Internship · Engineering
AI & Automation Intern
Ship an LLM feature a business can rely on, which mostly means learning when not to use one. Measured with an evaluation set, not a demo.
- Location
- Remote, 4 hours of overlap with 11:00 to 19:00 Pakistan time
- Hours
- 20 or 40 hours a week
- Length
- 8 weeks, fixed cohort start
- Compensation
- Unpaid. If your work ships to a client or into our own products, we pay you for it
- Applications close
- 31 October 2026
- Stack
- TypeScriptNestJSLLM APIsEvaluation harnesses
About the role
You join a cohort of nine interns building SiteDay, our field production software for sitework contractors. Your headline feature reads a contractor's estimate spreadsheet and proposes which column is the quantity, the unit and the price, so nobody retypes four hundred rows. Get it wrong silently and every budget in the product is wrong, so the feature is designed the other way around: the model proposes, the user confirms, and everything is measured. The eight weeks run in three phases: weeks 1 and 2 you learn the codebase and help write the specification, weeks 3 to 6 you build your slice in parallel with the other tracks, and weeks 7 and 8 are integration, release and a demo you run yourself.
What you will learn
- Why the first move is a deterministic matcher, and the model is the fallback
- Structured output with schema validation and retries, because malformed JSON happens
- Evaluation sets: a labelled corpus, per-field accuracy, and a report per prompt version
- Keeping a human in the loop where a silent error would be expensive
- Handling someone else's commercial data: what is sent, what is redacted, what is logged
- Cost, latency, caching, ceilings and a path that works when the provider is down
What we look for
- Python or TypeScript, and one thing you have built with any web API
- Curiosity about language models and comfort reading documentation
- Git basics
What you would build
Spreadsheet column mapping with a confidence score per column, built on a deterministic matcher first and evaluated against 30+ real-shaped estimate sheets. An impressive demo with no evaluation set fails this track; a modest feature with a measured accuracy number passes it well.
How the 8 weeks run
- Weeks 1 to 2: learn the codebase, read the specification, and ship one small change. The design intern produces the wireframes everyone else builds from
- Weeks 3 to 6: build your slice in parallel with the other eight tracks, reviewed weekly, with a mid-point review in week 4
- Weeks 7 to 8: integration, regression, and a release you demo to the company yourself
- Every week: a written daily standup, a Monday cohort sync, a 30-minute 1:1 with your mentor, and a Friday demo