Careers

Founding AI Engineer

San Francisco

Make vision-language models reliable enough to guide a technician through real work, live.

About the role

Data centers and power grids never stop, and the technicians who keep them running can't wait for better tools. Alcor builds wearable AI that sees what they see and talks them through the job, live.

The people who do this work rely on knowledge that was mostly never written down, and the most experienced of them can't be on every site. Our model is how that expertise reaches the technician in front of the equipment. In a data hall or a substation, a benchmark score decides nothing. What matters is reading the right thing on the right part, noticing when the work goes off track, knowing when the model is unsure, and answering fast on a weak network.

You own the model layer that makes that true, end to end: the data, the evals, the inference path, and how the model behaves on a real site. It is not a component someone else scoped. When the model is wrong in the field, you find out why and you fix it.

How do you build an eval from a few hundred field clips that predicts how the model will do on the next site? How does the model tell the right component from the identical one beside it, in bad light and in motion? When should it speak up mid-task, and when should it stay quiet? How do you make "I'm not sure" more common than a confident wrong answer? What do you fine-tune, what do you leave to a tool call, and what has to answer even when the network doesn't?

You'll work with the systems engineers who get frames and audio to your model, the full-stack engineers who put it in the product, and the forward deployed engineers and deployment strategists who bring back every failure from the field. You'll see your model at work on real industrial sites, next to the technicians it guides.

Every confident wrong answer costs a technician's trust. We work with that urgency, in person in San Francisco.

What you'll do

  • Build the multi-step visual reasoning loop that follows real jobs and inspections and calls the tools it needs.
  • Build evals from small amounts of real field data, decide what the model is judged on, and run the loop that turns field failures into better models with every release.
  • Own failure analysis: trace each wrong answer from the field to its cause, and fix it in the data, the model or the system around it.
  • Fine-tune, quantize and serve open-weight vision-language models within latency and connectivity limits, at the edge and on-premise.
  • Build real-time voice and video interaction: conversational guidance and proactive alerts.
  • Decide where to steer the model, prove or drop an approach on real data, and make the model and infrastructure choices the rest of the team builds on.

What success looks like

  • Every model change ships against an eval built from real field data.
  • Confident wrong answers become rare, and the model says when it is unsure.
  • Answers stay fast when the network does not cooperate.
  • Each new site needs less new data than the last.

What we're looking for

  • 3–7 years of relevant experience in applied computer vision or vision-language models, taken through to real users: training, fine-tuning or adapting the models, or building the path that serves them.
  • Experience as a founder or one of the first engineers at a startup shipping AI, vision or robotics products.
  • End-to-end ownership of a vision model's behavior in production, beyond API calls and beyond a component someone else scoped: the evals, the failure analysis and the metrics that decided each release.
  • A habit of judging a model by the errors that matter in use (confident wrong answers, misses, latency), not only by benchmark accuracy.
  • Expert Python and deep learning frameworks, and production experience serving models under latency and hardware limits.
  • In person in San Francisco.

Nice to have

  • On-device or edge inference
  • multi-step agents with tool use in production
  • egocentric or wearable video
  • robotics or driving perception combined with modern VLM work
  • inference and GPU optimization
  • a main-venue paper on top of shipped work
  • data and labeling strategy for rare failures
  • guardrails and fallbacks for model outputs and tool calls

Compensation

$180K–$255K base · 0.25–0.75% equity

This is the base salary range we expect to pay for this role in San Francisco. Where your offer falls within it depends on your experience and skills.

Apply

no form. Write to contact@alcor-labs.com with the hardest thing you have built or deployed.