Skip to content
View Accidental-MVP's full-sized avatar

Highlights

  • Pro

Organizations

@Mosaic-Exchange

Block or report Accidental-MVP

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Accidental-MVP/README.md
Uday Parmar — I build AI systems that go into production and stay there.

Portfolio Résumé LinkedIn Email Open to Summer 2027

Final-year Computer Science (AI) at McGill, graduating May 2027. I joined a royalty company listed on the TSX and NYSE as its first in-house developer and took its AI and data systems from an empty repository into production over one summer — which in practice meant learning to read mining technical reports.

I'm most useful somewhere the engineering touches a domain with real rules in it: finance, compliance, anything where being wrong has consequences and the right answer has to be traceable.


Things you can open right now

uday-parmar.vercel.app

Portfolio
Ten projects, each with the problem, the architecture, and what I'd change. Next.js 16, 19 static routes.
taxmapca.com

TaxMapCA  Live
Four questions, one exact Canadian sales-tax rate, a government source behind every answer. No model in the decision path — a tax answer that can hallucinate is worse than none.
TabScribe on the Chrome Web Store

TabScribe  Users Rating
Research workspace running entirely on-device through Chrome's Gemini Nano. Summarises, rewrites, translates and cites with the network off.
CodeRoast on the VS Code Marketplace

CodeRoast  VS Code
AI code review delivered through the editor's own diagnostics pipeline. Linters catch syntax and miss judgment; tools that catch judgment write reports nobody opens.

The install counts and ratings above are live endpoints — they read from the stores, not from me.


Things worth reading the code of

VectralQ — enterprise search on Postgres instead of a vector database. Three retrieval lanes — weighted BM25, pgvector, and a phrase lane scored separately from bag-of-words — fused with Reciprocal Rank Fusion over ranks rather than scores, then reranked by a cross-encoder. Tenant isolation enforced by Postgres row-level security with FORCE: a query that forgets the tenant returns nothing instead of returning someone else's documents.

Into McGill TechAccel within a month, piloted, shut down on weak demand. The retrieval worked; the market didn't. Killing it was the right call.

ARMM — a market maker that decides what kind of market it's in before it decides how to quote. BAND/DRIFT/EVENT classification from rolling quantiles, EMA persistence and robust volatility, with hysteresis and an event cooldown so it can't whipsaw itself. Inventory is a steered control variable, slew-rate-limited so the strategy can't chase its own signal. Solo in 24 hours — 2nd place, National Bank Challenge.

wtf — diagnoses terminal errors against your actual machine: your Python, your installed packages, your requirements.txt, whether your virtualenv is even active. Pasting a traceback into a chatbot throws all of that away. Runs fully offline with no API key.

DocuMint — README generation that survives a real repository. The chunker and context-aware reader are the project; everything else is plumbing around deciding what to put in front of the model.


Stack

Python TypeScript FastAPI Next.js PostgreSQL Azure Docker


I also built a Chrome extension that files a public GitHub issue every time I get distracted. It worked, which was the worst part.

Pinned Loading

  1. TabScribe TabScribe Public

    Offline-first Chrome extension that captures web snippets and turns them into cited research drafts, running entirely on Chrome's built-in Gemini Nano.

    JavaScript

  2. VectralQ VectralQ Public

    Enterprise search built on Postgres instead of a vector database: RRF fusion over BM25, vector and phrase lanes, cross-encoder reranking, and tenant isolation enforced by row-level security.

    PLpgSQL

  3. ARMM ARMM Public

    Forked from mattcauf/NBC_hackathon

    Market maker that classifies the book into regimes before it quotes, treating inventory as a steered control variable. Built solo in 24 hours; 2nd place, National Bank Challenge.

    Python

  4. wtf wtf Public

    Wraps any command and, when it fails, diagnoses the error against your actual environment — your versions, your packages, your virtualenv. Works offline with no API key.

    Python

  5. DocuMint DocuMint Public

    AI README generator that survives a real repository. A context-aware reader and chunker decide what to put in front of the model when the codebase does not fit the context window.

    Python

  6. code-roast code-roast Public

    VS Code extension that reviews your code with Gemini and reports findings through the editor's own diagnostics pipeline. The humour is the delivery mechanism.

    TypeScript