Engineering robust multi-agent systems, scalable machine learning pipelines, and secure model governance architectures. Specialized in LangGraph workflows, real-time data ingestion, and OSFI E-23 compliance.
Coordinated via Google ADK (Agent Development Kit), this multi-agent supervisor orchestrates specialized workers using Model Context Protocol (MCP) interfaces to enforce model compliance with Canadian **OSFI Guideline E-23** model risk management standards.
Google ADK acts as the supervisor model core. It reads worker schemas, handles thread locks, validates JSON configurations, and manages validation feedback loops.
Monitors updates in regulatory guidelines (OSFI E-23, SR 11-7) and updates evaluation vector mappings.
Autonomous crawler and embedding processor that continuously monitors OSFI updates. It converts legal PDF text chunks into structured vectors in ChromaDB, flagging changes that require adjustments to compliance checklists.
Audit logs evaluating a high-performing fraud classification model (v4.1) for credit card application risk.
How the system coordinates from raw benchmarks up to continuous post-deployment monitoring.
| Layer | System | Function |
|---|---|---|
| Layer 5 | Model Risk Dashboard | Continuous post-deployment monitoring aligned to OSFI E-23 §4 — PSI drift detection, AUC performance tracking, demographic parity & equal opportunity bias metrics, monthly surveillance log. |
| Layer 4 | GovernanceOps | Automated promotion pipeline, evaluating deterministic rules pre-deployment. |
| Layer 3 | Governance Control Tower | Multi-agent orchestration via Google ADK coordinating auditing tasks. |
| Layer 2 | Live Governance Systems | Specialist agents executing model validation checklists & monitors. |
| Layer 1 | CanFraudBench / CanFinBench | Published regulatory datasets and calibration baseline benchmarks. |
Technical summaries of deep-dive systems, benchmarking architectures, and self-healing data infrastructure built for enterprise reliability.
A governance decision, made once by a risk officer, becomes a runtime-enforced AI control — verified end-to-end against live production infrastructure, not just in a test suite.
A two-service closed-loop AI assurance platform. GovernanceOps Inventory is the system of record for OSFI E-23 model governance decisions — risk tier, autonomy level, permitted/forbidden tools, confidence thresholds. GovernanceOps Agent is a zero-dependency Python runtime library that fetches that policy directly over HTTPS, enforces it inside a live agent process, and reports enforcement evidence back to Inventory, each event tied to the exact policy version and SHA-256 hash that produced it.
Multi-agent supervisor orchestrating specialized auditing workers using Google ADK.
A multi-agent supervisor coordinating specialist auditing agents via Google ADK to enforce model compliance with OSFI E-23 and SR 11-7 model risk management guidelines.
Continuous model monitoring aligned to OSFI E-23 §4 — drift, performance, and bias surveillance.
Continuous model monitoring aligned to OSFI E-23 §4: PSI drift detection, AUC performance tracking, demographic parity and equal opportunity bias metrics, and a monthly surveillance log. Ships with a synthetic 6-month credit model dataset — drift appears at month 4, bias emerges at month 5, and the compliance score drops from stable to 15/Critical as the story unfolds.
Open-source Canadian identity-fraud benchmark datasets and calibration drift monitors.
An open-source identity-fraud benchmark designed specifically for Canadian regulated environments. Features automated drift alerts and Expected Calibration Error (ECE) tracking logs to audit financial verification pipelines.
Multi-domain regulatory RAG framework mapping Canadian financial audit regulations.
A multi-domain RAG framework mapping 7 Canadian regulatory domains (including OSFI E-23 and SR 11-7 guidelines). Features access-controlled vector stores and strict citation verification checking pipelines to prevent LLM hallucinations.
AI Governance Compliance Analyzer — OSFI E-23. Upload a model card, validation report, or governance policy; get a structured E-23 compliance report with 100% critical gap detection rate and 3/3 injection attacks blocked in under 30 seconds.
Analyzes documents against 8 OSFI E-23 requirement categories and returns a formal finding register — not a narrative summary. Each finding carries regulatory source, severity rating, gap description, and remediation recommendation as discrete fields a model risk manager can annotate, override, and sign off on. Critical findings escalate to second-line review per the E-23 three-lines-of-defense model. Eval: 10 E-23 scenarios, gpt-4o-mini, July 2026.
Local-first document intelligence API converting unstructured financial PDFs into structured schemas.
A 3-layer document intelligence pipeline designed for regulated financial environments where data sovereignty is mandatory. It processes statements locally on CPU with zero cloud OCR outbound calls.
Named in honor of my late mother, Felicia Pienaah 'Daavi' — embodying absolute trust and safety.
A verification-first real estate marketplace solving the diaspora trust deficit in cross-border property investments through PostGIS geospatial mapping, raw SQL verification gates, and vector similarity search.
Multi-Agent clinical intelligence platform for secure radiology workflows & patient portals.
A multi-agent clinical intelligence platform orchestrating specialized AI workers to ingest scans, draft impressions, generate billing codes, and simplify complex diagnostic terminology for patients.
Multi-agent AI system that analyzes business KPIs and generates growth strategies.
A production 6-agent LangGraph system built for the Kaggle × Google AI Agents Intensive (11,494 teams). It analyzes revenue-opportunity from funnel drop-offs, ranks marketing ideas via the ICE scoring engine, and generates content routed directly to Slack.
Hallucination tracking and calibration scoring across LLMs with Pareto frontier optimization.
A mathematical evaluation framework tracking hallucinations and calibration errors across open-source and commercial LLM backends, determining the optimal cost-accuracy trade-offs via multi-objective Pareto optimization.
Autonomous pipeline remediation monitoring system running test scenarios to validate fixes.
An automated self-healing pipeline processor deployed on AWS (App Runner + RDS) to audit data pipeline flows. Cuts time-to-resolution from hours to under a minute by spinning up isolated Docker containers to validate fixes before promotion.
Secondary system MVPs and hackathon entries built for algorithmic execution and local commerce validation.
XAU/USD Algorithmic trading engine with MetaTrader 5 integration.
An automated scalping bot executing positions on gold (XAU/USD) via the MetaTrader 5 API interface. Features trailing stop-loss, risk management boundaries, and local log alerts.
DTTP Hackathon Top 6 Hyperlocal Commerce MVP.
A hyperlocal merchant matching system built for the DTTP Hackathon, ranking in the Top 6. Facilitates neighborhood store listings, geolocation maps, and instant order broadcast.
“Great job! This individual and their entire team gave this project their very best effort! Much appreciative of the opportunity to work with him. Personally he really drove this project forward and served in a great 'quarterback role' for this project. We continue to collaborate!”