I'm Siyam — a data science student at COMSATS. I picked data science over AI, cybersecurity,
and computer science, not as a default but because it was the one that actually pulled at me.
A course assignment is what did it — the first time I built something and watched it actually
work on real data, not a toy example. I've been chasing that feeling since.
Off the laptop, I bowl. We finished third in the FBISE inter-college tournament against more
than 50 colleges — I'd like to think my spell had something to do with it.
COMSATS University Islamabad · data science · currently small-n, high-curiosity
Identity
Data × curiosity, n = 1
Status
the outlier, on purpose
hi. statistically significant at p < 0.05 — still doesn't fit in.
kind of my whole thing.
Stack
Python, mostly. Still expanding.
Live from GitHub
Language distribution — across public repos
fetching from the GitHub API…
Contribution activity — 2026 · — contributions
loading contribution graph…
Fetched live from github-contributions-api.jogruber.de — no widget, no auth token.
Each cell is one day; vermilion intensity scales with contribution count.
01 · production
NexDist
Problem A Pakistani medicine distributor was running sales, inventory, and ledgers through a legacy Oracle Forms system with no clean path off it.
Approach Built a mobile-first SaaS replacement — invoicing, inventory, customer ledgers, cash recovery — plus an Excel-based migration tool to move the first client's data over.
Result First client fully migrated off the legacy system: a PWA for field sales teams, a desktop view for office staff. Live since Feb 2026.
Next.js 15 · TypeScript · Supabase · Tailwind CSS · Vercel
private repository — client work, not public
02 · prototype
Search-Light
Problem Finding something specific in surveillance footage normally means watching it manually, or having pre-labeled training data — neither scales.
Approach Natural-language queries matched against CLIP (ViT-B-32) embeddings of sampled video frames, indexed in FAISS, ranked by cosine similarity. No labels needed.
Result 80% accuracy across a 10-query human-verified benchmark, ~0.33s average retrieval. Reliable on clear, large objects; documented to struggle on small objects under ~50px and on motion-direction queries — both named, with a real next step (SAHI tiling; a swap to VideoMAE or X-CLIP).
Python · Streamlit · CLIP (ViT-B-32) · FAISS · OpenCV
github →
03 · automated pipeline
Zameen-ML
Problem Pakistani property prices swing hyperlocally, but most listing platforms price off stale, manually-entered comps instead of continuously refreshed, location-aware data. (zameen — "land," in Urdu)
Approach A four-stage, fully decoupled pipeline: a defensively-built scraper pulls fresh Zameen.com listings weekly via a scheduled GitHub Action (rotating user-agents, Cloudflare-evasion); geopy + Nominatim resolve real coordinates per sector and compute Haversine distance to anchors like Islamabad's Blue Area; a regularized RandomForestRegressor trains on leakage-free splits, with log1p transforms to tame real estate's price skew; the trained model serves through a Streamlit app for sub-second inference.
Result 500 trees, depth-capped at 12, trained across 15 engineered features. The single biggest driver of predicted price isn't area or location alone — it's their interaction: area_x_loc carries 38% of the model's total predictive weight, more than double the next-closest feature. In Pakistani real estate, where you are multiplies what you have, it doesn't just add to it.
Python · scikit-learn (RandomForestRegressor) · geopy · GitHub Actions · Streamlit · joblib
github →
Drop a data point — a question, a project, a correction.
fbise inter-college — 50+ colleges, we finished 3rd.
my spell: wickets when it counted, runs nobody remembers me giving up.
— the outlier, on hover
