, or NeurIPS into systems that run on real traffic. Publications welcome, not required.