AI Systems Engineer · Applied ML Researcher
Toronto, Ontario
I design the systems around AI models: agent harnesses, specialized workflows, and evaluation infrastructure built for security, observability, and human control.
Across coding agents, smart-contract security, and applied ML research, I use cross-model and cross-harness evidence to separate model behavior from context, tools, orchestration, and product failures.
Education
Honours BSc in Computer Science
Current role
CredShields
Published at
IEEE MSWiM 2025
Focus
AI systems · evaluation
Selected work
Observability, agent infrastructure, security, and applied ML.
Local-first observability and evaluation across six AI coding harnesses.
Built read-only adapters and a provenance-aware SQLite ledger that reconcile duplicate sessions and revalidate transcript content from canonical local sources before serving it through a CLI, local dashboard, and read-only MCP tools.
A local observability prototype for AI-training systems.
Built a Rust telemetry sidecar, FastAPI control plane for anomaly detection and policy evaluation, and Next.js operations console. The verified local scope covers Apple Silicon telemetry and repeatable simulated faults.
A local Rust prototype for bounded delegation between ChatGPT and Codex.
Built a Rust MCP server with workspace-scoped context, managed worktrees, OAuth/PKCE, audit records, and approval boundaries that prevent model-controlled authorization without exposing a raw shell by default.
Undergraduate thesis on CodeBERT and graph-based smart-contract analysis.
Fine-tuned CodeBERT on 47,000+ smart contracts, reaching 99.89% binary accuracy and 89.31% multi-class accuracy on the thesis dataset. Also explored AST- and CFG-based graph models and template-based fixes for reentrancy and arithmetic overflow.
Additional work
A host- and model-decoupled workflow for structured founder intelligence.
Built nine bounded MCP tools that turn untrusted X evidence into reviewable decisions, with approval before paid collection, SQLite persistence, and atomic monthly credit controls.
A local-first guard for selected developer-tool installs.
Built a zero-dependency Go guard that selects mature versions, checks advisories, stages and scans supported artifacts, and fails closed when package identity or policy cannot be proven.
A multi-service Solana social prototype with AI-assisted discovery.
Architected and implemented a prototype spanning Next.js, Express and BullMQ, FastAPI retrieval and moderation pipelines, and Anchor programs for social and payment workflows.
Experience
Applied AI, ML research, and product work across industry and academic collaborations.
Lead the design and implementation of a backend AI system and agent harness for smart-contract security
Developed VAE and DDPM approaches for synthetic RSSI fingerprint generation in an industry-academic indoor-positioning project
Translated a web product into a Flutter app for iOS and Android
Built machine-learning pipelines for imminent-fracture-risk prediction
Skills
The languages, frameworks, and infrastructure I use most.
Publications & writing
An IEEE conference paper, with concise articles on AI systems, evaluation, and security to follow.
IEEE MSWiM 2025 · Conference paper
Co-authored research using autoencoders and VAEs for synthetic RSSI augmentation, reporting 6–25% lower RMSE and 11–400% higher R² in the paper's evaluation.
Articles coming soon
Concise notes on building, evaluating, and securing AI systems.
