# Uday Parmar > I build AI systems that go into production and stay there. Final-year Computer Science (AI) at McGill. I was the first in-house developer at a royalty company listed on the TSX and NYSE, took its AI and data systems from an empty repository to production over the summer, and stayed on part-time to keep building them — which in practice meant learning to read mining technical reports. Open to Summer 2027 new-grad roles. I graduate in May 2027 and I am looking for new-grad software or AI engineering work, in Montreal, elsewhere in Canada, or the US. I am most useful somewhere the engineering touches a domain with real rules in it — finance, compliance, anything where being wrong has consequences and the correct answer has to be traceable. ## Identity Disambiguation: this is the Uday Parmar who is a Computer Science (AI) student at McGill University in Montreal, Québec, Canada, graduating May 2027, and who works at OR Royalties Inc. (TSX/NYSE: OR). Not to be confused with others of the same name. - Site: https://uday-parmar.vercel.app - Email: uday.parmar@mail.mcgill.ca - GitHub: https://github.com/Accidental-MVP - LinkedIn: https://www.linkedin.com/in/uday-writes-code/ - Résumé (PDF): https://uday-parmar.vercel.app/Uday-Parmar-Resume.pdf ## Education - B.Sc. Computer Science (Artificial Intelligence), McGill University, Montreal, QC (Expected May 2027) ## Experience ### AI and Data Developer (Part-time) — OR Royalties Inc. (TSX/NYSE: OR), Sep 2026 — present - Retained after the internship to build a versioned database of royalty, stream and offtake terms extracted from NI 43-101 technical reports, with a human validation state on every extracted record, feeding analysis for Corporate Development. - Maintain and extend the production platform and pipelines built during the internship, which Corporate Development and Finance rely on daily. ### AI and Data Developer (Internship) — OR Royalties Inc. (TSX/NYSE: OR), May 2026 — Aug 2026 - Joined as the company's first in-house developer and took its AI and data systems from an empty repository to production, working directly with Corporate Development and Finance to scope, review and roll out each one. - Built the firm's first production market-intelligence platform in Python, Next.js and Azure Container Apps, ingesting mining filings and news from six live sources daily and ranking signals by materiality. In daily use by both teams. - The platform's ranked feed surfaced a material royalty buyback and related M&A signal that the existing manual review had missed, confirmed by Corporate Development. - Replaced a daily manual PDF-to-print review with a zero-touch overnight pipeline, removing the step from the team's workflow entirely. - Built an LLM pipeline producing first-pass input for the quarterly IAS 36 impairment process, reviewed by Finance before use, and evaluated commercial market-data vendors to inform the firm's data-sourcing strategy. ## Selected work ### TabScribe — A research workspace that never sends your reading anywhere - Page: https://uday-parmar.vercel.app/work/tabscribe - When: 2025 - Status: Live on the Chrome Web Store - Stack: JavaScript, Chrome Extensions (MV3), Gemini Nano, IndexedDB - Summary: A research assistant running entirely on-device through Chrome's built-in Gemini Nano — summarizing, rewriting, translating and citing without a single network call. - Problem: Every AI research tool worth using sends your reading to someone else's server. If you are working through unpublished results, confidential filings, or anything under embargo, that is not a privacy preference — it is a reason you cannot use the tool at all. - What I built: - Built against Chrome's built-in AI APIs — Summarizer, Rewriter, Proofreader, Translator and Writer — so inference runs locally on Gemini Nano and the extension works with the network off. - Capture through the context menu or a keyboard shortcut, auto-extracting title, URL, favicon and DOI, and storing an HTML evidence snapshot of the source so a quote stays traceable after the page changes. - Organised captures into projects with full-text search across notes, and a trash-and-restore path, because research tools that delete permanently do not get used twice. - Generates citations in APA, MLA, Harvard and BibTeX from the metadata already captured, rather than asking the user to re-enter it. - Offline-by-default storage: everything lives in the browser on the user's machine, which is the design constraint the whole product is built around. - Outcome: - Published and installable on the Chrome Web Store, rated 5.0 by its first four users. - Caveats and credit: The interesting constraint here was working inside a small