Retrieval / MLOps / Applied ML
Search that survivesmessy input. useful answers.
I build retrieval systems and the pipelines that keep them alive — including one that answers Roman-Urdu questions from Urdu-script documents, at 0.983 Recall@10 against 0.192, on no GPU.
IDF-weighted term overlap across 26 repositories and 3 roles, computed in your browser. A lexical baseline — the hybrid version is Raabta.
Projects · results · releases
Projects you can inspect, measure, and run.
Selected work
Three projects I’d defend in an interview.
26 repositories · live from GitHub
Machine learning
Raabta
Roman-Urdu questions, Urdu-script answers
Python · FastAPI · React · sentence-transformers
Machine learning
Prognostix
Predicts how long a jet engine has left
PyTorch · FastAPI · Docker · pytest
Systems
DataPrep Studio
A data cleaner that never uploads your file
Rust · TypeScript · Columnar engine · Windows
Ranked by what I’d defend in an interview, not by stars.
All 26 projectsMeasured results
Four figures, each with its baseline.
hover or tap a figure for the caveat
Accuracy, BBC News classification
vs TF-IDF + logistic regression
0.9830 macro-F1 on 419 held-out articles. Fine-tuning buys +2.0 macro-F1 over the linear model — for 1,700× the training time and a 268 MB artefact instead of 12 MB. On a latency budget, the honest recommendation is the linear model.
MRR@10, Roman-Urdu retrieval
vs previous retrieval, same questions
5.8× on the same 120 frozen development questions, after adding a romanized-title route; Recall@10 went 0.192 → 0.983. The 60-question test split is still locked and native-speaker review is pending, so treat this as provisional, not a published result.
kWh RMSE, industrial energy
vs same model on 11 PCA components
R² reads 0.9987, which flatters it — R² only compares against the mean. The number that matters is RMSE tripling when the feature space is compressed to 95% variance.
Topic purity, no labels seen
vs editorial sections recovered
LDA reproduced the newsroom's own section structure without ever seeing a label. Business and tech separate worst at 84% each — and that confusion is semantic, not model error. A telecoms merger is genuinely both.
Two of these four are cases where the model I’d recommend shipping is not the one with the best score. Open each figure to see which one, and why.
Run it yourself
Download the real thing.
3 artefacts · versions live from Releases
Point at a project to open its details. Tap once on touch screens.