My whole academic path at Université Grenoble Alpes: the
Licence MIASHS
in Bachelor/, and the
Master of Artificial Intelligence in Master/
as it is produced.
Source sits next to its rendered output, so everything reads without cloning.
The M1 is under way and Master/ fills as it goes.
Browse it as a site ↗, where the results are charted from the files that produced them and the Battleship bot can be watched playing a recorded game.
Graded work that runs end to end and reports a measured result.
Perfume satisfaction (R, Quarto). Binary classification over roughly 24 000 fragrances. LASSO logistic regression, a pruned decision tree, a random forest and kNN compared on ROC curves and confusion matrices; Naive Bayes and K-means were tried and rejected on stated grounds. Features map to 10 olfactory families and the 70/30 split is built against leakage. Read the report ↗, 24 pages, also served as a navigable HTML book. Group project.
Grenoble housing prices (R, Python). Hedonic pricing on DVF open data. The log transform is what makes the linear model work, lifting R² from 0.096 to 0.275. A random forest then cuts RMSE from €265,214 to €58,750, about 4.5 times better, by giving up exactly the interpretable implicit prices the hedonic method exists to produce. The write-up argues the two models answer different questions. Group project.
Battleship AI (Java). Four targeting strategies benchmarked head to head over 300 games each. Markov transition matrices win 74.3% of games and clear the board in 54.8 shots against 94.3 for uniform random, a 42% reduction. Monte Carlo costs far more computation for a slightly worse result, which is the interesting part. Pair project.
Strategy tournament (Prolog). Two strategies derived on paper, written as agents, and entered against thirteen other students'. They placed 7th and 8th of 16. Reading the 636-page log match by match shows the standings measuring something other than winning: the champion won 6 of its 15 matches, the agent that won 13 finished 14th, and both entrants beat the champion head to head. Across the field, finishing lower correlates with winning more matches. Rechecking the equilibrium behind them showed the Nash mixture strictly dominates the submitted one, earning more against every opponent by at least 0.1402 while still losing the head-to-head. Pair project.
Full-stack web application (Node.js, Express, SQLite). Nine practicals building up to a complete server: middleware, SQLite persistence, password hashing, sessions and authentication. Solo.
| R | tidyverse, caret, glmnet, ranger, rpart, pROC, deSolve. Machine learning, econometrics, statistics, analysis |
| Java | Object-oriented design, modules, GUI, tournaments under CI. Battleship AI, the OOP course |
| JavaScript | Node.js, Express, SQLite, sessions and authentication. Web |
| SQL, Prolog, Racket | Relational, logic and functional paradigms. Language modules, the Prolog tournament |
| Reporting | Quarto, R Markdown, LaTeX. Every report here renders from its source |
Full curriculum, semester by semester, in Bachelor/.
| Subject | ||
|---|---|---|
| Statistics | Estimation and testing | Bootstrap, maximum likelihood, Fisher information, chi-squared. Solo. |
| Statistics | Regression and ANOVA | Linear regression, one-factor ANOVA, non-parametric tests. Solo. |
| Analysis | Fourier and ODEs | Fourier series and transform, Euler and Runge-Kutta solvers. Solo. |
| ML | Classification practicals | Eight worked practicals on the Titanic data. |
| Economics | Advanced econometrics | Follows on from the housing-price study above. |
| Programming | Object-oriented programming | The Java course, 59 source files. |
| Programming | Functional · Logic · Relational | Racket, Prolog and SQL practicals. |
Code and writing are MIT; see LICENSE. NOTICE matters here: MIT covers my own work only. Practical subjects, handouts and published papers belong to their authors and are cited rather than redistributed.