Founding ML/Research Engineer

This hire turns Colare’s unique assessment data (expert-calibrated rubrics and recorded hardware-engineering workflows) into tasks and signals that models can learn from. The work: design and implement RL environments and evaluations for agentic CAD and ECAD work across mechanical and electrical engineering; build the graders, reward signals, and verification systems that score engineering work reliably; shape assessment data into datasets and environments frontier labs can use in research and production training runs; collaborate with the ME/EE engineers who define correctness and with lab teams who use the environments for post-training; and run experiments that help shape the research and product roadmap.

First month is heavy on data-pipeline work: Colare produces a lot of data and has plans to put it to use, including conversations with frontier labs and scaling startups that may acquire that data, so the early job is aligning what gets built to the data being produced. The second piece is automating the processes Colare is scaling (scoring, question creation), which calls for an understanding of training workflows (SFT and beyond, depending on direction). The bottleneck this solves: making those processes scalable faster, so Colare can onboard the demand it is seeing and free the talent-science team to spend their days reviewing submissions and driving better signal, instead of scoring and creating questions fully by hand. Six months out, the role likely changes substantially as the company grows.

Nain was specific that this is an ML engineer, not an AI applications engineer: he does not want someone coming in to do RAG or prompt engineering. RL is the frontier method for most of these projects.

What You’ll Own

• RL environments and evaluations for agentic CAD and ECAD work (mechanical and electrical)
• Graders, reward signals, and verification systems that score engineering work reliably
• Shaping assessment data into datasets and environments frontier labs can use in post-training
• Data pipelines that align what Colare builds to the data it produces, including lab and acquirer conversations
• Automating scoring and question-creation processes
• Experiments that shape the research and product roadmap

Requirements

Must-haves:

• Hard-tech grounding: ML/AI experience at a hard-tech company, OR close work with EE/ME engineers, OR an EE/ME educational background (needed to grasp the value of the work and apply ML to it)
• CAD / ECAD / 3D vision or similar hardware-adjacent ML knowledge (you will live in CAD/ECAD environments and work with ME/EE engineers, though not designing schematics yourself)
• RL environments experience from either side of the chain: building and selling them, or buying and using them for post-training

Strongly wanted

• Senior-level ML experience; the core of the work has been RL, data pipelines, evals, LLM-as-a-judge (not primarily computer vision or SFT)
• Hands-on with evaluations, grading, data pipelines, and ML infrastructure
• Track record of high initiative and ownership
• “Vision” (the decisive separator): takes initiative in the right direction. When they step outside what was assigned, it is exactly what should have come next. This is what moves a candidate to the top of the band.

Background / Logistics

• ~3 to 6 years full-time experience (not so senior they have lost the grind, not so junior they can’t apply the knowledge)
• Prefers scaling-startup backgrounds (Series D/E or smaller) who “get” ambiguity; big tech fine if the experience is strong (e.g., post-training at Gemini), accepting the trade-off
• In person SF, 996; relocation expected if not local · Visa: generally no; will consider a simple TN case for a Canadian/Waterloo grad

Who Will Thrive Here

• An ML engineer who has built or used RL environments and wants to work at the frontier of applying AI to real hardware engineering
• Someone fluent enough in hard tech (by work or degree) to see why CAD/ECAD evaluation matters
• A high-initiative builder who always knows the one thing that will move the needle next
• A scaling-startup type comfortable with ambiguity and a 996 pace