Full-stack engineer, applied AI. I build products end to end — the model, the API, the interface, the installer, and the release that puts it on someone's machine.

Right now that means computer vision that runs on-device, marketplaces that move real money, and native macOS tools. Computer Science student at Lovely Professional University, Punjab, graduating 2028. Open to software engineering internships.

Python YOLOv8 FastAPI WebSockets OpenCV PyInstaller

Watches any number of cameras — webcam, phone, or RTSP CCTV — and flags phones in real time with a cropped photo and location, for a human to confirm or dismiss. All inference runs on-device; video never leaves the building, which is what makes it deployable in a university at all.

Shipped as a native desktop app for macOS, Windows and Linux across 29 releases, built and signed in CI. Includes a fine-tuning pipeline so an institution can sharpen the model on its own footage, a searchable evidence log with CSV export, role-based accounts, and a command palette.

→ Source · Download it · Website

React TypeScript Supabase/Postgres Stripe Connect Vercel

A live commercial marketplace built on Stripe Connect. ~26,000 lines of TypeScript over 460 commits, 24 edge functions and 178 database migrations.

The part I would want to talk about in an interview is not the UI — it is the guardrails. Every database change goes through 13 CI workflows including migration linting, a strict schema audit, an automated review of edge functions, and smoke tests against a preview deployment. When your code moves other people's money, "it works locally" is not a standard.

→ vouchroot.com · source private (commercial)

React 19 TypeScript Supabase Row-Level Security GitHub Actions

A creator sets a price to receive a message. A visitor pays to open a conversation, the money is held until the creator actually replies inside their promised window, and it refunds automatically if they don't.

The interesting constraint was authorization: with no server of my own, correctness has to live in Postgres row-level security policies rather than in application code that a client could route around. Every number shown on a creator's profile is computed by a database view — none of it is hand-set.

→ Live demo

Status: the full loop works; Stripe is not wired in yet and escrow is simulated. The README says so too.

Swift AppKit Vision OCR Accessibility API

Copy a table out of a PDF. Paste it into Numbers and you get real cells; into Slack, readable lines; into a code editor, an array literal. One copy, three correct outputs, and you never picked a format.

Nine global bindings — OCR any region of the screen, pull the value a field is asking for, spread one record across a whole form, build a copy stack and pop it field by field. Hotkeys are validated against both the app's own bindings and the system's, so a shortcut can never silently shadow another. ⌘V is deliberately untouchable — no feature is worth adding latency to the most-pressed shortcut in computing.

Status: early — it builds and runs, and the README is explicit about which paths are not yet exercised against real apps.

Swift 6 ScreenCaptureKit Accessibility API

Hover a Dock icon and see live thumbnails of that app's actual windows, then click straight through to one — the Windows taskbar behaviour, done natively. A background cache warmer keeps thumbnails current so the strip appears instantly instead of rendering under the cursor.

React TypeScript i18n offline-first

Customers, measurements per garment, orders with advance/balance, due dates and photos. Fully bilingual English / ਪੰਜਾਬੀ, and entirely offline — no account, no server, no internet.

Built for a real shop, which set every constraint: the person using it does not read English comfortably, has patchy connectivity, and will not create an account. Those constraints made it a better piece of software than any of my own preferences would have.

→ Try it in your browser — no install, no account

JavaScript operating systems simulation

Classic schedulers decide once: a process is admitted, gets a slice, and keeps it. AdaptRes runs a feedback loop instead — it measures CPU and memory pressure every tick and, when CPU crosses 88% or memory 83%, throttles the lowest-priority running process until the bottleneck clears. Nothing is killed and high-priority work is never touched; a throttled process recovers on its own once pressure drops.

Live Gantt chart, utilisation graphs, starvation detection and CSV metrics export — and an off switch for adaptive mode, so you can run the same load with static allocation and watch it degrade.

→ Run it in your browser — no install

C++17

BFS, DFS, Dijkstra with path reconstruction, Prim's MST, cycle detection for directed and undirected graphs, and connected components — on a weighted adjacency list, in a menu-driven terminal program.

Languages — Python · TypeScript · Swift · C++ · SQL

AI / CV — YOLOv8 · Ultralytics · OpenCV · PyTorch · model fine-tuning · on-device inference · vision-language model fallbacks

Backend — FastAPI · WebSockets · Supabase · PostgreSQL · row-level security · Stripe Connect · REST

Frontend — React 19 · Vite · TailwindCSS · i18n · offline-first

Native — Swift 6 · AppKit · ScreenCaptureKit · Vision · Accessibility API

Infrastructure — GitHub Actions · Vercel · Docker · PyInstaller · signed multi-platform releases

• Ship the whole thing. A model in a notebook is not a product. Vigil is not finished because detection works — it is finished because someone can download a .dmg, drag it to Applications, and use it without a terminal.
• Put the guardrails in CI. Reviews catch what you thought to look for; pipelines catch what you didn't.
• Be honest in the README. Two of the projects above say plainly which parts are not finished. I would rather you trust the rest.

• Email — singhparamjot077s@gmail.com
• Live work — vouchroot.com · Vigil
• Location — Punjab, India