Jeremy MoebelEmail me

I'm an engineering student in Paris, at École des Ponts and in the M2 Data Science at École Polytechnique. From March to August 2026 I did research at the Harvard AI and Robotics Lab, and I build AI tools for companies as a freelancer.

Drag to turn it any way. Type R U F L D B to turn a face, Shift for prime.Drag to turn it any way.

Selected work

A few things I have worked on.

In productionFreelance, 2026

Reading shipping documents for a freight forwarder

With one other developer, for a global freight-forwarding company

Import staff drop a bill of lading or an air waybill into a web app. A vision model fills in the fields, a person checks them, and the file goes straight into the company's forwarding system with no retyping. In use across their offices since September 2026.

FastAPI, Next.js, SupabaseClaude vision2,300+ automated tests
In productionFreelance, 2026

Orders and invoices between two companies

For a luxury furniture manufacturer

When the sales office confirms an order by email, the workflow checks the payment in QuickBooks, has an LLM read the quote, checks every value against QuickBooks and Business Central, and creates the order once a person approves. Invoices then flow back on their own. Live since September 2026.

n8n, JavaScript157 tests3 human checkpoints
Research, 2026Harvard AI and Robotics Lab

Comparing ways to post-train an LLM

ML research intern, Harvard Medical School / Mass General Hospital

I maintained the lab's benchmark for post-training Llama-3.1-8B with SFT, DPO, GRPO, PPO and REINFORCE++ on A6000 and H100 GPUs, with every run evaluated over three seeds on seven benchmarks. I also worked on multi-agent reasoning and diffusion image editing, which became three preprints.

Llama-3.1-8BTRL, DeepSpeed ZeRO-33 seeds × 7 benchmarks
ShippedAuditoo, Station F, 2025

An OCR agent for field technicians

AI engineer intern, Paris

I scoped it with technicians on site, then built and shipped it: point the phone at an equipment label and the diagnostic form fills itself. I also picked the team's speech-to-text model from 8,000 real clips.

OCR + structured extraction8,000 clips over 50 daysWER, latency, cost

How I got here

One key per stop. Press them, they play.

2020

Lycée Louis Pasteur, prépa MPSI/MP

Classes préparatoires in maths and physics. First of 45 in maths in the final year, top 5% in the Mines-Ponts exam.

2023

École des Ponts, MEng

Mathematical and computer engineering. A+ in deep learning, machine learning and convex optimization.

2024

Zumy, iOS app

Started Zumy, an app for rating music, built on my own in SwiftUI and Firebase. In public TestFlight beta since January 2026.

2025

Auditoo, Station F

AI engineer intern. Shipped an OCR agent to production and chose the team's speech-to-text model.

2026

Harvard, AI and Robotics Lab

ML research intern from March to August: an LLM post-training benchmark, and three preprints with the lab.

2026

Freelance, since July

A document platform for a freight forwarder and an order and invoicing workflow for a furniture manufacturer, both in production.

2026

Polytechnique, M2 Data Science

Computer vision, generative models, reinforcement learning, optimal transport.

2027

April, your team?

Open from April 2027 for a 6-month internship: forward deployed, applied AI or ML engineering.

Off the clock

Piano and puzzles.

Jeremy Moebel, in notesJEREMYMOEBELJEREMYMOEBEL

Each letter is a note, with the French cipher behind Ravel's and Debussy's 1909 pieces on the name of Haydn.

Classical piano

I studied piano at the Conservatoire Claude Debussy, up to the advanced level (the French troisième cycle). I arrange pieces, share the sheet music at jeremymoebelsheetmusic.com, and you can listen to my recordings on Spotify.

Speedcubing

I solve Rubik's cubes. The one at the top of the page works too: scramble it, then press Solve.

Get in touch.

I'm looking for a 6-month internship from April 2027, in applied AI, ML engineering or a forward deployed role.

jeremymoebel@gmail.com