Machine Learning (Intern)

Machine Learning Intern — On-Device Models & Spreadsheet Intelligence

Type: Internship — Part-time, self-paced
Duration: 3–6 months
Location: On campus + remote
Eligibility: Current students (.edu email)

You build machine learning that has to perform, not impress: on-device functions that make the product measurably better at real work — smarter spreadsheet and table features, better extraction and suggestion quality, faster responses — on a modest consumer laptop, with no cloud and no paid APIs. Every function you ship is benchmarked against a baseline with real data from real tasks. If it doesn't beat the baseline, you write down why, and that's still a deliverable.

Responsibilities

  • Table & spreadsheet intelligence: prototype ML-driven functions for the table editor — formula suggestion, natural-language-to-formula, data cleaning, categorization, duplicate and error detection
  • Performance benchmarks: define and run accuracy and latency benchmarks for every function on modest consumer hardware (8GB-class machines), and report honest numbers
  • Baseline discipline: every prototype is compared against a simple non-ML baseline; document when the simple version wins
  • Real-data evaluation: test on your own real coursework datasets and program-collected usage data — no toy datasets, no cherry-picked examples
  • On-device constraints: work within local-first limits — models run on the user's machine, quantization-aware, no data leaves the device
  • Chart & data recommendations: prototype chart-type and visualization suggestions from table data
  • Document-to-table extraction: improve pulling structured data out of long-form notes into tables
  • Error analysis & handoff: catalog failure modes, propose fixes, and write an eval writeup (≥10 test cases) engineering can act on

Qualifications

  • Currently enrolled student (CS, data science, statistics, engineering, or adjacent — or demonstrated equivalent skill)
  • Working knowledge of Python and core ML tooling (scikit-learn, PyTorch, or similar)
  • Comfort benchmarking: you can design a fair test and report numbers you didn't like
  • Clear technical writing — eval writeups and benchmark reports are the deliverables
  • Rigorous honesty about negative results: a function that fails its benchmark is a finding, not a failure of you
  • Basic spreadsheet/data-wrangling literacy — you'll be working on table features, so you need to know what tables do

Preferred

  • Course or personal projects in NLP, extraction, recommendation, or structured prediction
  • Hands-on with local/offline model setups (Ollama, llama.cpp, MLX, LM Studio)
  • Familiarity with model quantization and small-model tradeoffs (GGUF, 4-bit, etc.)
  • Spreadsheet power-user — formulas, pivot logic, data cleaning
  • Published a benchmark, eval report, or ML project writeup

Skills & Keywords: machine learning, applied ML, Python, scikit-learn, PyTorch, NLP, natural language to formula, data extraction, spreadsheet automation, recommendation systems, benchmarking, model evaluation, on-device inference, quantization, GGUF, small models, local-first AI, error analysis

COTID AI — Internship Program

Individual, campus-based, self-paced internship running 3–6 months. You claim one element of the product and one role, use COTID free on your real coursework for the full program, and every finding you ship feeds the program — your campus work is the work.

Work

  • Core loop: real usage → captured evidence → improvement brief → feedback report on your element
  • Deliverables reviewed asynchronously by a mentor against a shared rubric
  • Your discipline's lens makes your findings different from every other intern's

What you earn

  • Free COTID for the full 3–6 month program
  • Certificate of completion
  • Opportunity to continue into an official cohort program
  • Not continuing into the cohort? Option to purchase COTID at a discount at the end of the program
  • Portfolio-grade deliverables: benchmarked ML prototypes, eval writeups (≥10 test cases), and benchmark reports with honest numbers

Apply

Apply through the COTID internship application — pick your role and element, and tell us why.

Apply for this job

Source: Wellfound.

Report this job

Wrongly listed? Tell us why.