// shipping models to production since 2018

JonathanNguyen

Staff AI Engineer at

Deep LearningLLM Systems

Staff AI Engineer

I build LLM and computer vision systems that run at scale — decks generated from a prompt, fraud detection across a thousand retail checkouts, and real-time bidding at a hundred thousand requests a second.

Years in ML: 8+
Companies: 6
Kaggle Gold Medal: 1×

// things I've built

Selected Work

AI Slide Generator

2026 — present

A prompt goes in, a polished and structured PowerPoint deck comes out. Built at UpSlide for the people who make client-ready decks all day.

Built the generation pipeline and the Office integration end to end.
The hard requirement isn't fluency, it's consistency: slides have to be reliable and hold up against professional design standards on every run, not just the demo.

Python
LLMs
PowerPoint / OOXML
Office Add-ins

BetterJobsHunt

2026 — present

An AI job-search workspace for people who'd rather send ten targeted applications than a hundred blind ones. Scores job fit, rewrites experience into STAR bullets, and tracks the whole pipeline in one place.

Fit and ATS scoring with keyword gap analysis, so a candidate sees what's missing before they apply rather than after the rejection.
Human-in-the-loop by design: the model drafts and rewrites, but nothing is auto-submitted — every application needs approval before it leaves.
Pipeline tracking with follow-up reminders, company research briefings and interview prep packs, plus pattern analysis across outcomes.
Subscription product with a free tier and credit-based usage, live and taking customers.

LLMs
Resume Parsing
ATS Matching
SaaS

Site

Self-Checkout Fraud Detection

2024 — 2026

A deep learning computer vision system that detects and tracks fraud at self-checkout across Carrefour's European estate.

Runs across more than 1,000 registers, targeting a 2% revenue improvement.
Built for the messy end of CV: varied store lighting, occlusion, and a false-positive budget that has to keep cashiers' trust.

Python
PyTorch
Computer Vision
GCP

JovyanAI

2024 — 2025

An AI copilot for data science and machine learning work — code assistance and faster experimentation for people who live in notebooks. Cofounded as CTO.

Reached €2,000 monthly recurring revenue.
Full-stack LLM system on Google Cloud: RAG pipelines, multi-agent architectures, Next.js and TypeScript.

LLMs
RAG
Multi-Agent
Next.js
TypeScript
GCP

Site

NMR Property Prediction

2020

A machine learning strategy for predicting magnetic interaction between atoms in a molecule — a Kaggle gold medal that became a peer-reviewed paper.

Gold medal against roughly 3,000 participants.
Published in PLOS ONE as a machine learning strategy for finding NMR property prediction models.

Python
PyTorch
Graph Neural Networks

Paper

Vertex AI MLOps Platform

2024

An internal platform that automates the end-to-end ML pipeline, from training through deployment and monitoring.

2 highlights

Took the deployment cycle from three weeks to one day.
Made reproducibility the default rather than a per-project effort — the same practices now apply across the team's models.

Vertex AI
GCP
Terraform
Docker
Airflow

Real-Time Bidding & Ad Network

2021 — 2024

An ad network built from scratch at Voodoo, plus the ML that decides what to bid, for a portfolio of mobile games with hundreds of millions of players.

3 highlights

Generated €1M in monthly revenue as the layer between advertisers and publishers.
Served recommendations and bids at over 100,000 impressions per second across 40 million daily active users.
Automated user-acquisition bidding for 200+ games against a $100M annual budget, replacing a manual process it improved on by 60%.

Python
Recommender Systems
Real-Time Bidding
Spark
AWS

Vietnamese Document & Rental Assistant

2022 — 2024

Two production NLP systems at DopikAI: a rental-enquiry chatbot and on-premise information extraction for Vietnamese documents.

2 highlights

The chatbot combined RAG with text-to-SQL to answer apartment questions in context, at over 1,000 requests a day.
Document extraction ran on-premise, which shaped every decision about model size and inference cost.

Python
RAG
Text-to-SQL
NLP
Docker

Site

Multilingual Speech Recognition

2020 — 2021

Speech-to-text R&D at BNP Paribas Cardif across Spanish, Italian and French, built to make call auditing affordable rather than to chase a benchmark.

3 highlights

Halved both the volume of audited calls and the cost of auditing them, with a recognition and text-audit pipeline built end to end from annotation through training.
Reached word error rates below 15% on very limited internal data, which is what made unsupervised and self-supervised methods the point rather than a nice-to-have.
Led the ML side of the work, including the research track on self-supervised learning for speech.

Python
PyTorch
Speech Recognition
Self-Supervised Learning

Document OCR Pipeline

2018 — 2020

A production document-extraction pipeline at BNP Paribas Cardif: find the layout, find the text, read it, and do it at the volume an insurer actually receives.

