26 projects
Everything I've built
Pulled live from GitHub, so this stays current whether or not I remember to update it. Every metric shows what it beat and what it cost — hover any number to see the rest of the story.
Live from the GitHub API13 with demos live or queuedSorted by weight, not recency
+0.792 absolute on the same 120 frozen development questions, after adding the romanized-title route. Retrieval measurement on a title-oriented set — not a claim that 98.3% of open questions get a correct final answer. The 60-question test split is still locked and unused.
Machine learning
ITSimplera Internship — Six Weeks, Six Deliverables
From first EDA to a fine-tuned transformer
Fine-tuned DistilBERT, 0.9830 macro-F1 on 419 held-out articles. Costs 1,700× the training time of the linear baseline for +2.0 macro-F1.
Ingest, train, evaluate, serve, monitor — with a drift check and alerting. Model performance numbers aren't published yet; that's the next thing this repo needs.
125,225 rows kept, measured on a 13th-gen Core i7-13700H with 15.7 GB RAM. Tested up to one million rows — larger files are expected to work but have not been measured, and peak memory use isn't instrumented yet, so neither is claimed.
Title, author, ISBN, publisher, barcode, call number, branch, language, year, subject, recommendation, compound filters and general full-text — plus fuzzy fallback for misspellings. Intent detection is regex and keyword based, not a learned classifier.
90%+ accuracy on the CDC cohort. The point of this project is the derivation, not the score — scikit-learn would beat it in three lines, which is exactly why writing it by hand was worth doing once.
Trained on listing asking prices, not sale prices, so it predicts what sellers ask rather than what property fetches. Listings were collected for a course project; anyone rebuilding this should check Zameen's terms first.
Implemented from the specifications for learning. Never deploy hand-rolled cryptography — this exists to prove I understand what the audited libraries are doing, not to replace them.
The easiest thing here to put in front of a stranger, and the most obvious gap in this portfolio until it's deployed.
Educational build — a real file system needs journalling, concurrency control and crash consistency, none of which this has.
Which is most of why it earns a place on this list — everything else is source you have to read. It's a presentation layer, though: the engineering claims are in the projects it links to, not in it.
Role separation is enforced in the API layer. It hasn't had a security review, so treat the access control as coursework-grade.
Dijkstra on a static graph across 50+ nodes. Real delivery routing needs live traffic and time windows, which turns it into a different problem entirely.
RLE only compresses runs — on text or already-compressed data it can produce output larger than the input. It's the right algorithm for the exercise, not for general use.
Forecast accuracy is the upstream provider's; this project owns the pipeline and UI.
Ten hand-constructed scenarios — enough to show the selection logic works, not enough to claim it generalises.
Machine learning
Hasnat Khan010 github io
Hasnat Khan's GitHub-style AI and machine learning portfolio