Full-time
Research Scientist, Continual Learning
San Francisco, California, United States · On-site
$250,000 - $320,000
Research Scientist, Post-Training & Continual Learning
Most deployed AI agents are only as good as the day they shipped. Every mistake they make in production carries useful information, and almost none of it gets used. This team wants to change that. They’re building agents that improve from their own work: every correction, override and failed task feeds back into training, so each version beats the last without waiting for an engineer to retrain it.
They’re an early-stage AI company automating complex, multi-step operational work for large enterprises. In that setting, a mostly-right answer still costs real money. The founders have built and sold several companies before, the engineering team includes multiple former CTOs, and they already have paying customers.
You’d own how the system learns and how its progress is proven. The work covers:
Benchmarks built from real work and graded by the people who do it, scored on whether the whole task was done correctly
Publishing the research behind the results, alongside academic collaborators
Post-training small open-weight specialist models with SFT, preference optimisation, RL with verifiable rewards and distillation, including per-role models and per-customer adapters
A continual-learning loop that turns production corrections into training signal, with humans reviewing only the decisions that need it
Reward models, verifiers and simulated users built from real interactions
Scoring new frontier models as they’re released, and deciding which model handles which step based on measured accuracy, latency and cost
You’ll have post-trained language models that reached real users, and you’ll be comfortable defending your results under scrutiny. You know how to build training data from messy production logs, read agent traces and name the failure mode. Experience building benchmarks or eval datasets that others used, LLM-as-judge calibration, active or online learning, or multi-GPU training and serving (FSDP, vLLM) is a strong plus, as is time at an early-stage startup.
📍 San Francisco, onsite. Visa support available
💰 Up to circa $320,000 base plus meaningful equity given their early stage
✅ Repeat founders with multiple exits
If it sounds relevant, we’d be happy to talk it through.
All applicants will receive a response.
Salary
$250,000 - $320,000
Location
San Francisco, California, United States · On-site
Contract
Full-time
Your consultant
Marc Powell

Full-time
Research Scientist, Continual Learning
San Francisco, California, United States · On-site
$250,000 - $320,000
Research Scientist, Post-Training & Continual Learning
Most deployed AI agents are only as good as the day they shipped. Every mistake they make in production carries useful information, and almost none of it gets used. This team wants to change that. They’re building agents that improve from their own work: every correction, override and failed task feeds back into training, so each version beats the last without waiting for an engineer to retrain it.
They’re an early-stage AI company automating complex, multi-step operational work for large enterprises. In that setting, a mostly-right answer still costs real money. The founders have built and sold several companies before, the engineering team includes multiple former CTOs, and they already have paying customers.
You’d own how the system learns and how its progress is proven. The work covers:
Benchmarks built from real work and graded by the people who do it, scored on whether the whole task was done correctly
Publishing the research behind the results, alongside academic collaborators
Post-training small open-weight specialist models with SFT, preference optimisation, RL with verifiable rewards and distillation, including per-role models and per-customer adapters
A continual-learning loop that turns production corrections into training signal, with humans reviewing only the decisions that need it
Reward models, verifiers and simulated users built from real interactions
Scoring new frontier models as they’re released, and deciding which model handles which step based on measured accuracy, latency and cost
You’ll have post-trained language models that reached real users, and you’ll be comfortable defending your results under scrutiny. You know how to build training data from messy production logs, read agent traces and name the failure mode. Experience building benchmarks or eval datasets that others used, LLM-as-judge calibration, active or online learning, or multi-GPU training and serving (FSDP, vLLM) is a strong plus, as is time at an early-stage startup.
📍 San Francisco, onsite. Visa support available
💰 Up to circa $320,000 base plus meaningful equity given their early stage
✅ Repeat founders with multiple exits
If it sounds relevant, we’d be happy to talk it through.
All applicants will receive a response.
Your consultant
Marc Powell

Full-time
Research Scientist, Continual Learning
San Francisco, California, United States · On-site
$250,000 - $320,000
Research Scientist, Post-Training & Continual Learning
Most deployed AI agents are only as good as the day they shipped. Every mistake they make in production carries useful information, and almost none of it gets used. This team wants to change that. They’re building agents that improve from their own work: every correction, override and failed task feeds back into training, so each version beats the last without waiting for an engineer to retrain it.
They’re an early-stage AI company automating complex, multi-step operational work for large enterprises. In that setting, a mostly-right answer still costs real money. The founders have built and sold several companies before, the engineering team includes multiple former CTOs, and they already have paying customers.
You’d own how the system learns and how its progress is proven. The work covers:
Benchmarks built from real work and graded by the people who do it, scored on whether the whole task was done correctly
Publishing the research behind the results, alongside academic collaborators
Post-training small open-weight specialist models with SFT, preference optimisation, RL with verifiable rewards and distillation, including per-role models and per-customer adapters
A continual-learning loop that turns production corrections into training signal, with humans reviewing only the decisions that need it
Reward models, verifiers and simulated users built from real interactions
Scoring new frontier models as they’re released, and deciding which model handles which step based on measured accuracy, latency and cost
You’ll have post-trained language models that reached real users, and you’ll be comfortable defending your results under scrutiny. You know how to build training data from messy production logs, read agent traces and name the failure mode. Experience building benchmarks or eval datasets that others used, LLM-as-judge calibration, active or online learning, or multi-GPU training and serving (FSDP, vLLM) is a strong plus, as is time at an early-stage startup.
📍 San Francisco, onsite. Visa support available
💰 Up to circa $320,000 base plus meaningful equity given their early stage
✅ Repeat founders with multiple exits
If it sounds relevant, we’d be happy to talk it through.
All applicants will receive a response.
Your consultant
Marc Powell


