Skip to content

AI Developer / Toronto, Ontario

Samarpartap Passey

I build agentic systems that run in production.

MCP servers, evaluation harnesses, and voice agents. They call real tools, hold real credentials, and have to be right when nobody is watching. Python and TypeScript.

See the workGitHub
Samarpartap Passey, AI Developer

Selected work

Four things I built, and what they prove

All four are public. Two of them are live and serving right now, and every number below comes straight from the repository.

maple-procure

An MCP server over Canada's public procurement data

Five tools that let a model search open federal tenders and awarded contracts, over both stdio and streamable HTTP. Scoped bearer tokens, a structured log line for every call, and one write that cannot be made to happen on a single call.

Python · FastAPI · MCP · SQLite FTS5 · Render

Repositoryfor maple-procureLive servicefor maple-procure

Evidence

  • Deployed and callable. A read-only demo token is published in the README so anyone can try it
  • A clean clone rebuilds the database from the official government source in about five seconds
  • Every result carries the file it came from and when that file was ingested
  • Exports return a preview first. Writing needs a second call carrying a server-minted handle

ledger

A grounding eval for RAG, at the level of the individual claim

Splits an answer into separate claims and verifies each one against only the chunk it cites. The verifier never sees the question or the rest of the answer, so it cannot be led to the conclusion.

Python · Claude · RAG · Evals

Repositoryfor ledger

Evidence

  • Measured 81.4% claim-level grounding across 86 claims, with zero fabrications
  • Detects false refusals, a failure mode that citation metrics miss entirely
  • Verifier validated against 25 hand-labelled claims, so every figure it reports is a floor

all-quote

A quote agent that proves what it saw

One intake, 79 distinct Ontario rate sources, and an evidence-backed terminal status for every one. A planner decides, an executor acts, and human checkpoints are first-class states in the design rather than errors to recover from.

Python · Claude · Playwright

Repositoryfor all-quote

Evidence

  • Every artifact hashed with sha256 and tagged observed or derived, keeping what the agent saw separable from what it inferred
  • Driver profile held in an encrypted vault, with personal data redacted before any evidence is written
  • Safety rules written down as a design artifact, covering retry limits, blocked flows and payment boundaries
  • More than 30 recorded HTML fixtures, so the parsers are tested against real pages

Sage

A voice-first interview coach, live and in use

Real-time voice on Vapi, with Claude Sonnet 4.6 planning each session and scoring it afterwards. The security design is the interesting part: a voice agent hands a lot to the browser by default, and this one hands it almost nothing.

Next.js · TypeScript · Claude · Vapi · Deepgram · Supabase

Repositoryfor Sagesageinterview.cofor Sage

Evidence

  • The browser holds only a 120-second JWT, scoped to exactly one capability and locked to the request origin
  • Prompts are assembled server-side and never reach the client
  • Transcription primed per session with named entities pulled from the interview plan
  • Supabase behind row-level security, every policy scoping rows to their owner

About

A little about me

I'm Samar. I build products, I chase new technology the week it lands, and I ship the results.

Most of my time goes into experimenting. When something new arrives, a model, a protocol, a framework, I want to have built something real with it before the opinions have settled. That is how a curiosity about MCP became a deployed server running against live government data, and how an interest in retrieval became a working evaluation harness.

I build at hackathons whenever I get the chance. I shipped a Claude-powered app in a single evening at the 7shifts × Anthropic Build Night. I led backend architecture on a four-developer team that delivered a machinery scheduling system to a manufacturing client. And I taught a workshop on building an AI interview agent to a room of people who had never written a line of code.

The through line is that I like making things that actually run. Not prototypes, not slide decks. Things with real users, real credentials and real data behind them.

Based in Toronto, Ontario. Open to AI engineering roles in the Greater Toronto Area.

Languages
Python · TypeScript · C#
AI systems
MCP · Anthropic SDK · Agentic design · LLM evals · RAG · Prompt design
Backend
FastAPI · Node · SQLite · Supabase · REST
Frontend
Next.js · React · Tailwind
Practice
pytest · Playwright · Jest · Git · Vercel · Netlify

Outside work

When I'm not building

Two things reliably pull me away from a keyboard.

Sound

Albums, front to back

I listen properly rather than on shuffle, and I have opinions about it. Most of what is on this page was built with something playing.

Weekends

Cricket and football

Both with friends, most weekends, and competitive enough to argue about the result afterwards. It is the part of the week that happens away from a screen.

Contact

If any of this is the kind of thing you're building, I'd like to hear about it.

samarpartapp@gmail.com