Vision AI Engineer - Intern

Train the models that keep people alive in the world's harshest workplaces.Misti AI · Remote or hybrid (London / Lima)
ABOUT MISTI AIWe turn the cameras already installed at mines and industrial sites into real time safety systems. Our vision models run on Nvidia Jetson devices at 4,000 metres, in dust, in heat, on networks that drop for hours at a time. No cloud round trip, no second chances. When the system works, a truck stops, an alert fires, and someone goes home.We are a small team across London, Lima and Delhi, live on real sites today, backed by Fuel Ventures and part of NVIDIA Inception.
THE ROLEEvery vision model is downstream of a data engine. Ours is an AI annotation agent that turns raw site footage into training data, and this internship sits right on top of it.You will do two things that feed each other. You will build the agent, shipping features across a Python backend and a modern web frontend. And you will run the data through it, producing and reviewing the labeled datasets our models train on. Improve the agent on Monday, feel the difference in your own work on Thursday, watch the retrained model run on a live mine a few weeks later.Be clear on the trade before you apply. This is not a labeling job with a nice title, and it is not a pure engineering job either. You will build the tooling and you will do the human judgment work that keeps it honest. The strongest computer vision teams in the world guard their data engine more carefully than their model code. If you already understand why, you will get more out of three months here than most people get from a year somewhere bigger.
WHAT YOU'LL DOBuild the annotation agentShip features and fixes across a Python backend and a React frontend.Strengthen the human in the loop experience: cut the busywork, surface the cases that genuinely need a person.Work with cloud GPU infrastructure and real production data, never a toy dataset.Own the data that trains our modelsProduce and review labeled datasets from real mining and industrial footage: PPE, restricted zones, machine and human segregation.Make the calls automation cannot. Ambiguous frames, weird lighting, occlusion, the errors that quietly break a model three weeks downstream.Hand off clean, training ready datasets and flag where the workflow is losing time.
WHAT WE'RE LOOKING FORSolid Python, and comfort reading and changing a real codebase you did not write.Real interest in computer vision and current SOTA models. Coursework, side projects and self study all count equally.A careful eye. Sloppy labels fail silently, and you are one of the last humans in the chain.Bias to action. Small team, no one queuing work up for you.
NICE TO HAVECommand line, Git, and the ability to get yourself unstuck on a remote Linux GPU box.Anything agentic: tool calling, LLM orchestration, multi step pipelines.PyTorch, VLMs, LLM APIs, React, AWS (S3 / EC2), or annotation tools like Roboflow.Any dataset building or ML project experience. Personal projects absolutely count.Spanish or Hindi is a bonus, not a requirement. Our team spans three countries.
WHY THIS ONEYour work ships in weeks, not quarters, into systems where the stakes are physical.You see the whole loop: data, models, product. Most internships hand you one narrow slice.You work directly with Jalaj, our CTO and cofounder, not through three layers.[Strong interns here have a path to a full time offer.]
HOW TO APPLYEmail [email protected] with:Your CV or LinkedInA link to one thing you built (repo, demo, writeup, anything)Two sentences on what was hard about itThat third line is the one we actually read.
Even if you do not tick every box, great engineers grow into roles. We would love to meet you.