Experience

Agentic AI for software security: LLM-driven static analysis, exploit reproduction, and automated repair — and the evaluations that decide whether any of it works.

• First author of ZeroFalse: a multi-stage LLM pipeline for static-analysis triage, evaluated across 10 frontier LLMs; best F1 0.912 on the OWASP Java Benchmark and 0.837 on real-world CodeQL alerts.
• Designed CVerify, an agentic CVE-reproduction pipeline with a differential oracle: an exploit counts only if it succeeds on the vulnerable build and is blocked on the patched one.
• Leads evaluation methodology on a line of LLM financial-statement auditing: independent reproduction of headline results, and no-model leak baselines that test whether a benchmark measures what it claims to.
• Co-author on two papers under review at MSR 2027. Teaching assistant for Database Systems, Software Development Methods, Digital Systems Design, and Digital Logic.
• Volunteer project co-lead and mentor (Sep 2025 – Apr 2026): two undergraduate co-op teams and two Master of Cybersecurity lab teams; all five students of one co-op team went on to co-author P2Patch.

Simon Fraser University

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.

• Designed and led P2Patch: an exploitation agent re-attacks each accepted patch, and every execution-confirmed bypass becomes a permanent repair requirement. Held-out fix-closed coverage rose from 78.1% to 84.7% against single-pass repair on the same agent and model (101 CVEs, Holm-adjusted p = 0.029), at about 4.6× the cost.
• After the appointment, built a security-patching benchmark from vulnerability chains (PatchChain): 149 tasks from 106 projects in 9 languages, each graded in a network-isolated verifier against the later fixes developers had to ship — not just the reported bug. The agent sees only the original advisory.
• Evaluated four frontier models in one fixed coding-agent harness over 1,788 runs. Complete-patch rates were 32.7–45.4%; a grader that tests only the reported bug would score the same patches about 2× higher — 51% of patches that fix the reported bug without breaking tests leave a gap developers later fixed under a new CVE.
• Analysed every failed run with instructed model coders: most failures came after the agent understood the bug — the code left vulnerable had been on screen in 79% of them — and derived six checkable design rules for patching agents from the failure classes.

University of Edinburgh

Built the agent layer of an LLM-powered IDE — subagents, multi-agent orchestration, the tool lifecycle — and the evaluations that decided what shipped.

• Ran the multi-model orchestration study behind Fabric’s published March 2026 benchmark: 97.8% on Aider Polyglot at $4.95, against 98.7% at $26.79 for the frontier baseline — 99% of the accuracy at 18% of the cost.
• Shipped a six-tool subagent system and a test-first, RED→GREEN multi-agent DAG orchestrator with a live control dashboard.
• Designed the prepare → permission → execute tool lifecycle, and an MCP server that lets agents drive the test-and-break loop.
• Ran a controlled SWE-bench experiment on graph-augmented code retrieval over an ~18k-line subsystem: no measurable gain, so it was not shipped.

Farpoint Technologies

Evaluated AI/ML capabilities — recommendation, personalization, vision-based form analysis, LLM coaching — and ran the user-research → spec → engineering-handoff loop.

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.

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.

SnappFood

Delivered a working beta in six months with a three-person founding team: Django and PostgreSQL with Celery, RabbitMQ and Redis behind a React front end, deployed with Nginx and Docker. Grew into tech-lead duties as the team reached 3–5 engineers.

Real-time pipeline from London Stock Exchange WebSocket feeds through Elasticsearch to a seq2seq LSTM trading model at tick-level granularity, evaluated head-to-head against the firm’s human traders.

AngularJS front end for a Slack-like business messenger on Django REST: live-session state sync, reconnect and backoff, presence and read receipts, and optimistic-UI reconciliation. First professional web role.