Innovation & Research

Overview

FinanceOS is an AI-powered workspace built for Canadian financial advisors. When an advisor selects a client, the system surfaces their full financial picture: accounts, balances, contribution room, tax documents, and action items. A per-client Knowledge Base acts as the advisor's living memory, storing goals, family context, and preferences that AI agents reference in every interaction.

Multi-Agent Architecture

• Query Classifier: Analyzes each message and routes to the optimal path: direct lookup, Knowledge Base update, or selective agent dispatch.
• Context Agent (Claude): Reads client profile, knowledge base, documents, and conversation history to draft personalized communications.
• Quant Agent (GPT-4o): Writes and executes Python for financial calculations (tax brackets, contribution optimization, CESG, HBP comparisons).
• Compliance Agent (Claude): Audits analysis against CRA rules, CIRO suitability requirements, and contribution limits with hard-coded regulatory thresholds.
• Shadow Backtest: Proactively scans all clients for opportunities like idle cash, RRSP deadlines, OAS clawback risk, and RRIF minimums.

Technical Stack

Frontend: Next.js (App Router), React, TypeScript, Zustand, WebSocket

Backend: FastAPI, Python, SQLite, Pydantic

AI: Claude, GPT-4o, multi-agent orchestrator with code execution sandbox

Deployment: Railway (backend + frontend services)

Design Principles

Human-in-the-loop at every decision point. Agents never fabricate numbers; missing data outputs "UNKNOWN." All tax rules, contribution limits (RRSP, TFSA, FHSA, RESP), and regulatory references are Canadian-specific and hard-coded as ground truth.

Overview

Most AI financial tools answer questions. This one stands between you and your next bad trade. Describe a trade you're considering and a visible processing pipeline kicks off: portfolio analysis, behavioral bias detection, then four specialized agents deliberate in parallel. A fifth synthesis agent reads all four and produces a structured Pre-Trade Brief with a clear signal.

Agent Committee

• Portfolio Impact: Analyzes concentration, account type implications, and portfolio-level risk from your actual holdings.
• Behavioral Risk: Detects bias patterns (loss aversion, disposition effect) from your real trade history with specific dates and prices.
• Devil's Advocate: Argues rigorously against the trade to surface blind spots you might be ignoring.
• Tax Analyst: Calculates Canadian tax consequences from your adjusted cost base.
• Committee Synthesis: Reads all four outputs and produces a PROCEED / CAUTION / HIGH RISK signal with a primary concern.

Technical Stack

Frontend: Next.js, React, TypeScript

AI: Claude claude-sonnet-4-6, 5 parallel streaming agents via Anthropic SDK

Analysis: Pure TypeScript modules for portfolio math, bias detection, and tax calculations

Privacy: No server-side data storage. Portfolio data stays in the browser.

Design Principles

Specialized agents with adversarial roles produce sharper thinking than a single AI asked to "analyze from all angles." Every brief ends with "THE DECISION IS YOURS" and an explicit list of what the AI cannot know. Built for the Wealthsimple AI Builders Competition, March 2026.

Overview

Refactored an academic assignment into a modular CLI application that orchestrates transcript ingestion, linguistic cleaning, and phoneme projection. The system mirrors original directory hierarchies to maintain experiment traceability while keeping the codebase lean for portfolio presentation.

Pipeline Innovations

• Transcript mirroring: Recursively synchronizes cleaned and phoneme outputs with source corpus layout for reproducible experimentation.
• Regex-driven normalization: Precompiled pattern suite strips metadata, diarization artifacts, and noise while preserving contractions and speaker intent.
• Pronunciation intelligence: CMUdict mapping resolves 94.7% of tokens to deterministic ARPAbet sequences with graceful handling of out-of-vocabulary words.

Technical Architecture

Languages & Tools: Python, argparse, pathlib

Structure: Modular pipeline package with reusable I/O helpers, phoneme transformers, and automated directory bootstrapping.

