MIT · Computer Science & Engineering
I like to train small language models on modest hardware when free, and spend the rest of my time on full-stack builds, ML tooling, and eval infrastructure.
- CTO at Sciloop. Sourcing and curating the STEM data that moves the needle on frontier model training
- Pretraining small encoder/language models from scratch under hard parameter and GPU-hour budgets
- Writing MCP servers and multi-agent pipelines to make LLM-assisted engineering less painful
- Open to talking with labs and teams who need better STEM data for training. Reach me at osmansiddiquer@gmail.com
| benchtop-bert |
A BERT-style encoder written from scratch: attention, blocks, embeddings, MLM head, all hand-rolled,
pretrained on a single 4 GB laptop GPU under a hard 30M-parameter cap.
27.9M params, 3B tokens, 35.4 GPU-hours, 85–86% SST-2 (vs. 92.7% for BERT-base at 4× the size).
Ships with a full technical report and a reproducible checkpoint lineage.
Python · PyTorch · from-scratch transformer |
| Fluxion |
A Desmos-like grapher built specifically for ODEs: it types, solves, and plots first- and higher-order
equations, coupled systems, implicit and polar curves, then lets you inspect extrema and intersections.
Live →
TypeScript · numerical methods |
| LitFinder |
A modular pipeline that takes a natural-language query, generates search criteria, pulls open-access
papers from Semantic Scholar, chunks and embeds them, then reranks for relevance, built to source
STEM literature for Sciloop's data pipeline.
Python · Gemini · FAISS · VoyageAI |
| 2D Inverse Kinematics |
Real-time inverse kinematics for a two-link robotic arm built for a friend's hardware project:
solves joint angles from a target x/y coordinate with both motors constrained to the base.
JavaScript · robotics |
| Pong |
My earliest project still online: a console Pong built in plain C against the Windows console API,
focused on game physics and paddle/ball control. Years later I compiled it to WebAssembly and
embedded it in my portfolio, running live on a WebGL terminal.
C · Windows console API |
| Web Development Progress |
Where I learned HTML, CSS, and JS from scratch: a growing site built page by page with the MDN
docs open in the other tab, kept online as a record of that first pass at web fundamentals.
Live →
HTML · CSS · JavaScript |
Thanks for stopping by. Open an issue on any repo above if you want to talk shop.