Skip to content
Low-key black and white portrait of Théo Gachet in a suit

Open to new opportunities

Théo Gachet

AI engineering · ML research · Space exploration · Quant finance

I ship LLM agents and ML systems on financial data, build systematic equity research tooling, and publish on reliable machine learning. Double MSc, ranked 1st at both Polytechnique Montréal and Mines Saint-Étienne.

Production AI
≈ 1 s

to answer about 80% of analyst queries, against 30 to 45 s before, with the agent I built at La Caisse.

Investing
$100K → $140K

student-managed portfolio grown in one year across about 40 holdings, as Vice-President of PolyFinances.

Research
3 papers

one accepted at SPAICE 2026 and two first-author preprints targeting RA-L and T-RO.

Experience

2023 – 2026

Education

Ranked 1st, twice
FrenchNative
EnglishFluent · TOEIC 970/990
ItalianAdvanced
ChineseBasic
ArabicBasic

Publications

1 accepted · 2 preprints · 1 thesis

Skills map

  • Python
  • C
  • C++
  • SQL
  • Julia
  • MATLAB
  • PyTorch
  • TensorFlow
  • Databricks
  • LangGraph
  • LangChain
  • MLflow
  • Docker
  • Git
  • Linux
  • Azure DevOps
  • AWS
  • Streamlit
  • MiniZinc
  • LaTeX
  • HTML / CSS
DomainSkillProject or role

Click a domain to zoom in, any node to see its links

Selected projects

Four in detail, eight in brief

La Caisse (CDPQ) · 2026 · In production

Hybrid conversational agent over a Databricks lakehouse

Problem
Investment teams relied on manual SQL work to answer recurring questions about financial data.
Built
A deterministic SQL engine for the known question families, with an automatic Databricks Genie fallback for the rest. Every answer is traceable to its query. Evaluated in MLflow with deterministic scorers and LLM judges.
Result
About 80% of queries answered in about 1 s, against 30 to 45 s before.

Databricks · LangGraph · MLflow · SQL · Azure DevOps

Learn more
Question in plain language
Router
Deterministic SQL engineknown families · ≈ 1 s
Databricks Genie fallbackeverything else
Answer, with the query behind it
Request routing, schematic

Personal project · In progress

Systematic equity research pipeline

Problem
A backtest is only worth its controls. The goal is a small-scale fundamental equity pipeline where look-ahead, survivorship and data-snooping biases are handled by construction.
Built
Point-in-time S&P 500 universe with EDGAR fundamentals as first reported. Sector-neutral scores turned into alphas, a factor risk model with Ledoit-Wolf shrinkage and a 130/30 optimiser in CVXPY. Walk-forward backtest, costs always included.
Status
In progress. No performance figure is published before the holdout is opened, once.

Python · pandas · DuckDB · CVXPY · pytest · GitHub Actions

Learn more
Point-in-time data
Signals
Alpha
Risk model
Optimiser
Backtest
2010–11 warm-up
2012–2022 research
2023–26 holdout, locked
Stages and time split enforced in code

AWS and National Bank of Canada datathon · 2024 · 1st place

AlexIA: generative AI for financial analysis

Problem
The datathon brief: rethink the tools of financial analysis with generative AI at the core.
Built
One platform for analysts: central access to key indicators, sentiment analysis with Amazon Comprehend, LLM summaries of reports with RAG, portfolio optimisation and governance profiles of boards.
Result
1st of 160+ participants. Presented to a team of CDPQ analysts in December 2024.

AWS · Amazon Comprehend · LLMs · RAG

Learn more
AlexIA
Key indicators
Sentiment analysis
Report summaries, RAG
Portfolio optimisation
Board profiles
Reporting
Modules, schematic. To be replaced by a screenshot

Mila · MIST Lab · 2025–2026 · First-author preprint

AURA*: planning under perception uncertainty

Problem
A planner that treats every uncertain cell as forbidden stops moving. One that ignores uncertainty takes risks it cannot see.
Built
A path planner that turns calibrated perception uncertainty, from 2D imagery and 3D LiDAR, into a continuous risk field and searches it with an adaptive A*.
Result
100% mission success where the geometric baseline falls to 0.4%. Maximum epistemic risk down 94.9% for about 9% extra path length.

Python · PyTorch · SegFormer · PointNet++ · conformal prediction

Learn more
Shortest path (dashed) against a risk-aware path. Illustration, not experimental data

Honors & awards

Certifications

A selection, each links to its credential

Recommendations

Four letters, as PDF

Contact

theo.gachet.pro@outlook.fr

For roles in AI engineering, ML research and quantitative finance. Based between France and Canada.

Off the clock

Notebook

Short notes on what I read, watch and play.

Paris--:--
Montréal--:--

Between two time zones. Whichever one you write from, I answer.

© 2026 Théo Gachet Hand-coded in HTML, CSS and JavaScript Last updated November 2026