AI/ML engineer & researcher
I build LLM agents and machine-learning systems, and the evaluations that show whether they actually work. Currently an M.Sc. researcher at Simon Fraser University. Before that: multi-agent orchestration for an agentic IDE, production ML for 10M+ users, and a trading platform I co-founded.
Open to AI/ML engineering and research roles · Burnaby, BC, Canada
Worked and studied at
Research · first author · 2026
A patch that blocks one exploit has not necessarily fixed the bug. P2Patch has an exploiter agent attack every accepted patch, and turns each bypass it finds into a new requirement for the final fix.
Raised held-out fix coverage from 78.1% to 84.7% over single-pass repair, across 101 CVEs.
Read moreFarpoint · Research Engineer Intern · 2025–26
An LLM-powered IDE where work is split across cooperating agents. I built the agent layer (subagents, a test-first DAG orchestrator, the tool-permission lifecycle) and the evaluations that decided what shipped.
The model-routing study I ran reached 99% of frontier accuracy at 18% of the cost on Aider Polyglot.
Read moreResearch · first author · 2025
Static analyzers flag far more than they should. ZeroFalse gives an LLM the code context and dataflow behind each CodeQL alert, then measures how well ten frontier models separate real issues from noise.
Best F1 of 0.912 on the OWASP benchmark; CWE-specific prompting added up to +0.26 F1 on real-world code.
Read moreSnappFood · Data Science & ML · 2019–20
On the customer-experience team of Iran's largest food-delivery platform, I worked on the models behind arrival-time estimates, churn prevention and vendor-fraud detection.
About 27% more accurate delivery estimates and 13% lower monthly churn, on a platform with 10M+ users.
Read more
Graduate Researcher and Teaching Assistant Simon Fraser University · M.Sc. Computing Science · Dr. Mohammad Tayebi Agentic AI for software security: LLM-driven static analysis, exploit reproduction, and automated repair — and the evaluations that decide whether any of it works.
Visiting Researcher (Mitacs Globalink) University of Edinburgh · School of Informatics · Mitacs Globalink Research Award Three-month visiting appointment, co-supervised by Marc Juarez and Adriana Sejfia with Dr. Mohammad Tayebi (SFU), on one question in automated security repair: whether a fix that blocks the demonstrated attack has actually closed the weakness behind it. Continued remotely with Dr. Marc Juarez after the appointment.
Research Engineer Intern (Applied AI & LLM Systems) Farpoint Technologies · LLM-powered agentic IDE Built the agent layer of an LLM-powered IDE — subagents, multi-agent orchestration, the tool lifecycle — and the evaluations that decided what shipped.
Technical Product Manager EnerGym · Fitness and health-tech platform Evaluated AI/ML capabilities — recommendation, personalization, vision-based form analysis, LLM coaching — and ran the user-research → spec → engineering-handoff loop.
Co-founder & Software Engineer Pabla · Crypto social-trading platform · ~40k users in 18 months Co-founded Iran’s leading crypto social- and copy-trading platform, funded by the Finnova accelerator after a ~2-month MVP. Architected the trading engine: smart order routing across external exchanges, best-execution price aggregation, async Python and Celery, sub-second price fan-out over Redis Pub/Sub, and the copy-replication state machine with slippage controls and per-follower sizing.
Technical Project Coordinator — Data Science & ML SnappFood · Iran’s largest food-delivery platform, 10M+ users Customer Experience team. Adapted Uber’s DeepETA to motorbike delivery (~27% better ETA accuracy, 24% fewer delays); shipped a churn model over 3M+ users that cut monthly churn 13%; vendor-fraud detection that lifted CSAT 10% and NPS from 5 to 7; and the company-wide Octopus BI layer.
SnappFoodThreatEZ: Control-Aware Threat Modeling from Software Repositories
Convict: Finding Vulnerability-Inducing Commits by Execution, Not Attribution
Multi-agent orchestration, tool use and permissions, subagents, context management, and MCP servers.
Vulnerability detection, exploit reproduction, and automated repair with LLM agents in sandboxed environments.
Controlled ablations, LLM-as-judge harnesses, held-out oracles, significance testing, and benchmark construction.
Production models for ETA, churn and fraud; embeddings, clustering and deep learning.
Real-time, high-throughput services; queues and caches; cloud deployment and CI/CD.
From user research to shipped apps: web, desktop and dashboards. Co-founded two products.
From Alert to Patch: The Hard Parts of Automating Vulnerability Repair with LLMs Open to AI/ML engineering and research roles. Email me at mhsn.iranmanesh@gmail.com.
