matthieuvion,github

Ex. data scientist @Linkfluence (Radarly). Skillz in NLP, tabular data, time-series, APIs, but curious about anything, really.

Current project : synthetic data gen / classification :

lmd_viz : crafted my own 200k dataset on Ukraine conflict. Comments as a proxy for people engagement.
lmd_classi : manual labeling + baseline model few shots classification w/ SetFit.
WIP : synthetic comments generation w/ Mistral-7B. Next steps will be fine-tune and train a good old Bert for inference perf.

Cool stuff :

wzkd app & match2kd : reverse engineering matchmaking score from players' features only ; using XGB & streamlit App.
wzlight : light wrapper around COD WZ api (discontinuated since). Also available on PyPi.

My favorite AI/ML newsletter (should read it too!) :

Data Machina : not mine, not sponsored. Technical-but-simple enough and existed way before the LLM hype.

matthieuvion's Projects

databricks-xgb

Distributed XGB on Spark with MLFlow model tracking. Code & notes on compatibility of (3!) XGB variants with PySpark/Databricks/MLFlow. Ensure notebook works with Community Edition.

lmd_classi

Few shot classification on Le Monde comments using 🤗 SetFit

lmd_ukr

Custom dataset of Le Monde (236 k comments), 1-year-coverage-on-Ukraine & custom made associated API (need Subscriber credentials)

lmd_viz

https://matthieuvion.github.io/lmd_viz/ 236k comments of Le Monde on Ukraine. A proxy to measure people' engagement. Semantic search & SBERT models testing via Sentence-Transformer / Faiss

lstm-gpu-jupyter

LSTM (PyTorch) , time series prediction: a benchmark case w/ and w/o stationarity + scaling

match2kd

Save thousands of API calls. Custom model & dataset aiming at predicting a game difficulty score ("lobby avg kd") without calling players' games history stats and profiles.

spark-cluster

Steps to deploy a local spark cluster w/ Docker. Bonus: a ready-to-use notebook for model prediction on Pyspark using spark.ml Pipeline() on a well known dataset

Streamlit-based dashboard that collect, aggr. and visualize player's stats from Call of Duty Warzone1. Also deploy our model (cf. /match2kd repo) to predict game difficulty ("lobby kd") from game features.