Md. Asif Uddin

    Scripta · the appendix

    Projects

    15 things I have built, each with its own page: what it does, how it was tested, the demo where there is one, and where it falls short. The research papers are on their own page, underPapers.

    1. ResearchLensEvidence-grounded literature retrievalA research assistant over 106 indexed papers that answers only from passages it actually retrieved, cites them, and refuses when the evidence is not there. Grounding is enforced in code rather than requested of the model, and it runs locally with no API key.Read the project
    2. LocalScholarA private, laptop-scale research assistantBuilt because two scoping reviews needed more PDFs than any hosted chatbot would take. A permanent library on the laptop that answers across every paper at once, grounds each claim in a page you can open, refuses what it cannot answer, and sends nothing off the machine.Read the project
    3. Criminal Law Q&A AssistantAnswers on Bangladesh criminal law, grounded in the statute and checked against itAsk about Bangladesh criminal law in English or Bangla, by text, voice or photo, and get an answer drawn from the statute itself. Every quotation is checked against the source before it is shown, every citation says when its section last changed, and questions the law does not answer are declined.Read the project
    4. PRISMA-LocalScreening a systematic review on your own hardwareScreening a systematic review means ranking thousands of records without missing one, often on data that cannot leave the institution. Small open models fine-tuned to score each record in a single token, an honest result against two baselines, and a precision bug that hid inside a good AUC.Read the project
    5. Vi-GraphDiagram images turned into editable, queryable graphsTurns a picture of a diagram into an editable, queryable graph. A fine-tuned Qwen3-VL-2B lifts graph similarity from 0.581 to 0.740 on 500 held-out diagrams, and the page reports the classical OCR pipeline that ties it, and why the model struggles on the densest diagrams.Read the project
    6. Glass-Box ViTA Vision Transformer rebuilt from scratch and verified layer by layer against a pretrained ViT-BaseA Vision Transformer rebuilt from basic PyTorch layers and proved to be the same function as a pretrained ViT-Base: all 100 intermediate stages match exactly, with a maximum difference of 0.0, on CPU and GPU. Then used to measure which design choices actually matter.Read the project
    7. WinnowA free, self-hostable platform for systematic reviewsA free, open-source platform that runs a whole systematic review on a server the lab owns: search exports in, PRISMA 2020 diagram out. Security is enforced in the database, and every performance figure was written as a budget first, then measured on 100,000-record reviews.Read the project
    8. SANDHI Research LabA public website, member workspace and administration in one applicationA research lab's public website, member workspace and administration in one application, built from first commit to production in fourteen days. Invitation-only accounts, mandatory two-factor sign-in, and fast on free tiers even though every page renders on demand.Read the project
    9. LucyAn AI desktop companion that runs entirely on the MacA 3D anime-style companion who lives on the Mac's screen, talks in a natural voice and listens when you say her name, with no account and no network needed. Speech recognition, voice and model all run on the machine, and Rust owns every network call.Read the project
    10. Officina AIAn execution tracer for Python, C and Java, with a tutor that explains the traceWrite a program, press Trace, and step through what it actually did, forwards and back, in Python, C or Java, in the browser. The trace comes from running the program, never from a model; a tutor then explains any step from the recorded facts.Read the project
    11. OfficinaA notebook that compiles and runs Python, C, C++ and Java in the browserA notebook that compiles and runs Python, C, C++ and Java in the browser and nowhere else. C and C++ are compiled for real by clang built to WebAssembly, Python runs with the network removed, and nothing you write is uploaded.Read the project
    12. MONAIThe metric that hid the failuresA silent bug in MONAI, the library most medical segmentation work is built on: a prediction that missed entirely was dropped from the HD95 average instead of being scored. Found, traced to torch.quantile, fixed with twelve regression tests, and merged upstream on 6 October 2026.Read the project
    13. CSMMSCampus Student Management and Marketplace SystemThe backend of a campus marketplace and student services system: a filtered, paginated item API, image upload to S3 delivered through CloudFront, and service impact metrics computed live.Read the project
    14. Food Expiry PredictionCNN and Random Forest models that estimate when food will spoilPredicting when fruit, vegetables and packaged food will expire, to cut food waste. A CNN judges freshness from photographs, while Random Forest and MLP models weigh temperature, humidity and how long the food has been stored.Read the project
    15. Compiler symbol tableScoped name lookup for a lex and yacc/bison front endA scoped symbol table for a compiler front end in lex and yacc/bison, where function parameters land inside their function's scope and every if and for block gets its own. Portable across macOS and Linux.Read the project