Peter Yousefi

AI Engineer · GenAI & Agents · Toronto, ON

I build AI systems that show their work.

Computer Science graduate focused on GenAI and AI engineering. Builds agent and LLM systems where every run is recorded, model output is checked before it is trusted, and results are measured rather than claimed. AZ-900 certified, currently studying for the AI-103.

View workRésumé PDF
Availability
Open to GenAI and AI engineering roles · Toronto (UTC-5)
Currently
Currently: studying for the AI-103 (Azure AI Apps & Agents)

Profile

Peter Yousefi

AI Engineer — GenAI & Agents

Core skills
PythonTypeScriptLLM applicationsAzure
Certs
AZ-900 earned · AI-103 in progress
Region
Toronto, ON (UTC-5)
Availability
Open to GenAI and AI engineering roles

Work

Selected projects

Seven projects, each starting from a failure I wanted to prevent. Results are stated as measured, including the weak spots.

Featured

Agent Flight Recorder

An execution control plane for AI agents

Agent jobs can fail after queuing, spend budget before timing out, or run twice after a crash, and leave no record of why. This control plane stores every attempt, event and cost, so a failed run can be inspected, retried or replayed without losing its history.

Outcome
Retries with backoff, dead-letter handling and per-run budget limits, with tracing from API to worker. Backed by 195 automated tests across unit, database, API, emulator and browser layers.

TypeScriptFastifyReactPostgreSQLAzure Service BusOpenTelemetryDocker
Live demoGitHub

CrawlOps

Reliability checks for AI web-research agents

A research agent can return a successful response with an empty, malformed or poorly sourced answer. CrawlOps grades every run against a written contract using seven deterministic checks, then labels it success, partial or failed with a recorded reason.

Outcome
Failures are classified and retries are bounded. Reliability trends come from stored runs, so reviewing them costs zero API calls.

TypeScriptFastifyReactPostgreSQLPrismaDockerAzure Container Apps
Live demoGitHubDemo video

AgentDrift

Detection of risky AI-agent data movement

Actions that are each permitted can still add up to data exfiltration. AgentDrift combines statistical baselines with a 15-minute window that links reads, staging and sends. Every report cites real events, and a human approves any simulated response.

Outcome
On 380 held-out synthetic runs, precision rose from 45.5% to 71.4% and false alarms fell from 60% to 20%. Recall stayed at 55.6%, and the results are published as measured.

PythonFastAPIReactAzure Table Storagepytest
Live demoGitHub

AegisOps

Human-governed incident response

Pulls alerts, metrics, logs, deploy history, and runbooks into one place for a simulated Checkout API outage: produces an assessment citing real evidence, proposes a reversible fix, waits for human approval, simulates the action, and logs everything in an append-only audit trail.

Outcome
AI output is treated as untrusted and validated against strict schemas, with no remediation without explicit approval. Synthetic data only; Azure deployment is design-only (Bicep templates, nothing provisioned) and the AI provider is a deterministic mock by default.

Next.jsTypeScriptFastAPIPostgreSQLDockerGitHub ActionsAzure Bicep
GitHubAzure-ready design

Falsifier

Exoplanet candidate triage

Runs a vetting pipeline on any Kepler or TESS target — ingest, detrend, period search, seven vetting tests — returning candidate / false positive / ambiguous with a reason. A triage tool, not a detection claim.

Outcome
Every number shown traces to a committed data artifact, with tests enforcing it, and the README states plainly what is unmeasured or blocked.

PythonFastAPIReactAstropypytestGitHub Actions
Live appGitHubIBM AI Builders Challenge

Continuity

AI storyboard generator

Takes a story, style guide, and character sheet; uses GPT to split the story into scenes, then builds each image prompt from all three so the same character stays consistent across the sequence. Includes a toggle to regenerate scenes without the character sheet for comparison.

Outcome
A review step lets you edit prompts before spending image credits. A Vercel Hobby serverless timeout was fixed by moving image generation to an Edge Function with a 60-second timeout.

Next.js 14TypeScriptReactOpenAI APIVercel EdgeTailwind
Live appGitHubDemo video

Lumina

AI campaign generator

Turns a business description into a ready-to-publish marketing campaign in minutes, with every field owner-editable before going live.

Outcome
Output is owner-editable before publishing, with a fallback path for API downtime.

PythonFlaskREST APIsIBM watsonx
GitHubDemo videoIBM AI Builders Challenge

Stack

Tools I work in

Core
PythonTypeScriptJavaScriptSQLGit & GitHub
AI applied
LLM applicationsOpenAI APIAzure OpenAIIBM watsonxStructured-output validationEvaluation harnessesAgent observability
Backend & data
FastAPIFastifyFlaskNode.jsPostgreSQLPrismaSQLAlchemyPydanticSQL reporting & validation
Web
ReactNext.jsTailwind
Cloud & testing
AzureDockerGitHub ActionsOpenTelemetrypytestVitestPlaywright
Currently learning
Azure AI AI-103Bicep/TerraformLinux & Bash

Experience

Where I've worked

  1. Sep 2025 — Dec 2025

    Teaching Assistant

    Trent University · Oshawa, ON

    • Replaced plain right-or-wrong marking with comments that name the exact concept behind each mistake.
    • Ran one-on-one debugging sessions that taught students to isolate bugs themselves, and wrote study materials for the topics they found hardest.

  2. Apr 2022 — Sep 2022

    HRIS Specialist (Intern)

    Concordia University · Montreal, QC

    • Wrote joined SQL queries against the live HR system to produce the employee and payroll reports its built-in tools could not.
    • Checked data replication during an ERP modernization and corrected discrepancies before they reached payroll.

  3. Sep 2021 — Dec 2022

    VP of Social / Social Director

    SCS & ECA, Concordia · Montreal, QC

    • Rebuilt the event calendar around member feedback and shared ownership across an 8-person committee, lifting active participation by 40%.
    • Won Event of the Year twice among 12 societies.
    • Point of contact for 200+ attendees and a volunteer team.

About

Peter Yousefi

Recent Computer Science graduate who likes turning messy real-world problems into things people can use, then making sure they hold up in production.

Strongest hands-on work started with data — writing SQL against a live HRIS during an ERP migration where a bad record reaches someone's paycheque. That habit shapes the AI projects here: they check model output, record what happened and keep a person in the loop. Built with AI-assisted tooling, with the effort going into design, validation and testing. Passed AZ-900, studying for AI-103.

Before this, coordinated international logistics and led a student society — part of why the focus stays on the person using a system, not just the system itself.

Education
Bachelor of Computer Science, Trent University, 2023 – 2026
Based in
Toronto, ON
Focus
GenAI / AI engineering / agents
Status
Open to opportunities

Certifications

  • Microsoft Azure Fundamentals (AZ-900)Earned
  • Azure AI Apps & Agents Developer Associate (AI-103)In progress
  • AWS & Databricks Cloud IntegrationsEarned
  • Python 101 for Data ScienceEarned
  • Data Workshop Level 2Earned

Contact

Let's build something.

Open to GenAI and AI engineering roles. Email is the fastest way to reach me.

peter@petery.orgLinkedInGitHub