Hi, I'm Nick.
I am
Three systems, walked through
Pre-launch builds on synthetic data. No real users, no production incidents.
LedgerLens: AI proposes, a human approves, the ledger stays correct.
Problem Advisers need answers from contracts they can trust, and corrections that never touch the books unchecked.
System Document-intelligence chat over advisory contracts with a verification gate, on top of a double-entry ledger.
- 102SEC contracts indexed
- 49question eval gate
- $0.0093median cost per answer
1. DashboardCorpus health at a glance: 102 of 102 contracts indexed, 817 extracted fields waiting for a person, and the golden-set eval result. 
2. 102 contracts, one pipelineEvery contract shows its steps: stored, parsed and PII-tokenised locally, fee terms extracted and verified, then indexed. 
3. Cited answerEvery figure in the answer links to the page it came from; the model does not calculate. 
4. Human approval gateA fee mismatch becomes a proposed balanced journal entry that waits for a person. Demo mode: nothing is written. 
5. Audit trailEach AI decision is logged with tool, model, prompt hash and inputs.
MediCoord AI: An AI triage that is not allowed to invent a clinic.
Problem Patients need a nearby, appropriate facility, and a language model alone will happily make one up.
System One LLM pass classifies severity, a deterministic PostGIS backend picks a real facility, a second pass writes the reply.
- 0.956DeepEval faithfulness
- 35/smessages per server
- 600+commits
1. Chat and facility mapThe city's 308 facilities are real records on a live map, next to the assistant. 
2. Follow-up questionsThe agent asks targeted clarifying questions before deciding severity. 
3. Severity and real facilityAn ESI-style urgency, a real nearby clinic with drive, cycle and walk times, and the route on the map. 
4. Agent graphThe two-pass design: tools return facts, confidence gates decide between follow-up, answer and human handoff. 
5. GraphRAG experimentA SNOMED-CT graph was built and measured, and the best retrieval did not give the best triage. That result is reported as found.
docciter: From PDF region to citable data.
Problem Answers over technical drawings are only useful if they can point to the exact page and region they came from.
System React and FastAPI workspace: draw capture and ignore regions on PDFs and OCR them. The cited-answer step is designed, not built.
1. WalkthroughThe annotation workflow in motion, recorded by the author. 
2. AnnotationCapture and ignore tools on a real construction drawing set; the inspector lists regions as they are drawn. 
3. Extraction hubWhere captured regions are meant to land as tagged entities with source and match status. Layout with sample rows. 
4. Corpus workspaceDocument list and ingestion status. The tile figures are design placeholders: the vector store is not connected. 
5. Citation UI previewChat with a source-grounding inspector. Honest status: no answering engine is connected yet; this previews the citation UI.
Areas of Expertise
Industry Credentials
Experience & Education
Independent Contributor
Aug 2026 — PresentContributing to BikeSpace, a free open web app that lets Torontonians report bicycle-parking issues, with 1,000+ reports collected and an open API and public dashboard used by city planners. Working as an independent contributor in a volunteer team of engineers, designers and community organizers.
Fullstack Software Engineer
Jan 2024 — Sep 2024Early-stage research startup applying ML and computer networking to real-world problems. Owned the full delivery cycle from API to UI in a fast, research-driven environment.
Master's — Computer Science
2021 — 2023Algorithms & data structures, software testing, TDD, architectural design patterns, Agile, data science and ML fundamentals.
Fullstack Engineer → Project Lead
Sep 2021 — May 2024Joined as an intern, grew into delivery ownership. Led client meetings, managed the full software lifecycle and drove technical decisions across Java backend and Angular frontend.
Data Analyst Program
Apr 2025 — Jul 2025Intensive training in data analytics — visualization, storytelling with data, exploration and data-driven decision making.
My Portfolio
LedgerLens: Fintech Ledger + Document Intelligence
Double-entry ledger plus cited document-intelligence chat for advisory contracts
MediCoord AI: Toronto Health Coordination
Toronto health-triage system: LLM triage, graph grounding, AWS data pipeline, load-tested on staging at about 35 messages/s per server
Postair: Content Platform with Hybrid AI Search
Content platform with 3-tier hybrid AI search and graceful degradation
docciter: Document Annotation & Citation MVP
PDF annotation workspace with OCR capture, built toward cited answers
Commitr: Developer Focus & Session Tracker
Developer focus tracker with GitHub-style activity heatmap
AI Evaluator: Agentic Observability Platform
Agentic RAG evaluation suite with HITL dashboard
AI Healthcare sickness severity classification
Gemini 2.5 Flash fine-tuned on Vertex AI for severity triage
Firo: Financial CSR Bot
Gen-AI assistant for banking CSR reps with confidence scores
Firo ML: User interaction topic classification
Finance topic classifier on 50k transcripts with SBERT
Remote System Monitoring
Streaming server metrics dashboard on Django + K8s + CodePipeline
Taskalert Full UI
Angular 19 microfrontend with native federation and NgRx
Power BI Dashboard on Public Dataset
Power BI dashboard with storytelling on public dataset
Reactive Springboot Skillrater App
Reactive Spring Boot with SSE and multi-threaded live voting
Insight Hub: Springboot Microservices
Microservices issue tracker with Kafka, Resilience4J and LDAP
Taskalert v2: Frontend Application
MEAN stack task manager with NgRx and PusherJS notifications
Taskalert v2: Backend Application
Django REST backend with RabbitMQ and WebSocket channels
Python AI Real‑Time Chat Message Filtering
Kafka-powered AI agent monitoring live chat for policy compliance
Python Asynchronous Web Data Scraper
Async Django scraper with Selenium for dynamic JS-rendered sites
Python Advanced NLP Text Processing
BERT-based text summarization and contextual keyword augmentation
Mobile Application UI & Interaction Design
High-fidelity Figma prototype with mobile interaction patterns