The mandate: Advance the LLM side of Retell's voice agents — reasoning, latency, and conversational quality in real-time systems. Explore new techniques, train and iterate on models, design evaluation frameworks for subjective conversational quality, and get the results into production. The JD body is identical to the Audio req. The differentiator is the specialization axis: this seat is the LLM-side hire — post-training, reasoning, instruction following, agentic behavior, inference efficiency. The Audio seat is the speech-side hire — ASR, TTS, streaming audio. Recruiters should source these as two distinct pools with modest overlap at multimodal. Both are founding research seats. There is no research org yet; whoever takes this defines what LLM 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; innovate on training paradigms, methods, and inference
-
Evaluation and benchmarking — novel eval frameworks, datasets, and metrics for complex real-world voice tasks
-
Research-to-production — work directly with engineering to deploy findings
-
Human feedback loops — methods for folding 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 required. PhD preferred. Equivalent research-level engineering experience considered — bypass the degree only for a genuinely research-grade portfolio.
-
Advanced ML research: LLM pre-training or post-training, transcription model training, TTS, or multimodal systems. Industry or academia.
-
PyTorch fluency, model architecture depth, and the underlying math. Round 4 tests all three live.
-
On-site in Redwood City — relocation fully covered Profile:
-
Goes 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- or co-author publications at NeurIPS, ICML, ICLR, ACL, EMNLP, COLM, or equivalent. For this req specifically, NeurIPS/ICML/ICLR/ACL/COLM carry more signal than Interspeech.
-
Post-training depth — RLHF, DPO, RLAIF, preference modeling, instruction tuning, reward modeling. This is the highest-value specialization for the role.
-
Inference efficiency — speculative decoding, KV cache optimization, quantization, distillation, streaming generation. Latency is a first-class constraint here.
-
Agentic LLM work — tool use, function calling, multi-turn state, long-horizon task execution
-
LLM evaluation research, especially for subjective or open-ended quality
-
Competition awards Location and visa:
-
On-site, Redwood City. 100% relocation provided.
-
Sponsorship: Yes — H-1B, TN, L-1, E-3, F-1 (OPT/CPT), and O-1. O-1 is listed on this req. Lead with it for published international researchers — most startups won't touch it. Anti-patterns
-
ML engineers who fine-tune off-the-shelf models and call it research. The PyTorch and theory round exposes this in ten minutes.
-
Prompt engineers with a "LLM researcher" title
-
Pure academics with no interest in production or weak engineering — the Backend + AI Practical round is real
-
Researchers who need a large team, mature infrastructure, and a handed-down agenda
-
Anyone who only wants to publish. Retell isn't a lab.
-
Remote requirements — no exception surfaced Who Will Thrive Here
Someone with real LLM research credentials who's tired of waiting in the compute queue and the publication cycle at a big lab. They want their post-training run in front of 50M calls next month. They'll derive the objective and then debug the inference server. Ex-frontier-lab researchers wanting founding scope, and strong NLP/ML PhDs who want production stakes and a latency constraint that makes the problem harder, are the two clearest profiles.