Data Sources: CHILDES `.cha` corpora, CMU Pronouncing Dictionary

Impact

Delivers normalized corpora with 98.8% utterance retention and phoneme projections covering 94.7% of lexical tokens, accelerating downstream language-modeling experiments while documenting rationale in portfolio-ready narratives.

Overview

AI parser that pairs handcrafted CFG rules with deterministic chart parsing to flag defective drafts for enterprise content teams.

Core Capabilities

• Adaptive CFG engine: Modular productions generalize to unseen phrasing while keeping interpretability front-and-center.
• Deterministic inference: NL Toolkit ChartParser transforms POS-tag sequences into transparent accept/reject decisions with confidence analytics.
• Evaluation telemetry: Automated precision, recall, and coverage reporting accelerates grammar iteration cycles.

Technical Stack

Languages & Tools: Python 3.11, NLTK ChartParser, standard library CSV utilities

Data: POS-tagged correspondence corpus curated for AutoML compatibility

Delivery: Dual CLI + Python module packaging for batch jobs and integrations

Impact

Provides audit-ready grammar enforcement that de-risks editorial pipelines.

Overview

This research project addresses one of the fundamental challenges in reinforcement learning: handling partial observability in continuous data streams. By integrating Real-Time Recurrent Learning (RTRL) with streaming Deep Reinforcement Learning algorithms, we enable agents to maintain and update memory of past observations in real-time without requiring full episode replay.

Key Innovations

• RTRL Integration: Implemented online gradient computation that eliminates the need for backpropagation through time, enabling true streaming learning
• ObGD Optimization: Developed adaptive optimization techniques specifically designed for online credit assignment in Partially Observable Markov Decision Process (POMDP) environments
• Sample Efficiency: Achieved 2.1× improvement in sample efficiency on MuJoCo continuous control benchmarks compared to baseline streaming algorithms
• Memory Management: Designed efficient memory architectures that balance computational cost with information retention

Technical Stack

Languages & Frameworks: Python, PyTorch, JAX, NumPy

RL Libraries: Stable-Baselines3, RLlib, Gymnasium

Benchmarks: MuJoCo, Atari, DeepMind Control Suite

Impact & Results

The research demonstrates significant improvements in both computational efficiency and learning performance. The streaming approach reduces memory requirements by 60% while maintaining comparable or superior performance to traditional experience replay methods. This work has implications for real-world applications where agents must learn continuously from streaming data, such as robotics, autonomous systems, and adaptive control.

Overview

Refactored coursework into an end-to-end product narrative that packages corpus normalization, deterministic splits, and model training into reproducible commands with JSON artifact tracking.

Core Capabilities

• Corpus pipeline: Automated validation, deduplication, and stratified train/dev splits seeded for reproducibility.
• N-gram factory: Extensible NGramLanguageModel abstraction with Laplace smoothing and cross-order persistence.
• Perplexity console: Evaluation harness that surfaces OOV-aware metrics before production deployment.

Technical Stack

Languages & Tools: Python 3.10, argparse, dataclasses, unit-tested modules

Artifacts: Structured JSON checkpoints, CLI logging with timestamps and corpus stats

Integrations: Designed for ASR/TTS pipelines and hybrid neural pairing

Impact

Reduced corpus prep time by 60%, improved dev-set perplexity 2.1× over unsmoothed baselines, and enabled shareable experiment playbooks for hiring panels and research peers.

Overview

DataMod is a comprehensive database management tool that provides a unified interface for interacting with both SQL and NoSQL databases. The application bridges the gap between traditional relational databases and modern document-oriented storage, offering developers and database administrators a single platform for all their data management needs.

