I'm a Tunisian high-school student (Sciences de l'Informatique, bac 2028) building applied-AI and health tools that know their limits. I work on applied AI, health informatics, agriculture, and climate risk — small systems with clear limits and readable behavior.

🔬 Honesty engineering — the boy who teaches machines to say: I don't know. My research and my projects ask the same question: what does a system owe the people who trust it, when its data cannot be trusted?

Python · JavaScript · Flask · React · Vite · SQLite · Git — small stack, shipped often.

• 🌧️ Rain2Risk — transparent relative flood-risk screening with documented data limits and tests. Live: rain2risk.vercel.app
• 🌾 CropGuide AI — agricultural decision support with explicit uncertainty, data audits, and tests.

Honest limits, printed on the label: both are exploratory software. CropGuide AI is not a substitute for agricultural expertise; Rain2Risk is a screening score, not an official warning or forecast.

AI-Based Clinical Decision-Support System Robustness Under Unreliable Patient Data — a narrative review and thematic synthesis, published as a preprint on Zenodo in 2026. Fifteen studies, six robustness themes, one conclusion: a clinical AI that cannot say "I don't know" is not safe yet.

Merged work in established projects — every entry links to the merged pull request.

• pandas — fixed a silent data bug: pd.to_numeric(..., errors="coerce") on Arrow-backed strings stored a plain float NaN instead of pd.NA, so the missing values escaped isna() — and next to an existing missing entry, a neighbouring value could be duplicated in their place. Root-caused inside to_numeric, fixed with regression tests, and merged into main for pandas 3.2.0, where I am credited in the release notes. → PR #71991 · closes #67949
• Microsoft Azure documentation — two merged fixes in Microsoft's official Azure docs: a duplicated word in the Azure Monitor metrics article (PR #128901) and in the Blob Storage CLI quickstart (PR #128902).

Currently under review: curl · scikit-learn · xarray · NVIDIA · lighteval

I compete on Zindi as koussayy — ranked #7,237 among 100,000+ AI builders (best rank #7,208) · 49 points · 56 submissions across 5 challenges (standings as shown on Zindi, 10 Nov 2026).

• Financial Stress Prediction Challenge (Beginner Track, October Edition) — predicting financial stress from survey data, scored on log loss + ROC-AUC. Rank #56 of 196, best public score 0.7360881, 25 submissions. Closes 29 Nov 2026.

• Bias Bounty: Mapping Equity Challenge (Humane Intelligence × Radiant Earth) — estimating Overture Maps coverage gaps across 9,794 US census tracts with a fully documented, reproducible, no-training methodology: a rule-based index, a wrong denominator found and fixed, then shrinkage toward regional means. Final standing #132 of 505, private score 0.085128854 (public 0.081706769), 31 submissions. Closed 1 Nov 2026; under code review, with rankings provisional until 22 Nov 2026.

• Green Rising Methane Innovation Challenge — launches 13 Nov 2026, closes 14 Feb 2027.

• AMD Global Physical AI Behavioral Safety Challenge — active, closes 31 Jan 2027 (registered; no submissions yet).

• AI Agents for RAN Challenge by ITU — starts 27 Nov 2026, closes 5 Feb 2027.

• 🏅 Zindi badges earned so far:

_{First Steps · Go-Getter · Zindian}

• AI in Healthcare Specialization · Stanford Online (Coursera)
• AI in Government · Oxford Saïd

_{Thanks for stopping by — the next build is already on the desk. · © 2026 Koussay Mehdouani}