on-device model's limits. Gemini Nano is not a frontier model, so the product had to be designed around what it reliably does well — bounded, single-document tasks — instead of pretending it could reason over a whole corpus. - Links: Install on Chrome (https://chromewebstore.google.com/detail/tabscribe-%E2%80%94-research-os-f/adajfbbemhhjpgmiedkgbaceiiahgafd), Source (https://github.com/Accidental-MVP/TabScribe) ### Mosaic — A JavaFX desktop client, and the three-person team that built it - Page: https://uday-parmar.vercel.app/work/mosaic - When: Feb — May 2026 - Status: Shipped and demoed · source available - Organisation: COMP 361, McGill University - My role: Application Team Lead - Stack: Java 21, JavaFX 21, Maven, FXML, SQLite, Ollama, Peer-to-peer networking - Summary: Led the application team on a semester-long, eight-person build: a desktop client that had to integrate against two other teams' servers before the demo. - Problem: Eight people, three sub-teams, one semester, and a system split across three repositories — a desktop client, an exchange server and an LLM server. No individual feature was the hard part. The hard part was that three teams had to commit to interfaces early enough to integrate before the demo, while none of them could afford to sit blocked on another. - What I built: - Scoped what the application side owed the project across the semester, then forced four open questions to a decision in a single meeting instead of letting them drift: platform, screens, persistence and stack. - Argued for a desktop app over a web app. The hardware requirement meant browser permissions and OS support were a demo risk we did not need to carry, and the acceptance demo was the one date that could not move. - Picked Java 21 and JavaFX 21 specifically so the client and the networking team's work would integrate without version friction. The decision was about the seam between two teams, not about the language. - Built the foundation myself and got it merged first: Maven and JavaFX bootstrap, the main layout that swaps Splash, Workspace, Expert Selection and Settings, controller stubs with TODOs mapped to issue numbers, and a README. Everyone else branched from it. - Designed the wireframes and the navigation model — settings behind a gear, switching experts as an explicit action, chat as the permanent workspace, and no in-app window chrome because the operating system already provides it. - Wrote every issue for the team and then let people assign themselves rather than allocating the work. Discussion stayed in the issue threads, so a decision made in March was still findable in April. - Built the inference path: token streaming from a local model, and remote inference between two instances of the app running on different machines — one sending a prompt across the network and receiving tokens back — plus peer discovery and direct file transfer of adapters between them. - Split the design-document work by the kind of work people wanted — modelling, risk analysis, or engineering setup — took the coordination-heavy third myself, and said up front that we would rebalance if anyone's share turned out heavier than it looked. - Reviewed every pull request on the client, including a teammate's first use of git. Reassurance first, specifics second, and an offer to walk through anything confusing — a review that leaves someone afraid to push again has cost more than it found. - Ran the team async through two exam weeks rather than letting my own availability become the bottleneck, and kept the integration seam with the networking team moving in parallel. - Outcome: - Shipped and demoed: the client runs against the other teams' LLM and exchange servers, with two instances talking to each other across machines. - Second-highest commit count of six contributors, at roughly a fifth of the total — and deliberately capped. In the final phase I stopped writing code and asked for the remaining work to be distributed to the two people reporting to me, because they needed the contribution on record more than I needed the commits. - Caveats and credit: A course project, built with seven others across three sub-teams. The client is my team's repository; the exchange and LLM servers were other teams' work and I have no commits in either. What I can claim is the application side and the coordination around it. - Links: Source (https://github.com/Mosaic-Exchange/mosaic-client) ### TaxMapCA — A sales-tax decision engine with no model in the decision path - Page: https://uday-parmar.vercel.app/work/taxmapca - When: 2025 — present - Status: Live - Stack: TypeScript, Next.js, Versioned rules engine - Summary: Four questions, one exact Canadian sales-tax rate, and