3 highlights

Cut the time and cost of extracting information from documents by 60%.
Detected layout and text with object- and text-detection models, then read it with in-house recognition models reaching word error rates below 5%.
Owned the deployment path as well as the models — the Docker environment, the API and the CI/CD pipeline.

Python
Deep Learning
OCR
Computer Vision
Docker
CI/CD

// where I've worked

Experience

Staff AI Engineer · UpSlide
Jun 2026 — Present
Paris, France
- Built an AI-powered slide generator that turns a user's prompt into a polished, structured PowerPoint deck, cutting the manual effort of building slides from scratch.
- Developed the generation pipeline and the Office integration, improving slide consistency and reliability and keeping output aligned with professional design standards.
- Python
- LLMs
- PowerPoint / OOXML
- Office Add-ins
Lead Machine Learning Engineer · Carrefour
Feb 2024 — Jun 2026
Paris, France
- Built and deployed a deep learning computer vision system that detects and tracks fraud across more than 1,000 self-checkout registers in Europe, targeting a 2% revenue improvement.
- Cut the ML deployment cycle from three weeks to one day by building a Vertex AI–based MLOps platform that automates the end-to-end pipeline, and established the team's practices for ML system design and reproducibility.
- Deployed a deep learning people-counting system for real-time queue monitoring in stores, giving staffing decisions actual data to work from.
- Led the ML team, delivering AI work across functional boundaries.
- Python
- PyTorch
- Computer Vision
- Vertex AI
- GCP
- MLOps
Cofounder & CTO · JovyanAI
Dec 2024 — Dec 2025
Paris, France
- Cofounded an AI copilot for data science and machine learning work, reaching €2,000 monthly recurring revenue.
- Built an assistant that automates data science workflows with intelligent code assistance, shortening the loop between idea and experiment.
- Designed and implemented the full-stack LLM system on Google Cloud — RAG pipelines, multi-agent architectures, Next.js and TypeScript front to back.
- Python
- TypeScript
- Next.js
- LLMs
- RAG
- GCP
Lead Machine Learning Engineer · DopikAI
Jan 2022 — Dec 2024
Paris, France
- Delivered a production Facebook Messenger chatbot using RAG and text-to-SQL to answer apartment rental questions in context, at over 1,000 user requests a day.
- Built on-premise deep learning for extracting structured information from Vietnamese documents.
- Python
- RAG
- Text-to-SQL
- NLP
- Docker
Senior Machine Learning Engineer · Voodoo
Dec 2021 — Jan 2024
Paris, France
- Built an ad network from scratch — the layer between advertisers buying placements and publishers selling them — generating €1M in monthly revenue.
- Deployed ML models and a recommendation system for real-time bidding at over 100,000 impressions per second across 40 million daily active users.
- Improved the manual bidding process by more than 60% with models predicting user lifetime value, install volumes and margins.
- Automated user-acquisition bidding across mobile ad platforms for 200+ games and a $100M annual budget.
- Built an internal LLM Slackbot that answers employee questions against internal and external sources.
- Python
- Real-Time Bidding
- Recommender Systems
- Spark
- AWS
Senior Machine Learning Engineer · BNP Paribas Cardif
Dec 2020 — Dec 2021
Paris, France
- Led the ML side of speech recognition R&D, recognising and translating Spanish, Italian and French speech to text.
- Halved the volume of audited calls — and the cost of auditing them — with a speech recognition and text audit pipeline built end to end, from annotation through training.
- Reached word error rates below 15% on very limited internal data using state-of-the-art unsupervised methods.
- Supervised the research track on unsupervised and self-supervised learning for speech.
- Python
- PyTorch
- Speech Recognition
- Self-Supervised Learning
Machine Learning Engineer · BNP Paribas Cardif
Apr 2018 — Dec 2020
Paris, France
- Cut the time and cost of extracting information from documents by 60% with a scalable, production-ready extraction pipeline.
- Detected document layout and text using state-of-the-art object and text detection models.
- Built in-house recognition models reaching word error rates below 5%.
- Owned the Docker environment, API and CI/CD pipeline for deployment.
- Python
- Deep Learning
- OCR
- Docker
- CI/CD

JonathanNguyen

Who I Am

Education

M.S.E, Financial Engineering

BSc, Economics and Management

Selected Work

Machine Learning

Languages

ML Frameworks

Cloud & Data

MLOps

Also On My Desk

Experience

Staff AI Engineer · UpSlide

Lead Machine Learning Engineer · Carrefour

Cofounder & CTO · JovyanAI

Lead Machine Learning Engineer · DopikAI

Senior Machine Learning Engineer · Voodoo

Senior Machine Learning Engineer · BNP Paribas Cardif

Machine Learning Engineer · BNP Paribas Cardif

Get In Touch

Elsewhere

Book A Meeting

Good for

# Jonathan Nguyen

## Built with

## Keyboard shortcuts

## A note on the assistant