I build AI applications and distributed systems, from LangGraph multi-agent assistants and MCP-powered retrieval to cloud database infrastructure. My work combines practical software engineering, applied machine learning, and evidence-based evaluation.
Multi-agent AIMCP + RAGDistributed databasesApplied ML researchCloud infrastructure
01 / Experience
Engineering across AI and infrastructure.
Oct. 2025 — July 2026AI SamuraiRemote - US
AI systems
AI Software Engineer
Built production media intelligence pipelines that summarized 100+ audio and video assets with OpenAI and Gemini APIs. Integrated LangGraph agentic workflows for multi-step reasoning and tool use, and developed Python services using AWS, codecs, and WebAudio.
Improved summary quality by 25% through backend processing and evaluation.
Benchmarked models, engineered prompts, and automated quality measurement.
Connected AI outputs to Tableau and Power BI for practical analysis.
PythonLangGraphOpenAI APIGemini APIAWSWebAudioTableauPower BI
May 2024 — Sep. 2025Huawei CanadaMarkham, ON, CAN
Cloud databases
Cloud Database DevOps Engineer
Developed low-level C, Python, and Shell software for Linux cloud database systems supporting enterprise SaaS and 5G deployments. Built and debugged replication, high-availability, and recovery components, with Databricks and GitLab supporting the work.
Automated DevOps workflows to improve deployment efficiency by 40%.
Built LLM-driven synthetic data generation from DDL schemas for testing and benchmarking.
Accelerated semantic search and data retrieval by 23% with Ollama, Vanna AI, and vector databases.
Turns customer questions into schema-aware database queries and cited policy answers. A LangGraph supervisor coordinates SQL and RAG agents, combining Oracle records with company policy PDFs while retaining context across follow-up questions.
2Specialist agents
11MCP tools in Oracle mode
SQL + RAGCombined evidence
The SQL agent discovers tables, columns, and relationships before generating queries. SQLGlot validation, bound parameters, a read-only Oracle account, and execution limits constrain access, with one repair attempt for recoverable errors. Chroma and OpenAI embeddings retrieve policy passages with page-level citations.
FastAPI and Streamlit power chat and PDF uploads. Includes a SQLite fixed-lookup alternative, verified data migration, automated tests, and live-model evaluation. Local demo using fictional customer data.
Fine-tuned Granite-7B with Red Hat × IBM InstructLab to detect political, gender, age, ethical, marketing, and research bias. The system identifies biased text, categorizes it, and explains the reasoning as advisory feedback.
~0.99Political-bias accuracy
3Models evaluated
6Bias taxonomies
Designed compositional skill taxonomies, generated and filtered synthetic Q&A data, then compared Granite, Mistral, and Merlinite using MT-Bench-style judging and labeled Kaggle datasets. The study also documented model-specific regressions, false-positive risks, and the limits of bias classification.
IBM InstructLabGranite-7BPythonGCPMT-Benchscikit-learn
Built a worldwide photo-geolocation model on an EVA-02 backbone, combining continent classification with gated coordinate regression to predict continuous spherical coordinates from Street View imagery.
96.1%Continent accuracy
38K+Training images
251 kmMedian error
Created a custom dataset from 12,600 locations with three angles per location, data augmentation, teacher forcing, and an angular loss objective. The model outperformed PlaNet's 1,131.7 km median error and produced a 683 km mean error worldwide.
PyTorchEVA-02PythonGoogle Street View APINumPyTransfer learning
Architected a serverless collaborative canvas for concurrent real-time interaction. API Gateway WebSockets and AWS Lambda handle persistent connections and pixel broadcasts, Redis provides connection state and atomic rate limiting, and CloudFront reduces read latency through edge caching.
U of T Software PlatformLed three developers to build a platform for 200+ students, bringing secure sign-in and real-time updates together with React, Flask, and MongoDB.Explore profile ↗
CrewiseAn expedition health platform with role-based access, built with React, TypeScript, Flask, and MongoDB and deployed on DigitalOcean.View code ↗
Remixed UNOA multiplayer twist on UNO for local-network play, built in Java and JavaFX with SOLID principles and reusable design patterns.View code ↗
Zip & UnzipCompress and restore text, images, and audio with a Python tool powered by Huffman trees and efficient data structures.Explore profile ↗
03 / Patent work
Database replication at the systems level.
Co-inventor 2025
Huawei · Cloud-native databases
DMA pull-based RDMA log synchronization
Co-inventor on a 2025 patent related to synchronizing database logs between master and replica nodes using a DMA pull-based RDMA approach. The work targets a critical distributed-database path: moving replication data efficiently while supporting high availability and recovery.
The public portfolio intentionally describes the technical area at a high level and does not disclose confidential implementation details.
Built agent workflows using planning, reflection, memory, tool use, and multi-agent coordination, with emphasis on reliable execution and evaluation.
2026 · Google
Agile Project Management
Applied Scrum and Kanban through backlog refinement, sprint planning, delivery metrics, risk management, and stakeholder communication.
2026 · Riipen / Red Hat × IBM
InstructLab Bias Detection
Designed bias taxonomies, generated synthetic data, fine-tuned Granite, evaluated model behavior, and documented the system in a technical white paper.
2024 · Coursera / DeepLearning.AI
Machine Learning Specialization
Covered neural networks, decision trees, clustering, anomaly detection, recommender systems, reinforcement learning, and practical model evaluation.
2023 · Microsoft
Azure Fundamentals
Validated core knowledge of Azure compute, storage, networking, identity, security, governance, monitoring, and cloud cost management.
2023 · University of Toronto
Dean’s List Scholar
Academic distinction recognizing sustained high achievement in the Computer Science Specialist program.
2021 · University of Toronto
Entrance Scholarship
Merit-based recognition awarded on admission for strong academic achievement.
2019 · Stephen Lewis SS
Most Valuable Swimmer
Athletic award recognizing competitive performance, consistency, leadership, and contribution to the swim team.
06 / Technologies
Tools across the stack.
Languages
Python · C · C++ · Java · TypeScript · JavaScript · SQL · Shell · Assembly
AI / ML
LangGraph · LangChain · MCP · RAG · OpenAI API · Gemini API · OpenAI embeddings · PyTorch · TensorFlow · Ollama · Model evaluation