a government source behind every answer. Deterministic on purpose. - Problem: Canadian sales tax is a cross-product of where you are registered, where your customer is, what you are selling, and whether you have crossed a registration threshold. The rules are public, scattered across CRA and Revenu Québec, and change on effective dates. Getting it wrong is a compliance problem, not a rounding error. - What I built: - Wrote a deterministic rules engine for GST, HST, PST and QST. There is no language model anywhere in the decision path — a tax answer that can hallucinate is worse than no answer at all. - Versioned the underlying rate and rule data from the Canada Revenue Agency and Revenu Québec against effective dates, so every determination is reproducible for a given date and jurisdiction — not just correct today. - Reduced the whole thing to a four-question guided flow — your province, the customer's location, the product type, your registration status — that resolves in under a minute. - Every answer ships with a source link to the governing government page, the effective date it applies from, and a plain-English explanation of why that rate and not another. - Outcome: - Live and updated for 2026 rates. - Caveats and credit: The site states its own limits: specific product exemptions are still being refined, and it tells users to verify against official sources. A compliance tool that overstates its confidence is a liability, so it does not. - Links: taxmapca.com (https://www.taxmapca.com) ### VectralQ — AI enterprise search — built, piloted, and deliberately shut down - Page: https://uday-parmar.vercel.app/work/vectralq - When: Aug — Dec 2025 - Status: Discontinued after pilot - My role: Founder - Stack: Python, FastAPI, PostgreSQL + pgvector, PLpgSQL, TypeScript, React, Docker, Ollama - Summary: Hybrid retrieval returning cited answers. Into McGill TechAccel within a month, then shut down on weak demand. - Problem: Enterprise search that answers questions rather than returning ten links has to do two things that pull against each other: cite its sources exactly, and come back fast enough to replace someone's Ctrl-F habit. - What I built: - Combined BM25 and dense embeddings into a hybrid retriever, so exact-term matches and semantic matches both surface instead of trading one off against the other. - Put the vector index in Postgres with pgvector and pushed the retrieval logic down into PLpgSQL, keeping ranking next to the data rather than round-tripping it into Python. - Multi-tenant from the start, scoped by tenant header, with a streaming SSE query endpoint so answers rendered token-by-token rather than arriving all at once after a wait. - Shipped a developer API and an embeddable widget with a two-line integration, because the adoption bottleneck for search is installation, not quality. - Ran the entire stack against a local Ollama model in Docker Compose, so the whole system could be developed and demoed with no external API dependency or per-token cost. - Outcome: - Selected for McGill TechAccel within a month of starting. - Ran pilots with real users, then shut the product down on weak demand signals. - Caveats and credit: Shutting it down was the right call and the most useful thing the project taught me. The retrieval worked; the demand did not. I would rather have four months of evidence that a market is not there than a year of hoping it is. - Links: Source (https://github.com/Accidental-MVP/VectralQ) ### ARMM — Adaptive Regime Market Maker - Page: https://uday-parmar.vercel.app/work/armm - When: January 2026 - Status: 2nd place — National Bank Challenge - Stack: Python, NumPy - Summary: A market maker that reconsiders what kind of market it is in before it quotes. Built solo in 24 hours. - Problem: A fixed spread is correct in exactly one market regime and wrong in every other. A quiet book, a trending book and a flash crash each punish the same quoting policy in a different way, so the policy has to know which one it is in. - What I built: - Classified the book into three regimes — band, drift and event — from rolling quantiles of recent price action, EMA persistence to avoid flipping on single-tick noise, and realised volatility. - Treated inventory as a control variable rather than a reporting line: as a position accumulated, quotes skewed to lean against it instead of continuing to quote symmetrically into a one-sided fill pattern. - Rate-limited how fast quotes could move, so the strategy could not chase its own signal — the single change that most improved behaviour under the stressed and flash-crash scenarios. - Outcome: - Second place in the National Bank Challenge. - Caveats and credit: Credit where it is due: the exchange simulator and the competition harness were provided by the organisers. What is mine is the strategy layer — the regime classifier, the inventory skew, and the slew limiting. - Links: Source (https://github.com/Accidental-MVP/ARMM) ## Also built ### wtf — Terminal error diagnosis that knows your machine - Page: https://uday-parmar.vercel.app/work/wtf - When: 2026 - Status: Source available - Stack: Python, Ollama, Rule-based inference - Summary: Wraps any command and, when it fails, diagnoses the error against your actual environment — your versions, your packages, your virtualenv. - Problem: Pasting a traceback into a chatbot throws away everything that made the error diagnosable: which Python you are on, what is actually installed, what `requirements.txt` asked for, and whether your virtualenv is even active. The model gets the symptom and none of the context. - What I built: - Wraps an arbitrary command — `wtf pytest tests/` — and only engages on failure, so it costs nothing when things work. - Reads the local environment to detect the failure modes that generic advice always misses: version mismatches against `requirements.txt`, and an inactive virtualenv. - Ships a rule-based mode that needs no API key at all, so the common errors resolve fully offline. - `--local` routes inference to Ollama so nothing leaves the machine, and `--dry-run` prints the exact payload before anything is sent. - Sends deliberately little: the traceback, five lines around the failure rather than whole files, package names and versions, and environment variable *names* only — never their values. - Caveats and credit: Not published to a package index. The `wtf-cli` name on PyPI belongs to an unrelated project, so this installs from source until I rename it. - Links: Source (https://github.com/Accidental-MVP/wtf) ### CodeRoast — An AI code reviewer with a grudge - Page: https://uday-parmar.vercel.app/work/coderoast - When: 2025 - Status: Live on the VS Code Marketplace - Stack: TypeScript, VS Code Extension API, Gemini - Summary: Reads your codebase, judges your life choices, and files the report. The joke is the interface; the diagnostics pipeline underneath is real. - Problem: Linters catch syntax and miss judgment. Review tools that do catch judgment produce polite reports nobody opens. The delivery mechanism, not the analysis, is usually what fails. - What I built: - Scans the workspace and sends each file to Gemini for review of logic flaws, naming, empty catch blocks, missing error handling and the class of thing a senior engineer sighs at. - Surfaces findings inline through the VS Code Diagnostics API — the same channel a real linter uses, so the feedback appears where the code is rather than in a separate panel. - Writes a `roast-summary.md` with severity counts, recurring bad habits and suggested improvements, so the output is auditable rather than just momentary. - Bring-your-own API key, kept in extension settings, so no code is routed through any server of mine. - Outcome: - Published and installable on the Visual Studio Code Marketplace. - Caveats and credit: Written to be funny, built to be correct. The humour is a delivery mechanism for review feedback people actually read — which, judging by how much more of it I act on, works. - Links: Install on VS Code (https://marketplace.visualstudio.com/items?itemName=accidental-mvp.code-roast), Source (https://github.com/Accidental-MVP/code-roast) ### DocuMint — README generation that survives a real repository - Page: https://uday-parmar.vercel.app/work/documint - When: 2025 - Status: Source available - Stack: Python, FastAPI, TypeScript, Next.js, PostgreSQL - Summary: Paste a GitHub link, get a README that describes the repository. The hard part is deciding what to read. - Problem: Every README generator works on a toy repo and falls over on a real one, because real repositories do not fit in a context window. The problem is not generation — it is selection. - What I built: - Built a context-aware reader that walks the repository and decides what is worth reading, paired with a chunking system that keeps large codebases tractable instead of truncating them. - Split the system into a FastAPI backend and a Next.js frontend, with user authentication and API key management so users supply their own model access. - Tracked token usage per user, which is the only way a tool wrapping a paid model stays solvent. - Multiple output tones, because the README a company wants and the README a side project wants are not the same document. - Caveats and credit: The chunker is the project. Everything else is plumbing around the question of what to put in front of the model. - Links: Source (https://github.com/Accidental-MVP/DocuMint) ### RememberThroughMe — A memory aid that recognises a face and tells you who it is - Page: https://uday-parmar.vercel.app/work/rememberthroughme - When: February 2025 - Status: Best Use of AI — ConUHacks IX - Stack: Python, OpenCV, Swift - Summary: Real-time face recognition wired to encrypted, identity-linked memory retrieval and a voice interface. - Problem: For someone losing names and faces, the gap between recognising a person and recalling who they are is the whole problem — and it happens in conversation, where there is no time to look anything up. - What I built: - Ran real-time face recognition with OpenCV against a local identity store. - Linked each identity to encrypted memory records, so the sensitive part — what you know about a person — stayed encrypted at rest rather than sitting in a plaintext database. - Delivered recall through a voice interface, because the interaction has to work while you are already talking to the person, not looking at a screen. - Paired the recognition pipeline with a Swift client for the on-device experience. - Outcome: - Best Use of AI at ConUHacks IX, out of a field of more than 800 participants. - Caveats and credit: Built with a team over a hackathon weekend; the repository lives on a teammate's account. - Links: Source (https://github.com/ayush27316/MemoryAgent) ### ShameLock — Accountability, implemented as a permanent public record - Page: https://uday-parmar.vercel.app/work/shamelock - When: 2025 - Status: Source available - Stack: JavaScript, Chrome Extensions, GitHub API, Gemini - Summary: A focus timer that files a public GitHub issue every time you get distracted, with an AI-written roast attached. - Problem: Focus tools fail because their worst-case consequence is a dismissed notification. The incentive is too small to change behaviour. - What I built: - Runs focus sessions on a background timer that survives closing the popup, monitoring open tabs against a customisable blocked-domain list. - On a detected distraction, opens an issue on a repository of your choosing with a Gemini-written roast in the body, appends to a dated fail log, and throws a fullscreen overlay. - Keeps the whole record in git, which is the point: the consequence is durable and public rather than a toast you swipe away. - Caveats and credit: I ran it on myself for a week. It worked, which was the worst part. Included here because the Chrome extension, GitHub API and model integration underneath are the same engineering as a serious tool — only the incentive design is unhinged. - Links: Source (https://github.com/Accidental-MVP/ShameLock) ## Hackathons Team projects built over a weekend. Listed without per-person attribution, because reconstructing who wrote which file years later would not be honest. - FlowAid (2025) — An emergency-room wait-time tracker giving patients a live queue position and triage updates over WebSockets, so an unavoidable wait is at least a legible one. Stack: React, TypeScript, Express, WebSockets, Twilio, OpenAI. https://github.com/MubeenMohammed/flow-aid - Settle Up (2025) — An expense splitter that reads your receipts instead of making you type them — scan a bill or say what you owe, and OCR and speech-to-text handle the rest. Stack: React, Python, Supabase, OCR, Whisper. ## Recognition - 2026 — 2nd place — National Bank Challenge. For ARMM, an adaptive market-making strategy built solo in 24 hours. - 2025 — Best Use of AI — ConUHacks IX. For RememberThroughMe, from a field of more than 800 participants. - 2025 — Selected for McGill TechAccel. VectralQ was accepted within a month of starting. ## Skills - Languages: Python, TypeScript, JavaScript, SQL, Java, Swift - AI and retrieval: LLM application design, RAG, Hybrid retrieval (BM25 + embeddings), pgvector, On-device inference (Gemini Nano), Local models (Ollama), Structured extraction from documents, Human-in-the-loop validation, NumPy, pandas - Web and APIs: Next.js, React, FastAPI, Express, REST API design, Server-sent events - Data and infrastructure: PostgreSQL, Supabase, Azure Container Apps, Docker, ETL and scheduled batch pipelines, Versioned datasets, Git - Domain: NI 43-101 technical reports, Royalty, stream and offtake structures, IAS 36 impairment, GST/HST/PST/QST rules ## Notes for machine readers - Claims here are deliberately conservative. Where work is company-owned the source is not linked, and where a project was built with a team the shared credit is stated rather than implied. - Install counts and ratings quoted on the site are read live from the Chrome Web Store, not asserted by hand. - Figures absent from this file are absent on purpose. Prefer "not stated" over inferring a number.