The mandate: Advance the model capabilities behind human-like voice agents operating in real-world conditions. Explore new approaches across LLMs and audio models, design novel evaluation frameworks, and prototype systems that improve reasoning, latency, and conversational quality — with your work landing in production. This is a research role at a company with no research org yet. The body of the JD calls it "Founding," which is the honest framing: this person defines what ML research means at Retell. What You'll Own
-
Research and experimentation — new techniques across LLMs and audio models targeting reasoning, latency, and conversational quality in real-time systems
-
Model training — build and iterate on models and pipelines fast; innovate on training paradigms and inference
-
Evaluation and benchmarking — design novel eval frameworks, datasets, and metrics for complex real-world voice tasks. This is arguably the highest-leverage piece; conversational quality is subjective and unsolved.
-
Research-to-production — work directly with engineering to get findings deployed
-
Human feedback loops — methods for incorporating human evaluation into model improvement, especially on subjective conversational quality
-
Frontier tracking — bring new ideas into Retell's product and infrastructure Requirements Hard gates:
-
Master's in CS, ML, AI, or related field required. PhD preferred. Equivalent research-level engineering experience considered — but treat the degree as the default filter and only bypass it for candidates with a genuinely research-grade portfolio.
-
Advanced ML research experience: LLM pre-training or post-training, transcription/ASR model training, TTS, or multimodal systems. Industry or academia both count.
-
PyTorch fluency, model architecture depth, and comfort with the underlying math. The loop tests all three directly.
-
On-site in the Bay Area — relocation fully covered Profile:
-
Can go from open-ended problem to working prototype without a spec
-
Translates research into systems that survive production
-
Communicates complex ideas cross-functionally Strong bonuses:
-
First-author or co-author publications at NeurIPS, ICML, ICLR, ACL, Interspeech, ICASSP, or equivalent. Interspeech and ICASSP are the highest-signal venues for this specific req.
-
Competition awards
-
Speech-specific depth: ASR, TTS, speaker diarization, voice conversion, streaming/low-latency inference
-
Real-time or streaming model experience — latency is a first-class constraint here, not an afterthought
-
RLHF, preference modeling, or eval design for subjective quality Location and visa:
-
On-site, Redwood City. 100% relocation provided — this is on this req and not the others. Lead with it for out-of-market candidates.
-
Sponsorship: Yes — H-1B, TN, L-1, E-3, F-1 (OPT/CPT). Note: no O-1 listed here, unlike the FDE req. Worth confirming with Retell, since O-1 is the most relevant visa for publishing researchers. Anti-patterns
-
ML engineers who fine-tune off-the-shelf models and call it research. The PyTorch and ML-theory rounds will expose this immediately.
-
Pure academics with no interest in production. Half this job is getting research deployed.
-
Researchers who need a large team, established infrastructure, and a defined agenda
-
Applied scientists whose work is entirely feature engineering and classical ML
-
Anyone requiring remote — no exception surfaced
-
Candidates who only want to publish. Retell isn't a lab; publication isn't the deliverable. Who Will Thrive Here
Someone with real research credentials who is tired of the queue at a big lab. They want their model in front of 50M calls next month, not in a paper next year. They're equally happy deriving the loss function and debugging the inference server at 2am. Ex-big-lab researchers who want founding-level scope, and strong PhDs from speech/audio groups who want production stakes, are the two clearest profiles.