Core Features

• Multi-Database Support: Seamlessly work with MongoDB (NoSQL) and SQL databases through a unified interface
• CRUD Operations: Complete Create, Read, Update, and Delete functionality with intuitive UI controls
• Query Builder: Visual query construction tool that generates optimized SQL and MongoDB queries
• Schema Management: Create and modify database schemas with validation and constraint support
• Data Visualization: Built-in tools to visualize data relationships and query results
• Import/Export: Support for CSV, JSON, and XML data formats for easy data migration

Technical Implementation

Backend: Python with PyMongo for MongoDB connections and SQLAlchemy for SQL database abstraction

Database Support: MongoDB, PostgreSQL, MySQL, SQLite

Architecture: MVC pattern with modular design for easy extension to additional database types

Security: Parameterized queries to prevent SQL injection, connection pooling for performance

Use Cases

DataMod is ideal for development teams working with hybrid database architectures, database administrators managing multiple database types, and students learning database concepts. The tool simplifies complex database operations while maintaining the power and flexibility needed for advanced use cases.

Overview

CitiWatch is an AI-powered public safety platform designed to assist law enforcement agencies in identifying and responding to potential weapon threats in real-time. The system leverages state-of-the-art computer vision technology to analyze video feeds and images, providing immediate alerts and comprehensive threat assessment tools through an intuitive web dashboard.

Key Features

• Real-Time Weapon Detection: YOLOv5-based image recognition system trained on extensive weapon datasets, achieving 94% accuracy in threat identification
• Interactive Dashboard: React-based frontend displaying live threat alerts, historical incident data, and risk analytics
• Global Risk Map: Geographic visualization of threat patterns using Mapbox GL, enabling strategic resource allocation
• Threat History: Comprehensive logging system storing incident details, timestamps, and associated metadata for investigation and analysis
• Multi-User Authentication: Role-based access control with Firebase Authentication, supporting different permission levels for various law enforcement roles
• Alert System: Configurable notification system with SMS and email integration for immediate threat response

Technical Architecture

Frontend: React.js, Redux for state management, Material-UI components, Mapbox GL for mapping

Backend: Flask (Python), RESTful API design, WebSocket for real-time updates

AI/ML: YOLOv5 (PyTorch), OpenCV for image preprocessing, custom training pipeline

Database: Firebase Realtime Database for live data, Cloud Firestore for historical records

Deployment: Docker containerization, Google Cloud Platform hosting

Achievement

★ 1st Place - AI Track, GovTech Hackathon

CitiWatch was recognized for its innovative approach to public safety, practical implementation, and potential for real-world impact. The judges praised the system's accuracy, user-friendly interface, and comprehensive feature set that addresses genuine law enforcement needs.

Overview

QRchive is a location-based mobile gaming platform that transforms the real world into a competitive playground. Players scan QR codes placed throughout their city to collect points, unlock achievements, and compete on global leaderboards. The app combines elements of scavenger hunts, geocaching, and social gaming to create an engaging outdoor experience.

Core Features

• QR Code Scanning: Fast and reliable QR code detection using Android Camera2 API and ZXing library
• Point System: Dynamic scoring algorithm based on QR code rarity, location difficulty, and time-based challenges
• Leaderboards: Real-time global and local rankings with weekly competitions and seasonal events
• Social Features: Friend system, team challenges, and in-app chat for coordinating hunts
• Achievement System: 50+ unlockable achievements with unique badges and rewards
• Map Integration: Google Maps integration showing nearby QR codes and player locations
• Profile Customization: Customizable avatars, themes, and player statistics

Technical Implementation

Platform: Native Android (Java), minimum SDK 24 (Android 7.0)

Backend: Firebase Realtime Database, Cloud Firestore, Firebase Cloud Functions

Authentication: Firebase Authentication with Google Sign-In and email/password support

Storage: Firebase Cloud Storage for user-generated content and profile images

Maps: Google Maps Android API with custom markers and clustering

Testing: JUnit, Espresso for UI testing, Robolectric for unit tests

Team Leadership & DevOps

As team lead, I coordinated a group of 6 developers using Agile methodologies with 2-week sprints. Implemented comprehensive CI/CD pipeline using GitHub Actions for automated testing, code quality checks, and deployment to Google Play internal testing track. Established code review processes, documentation standards, and conducted regular knowledge-sharing sessions to ensure consistent code quality across the team.