Things I've built — mostly to learn, sometimes to ship.
AI-powered YouTube highlight detector. Drop a URL, get the 5-7 most shareable moments with timestamps, suggested captions, and platform recommendations (TikTok, LinkedIn, Twitter). Built in one overnight session.
Paste a job description, get 12-15 tailored interview questions and ideal answers in your own voice. Categories: behavioral, technical, situational, culture fit, role-specific. Powered by Claude Opus.
Real-time snow conditions and scoring for 50+ ski resorts across North America. Rates mountains out of 10 based on fresh snow, snowpack, bluebird conditions, temperature, wind, and visibility. Features an adventure meter for Vancouver's North Shore mountains.
A mobile-first workout tracker built for strength & conditioning. Log sets, track progressive overload, and view exercise progress. Features an AI coach with e1RM tracking, post-workout analysis, personal records board, and Strava integration.
Sports science training load monitor with ACWR gauge, readiness scoring (sleep, soreness, stress, energy, mood), 40+ exercise presets with HR zone tracking, and evidence-based training recommendations. Features RPE distributions, monotony/strain trends, and load-vs-readiness correlation analysis.
Personal finance dashboard with budgeting, savings goals, loan tracking, and AI insights. Imports real bank statement CSVs with pattern-based categorization. Features net worth forecasts, subscription detection, a full transaction ledger, live market tickers, and multi-currency support.
A full-featured job application pipeline with Kanban board, timeline tracking, and three Claude-powered tools — cover letter generator with streaming, job fit analyzer with visual score ring, and tailored interview prep. Built for the job search that funded its own development.
AI-powered code learning tool with line-by-line annotations, concept tracking with confidence progression, learning stats dashboard with GitHub-style streak calendar, and analysis history. Supports 10 languages. Turns vibe-coding into deliberate learning.
Interactive rental market data, suburb comparison, and yield calculator for New Zealand property. Combines MBIE Tenancy Services bond data with Stats NZ demographics. Features Mapbox map explorer, market dashboard with trend charts, side-by-side suburb comparison, and gross yield calculator.
Travel blog and photo gallery documenting adventures around the world.
Computer vision rep counter that runs entirely in the browser. Uses MediaPipe Pose Landmarker to track body landmarks and count reps for squats, push-ups, curls, and more. Zero data leaves the device.
Real-time collaborative whiteboard with freehand drawing, shapes, sticky notes, and live cursors. Multi-user sessions powered by Liveblocks. Works on desktop and touch devices.
Live currency conversion with multi-currency comparison, historical rate trends, and support for 30+ currencies. Fetches real-time exchange rates from open APIs.
Holistic wellness tracker combining sleep, stress, energy, and activity data. Features ACWR-based load monitoring, pattern detection, day-of-week analysis, and a weekly radar comparison report.
CLI tool for managing AI agent configuration files. Diff configs, apply presets, use git hooks to keep agent prompts in sync. 14 built-in presets for common agent personas.
One-command CLI to create a GitHub repo, add Vercel integration, and deploy a Next.js app. Automates the entire setup process from local project to live URL in seconds.
Cycling endurance analysis built on published sports science research. Upload a Garmin FIT file to measure power durability — how well you sustain output as fatigue accumulates. Tiered scoring, multi-ride trends, decay curves across work bins. All client-side. Built by the researcher who published the methodology.
Sports science AI chat backed by peer-reviewed research — including Harrison's published papers in EJAP. Streaming answers on durability, CTL/ATL/TSB, HRV, VO2max, periodization, and carbohydrate metabolism. Hand-crafted knowledge base for precision over RAG hallucination risk.
Document intelligence demo — paste any document (contracts, research papers, product specs, reports) and ask questions in plain language. Get cited answers with exact quotes from the source. Privacy-first: documents are never stored on the server.
Real-time adventure scoring for New Zealand hiking tracks. Combines Open-Meteo weather forecasts with GeoNet seismic data to rate conditions out of 10. Interactive Mapbox map, track safety scoring, duration filters, and expandable condition breakdowns.
Multi-step LLM orchestration — visible. Drop in any document and watch a 4-stage AI pipeline run: Extract structured facts, Analyze risks and obligations, Synthesize an executive summary, Generate action items with owners and timelines. Each stage shows timing, token usage, and intermediate output.
AI-generated interview prep dossier. Enter a company name and role — get tech stack analysis, culture signals, smart questions to ask, red flags to investigate, and positioning advice. Useful prep in seconds, not hours.
Paste any research paper, abstract, or scientific text — get an instant structured breakdown. Plain English summary, key findings, methodology critique, limitations, practical applications, and follow-up questions. Built for scientists, students, and curious non-experts.
Interactive Retrieval-Augmented Generation pipeline — no black boxes. Load documents, run a query, and watch BM25 score each chunk, see which were retrieved and why, then get a Claude Haiku answer with inline source citations. Every step is transparent.
Paste any code, get a structured AI review in seconds: bugs, security vulnerabilities, performance issues, style suggestions — each with severity ratings and specific fixes. One-click generation of corrected code. Supports 12 languages.
Watch the agentic loop live: the AI reasons about which tools to call, executes them, and synthesizes results — all visible in the UI. Five real tools: weather, calculator, web search, timezone lookup, currency conversion. Shows the core pattern behind every production AI agent.
Paste any contract — employment agreement, NDA, SaaS terms, service agreement — and get an instant structured analysis: risk flags ranked by severity, plain-English summary of what you're agreeing to, key obligations by party, and specific negotiation leverage points.
3-stage autonomous AI pipeline for legal matter processing. Paste any matter description and watch: Stage 1 extracts structured data (parties, issues, urgency), Stage 2 streams 3 strategic investigation angles, Stage 3 generates a client update, billing narrative, and risk assessment — all in parallel.
Autonomous multi-step AI research agent. Enter any topic, and the agent plans a structured research agenda, researches each section in sequence with SSE streaming, and compiles a formatted report — all visible in real time. Shows the plan → execute → compile agentic workflow.
The Model Context Protocol, visualized in real time. Two-panel layout: chat on the left, live JSON-RPC protocol trace on the right. Watch Claude discover tools, call them with typed parameters, receive results, and synthesize responses — every protocol message visible as it happens.
A real, working MCP server — not a simulation. Built with the official @modelcontextprotocol/sdk via stdio transport. Install in Claude Desktop with one config line and ask Claude about Harrison's portfolio, research, projects, and job search. 6 tools with Zod-validated schemas.
Real-time streaming comparison of Claude Haiku vs Sonnet, side by side. Enter a prompt, watch both models stream simultaneously. Shows TTFT, tokens/second, and total time. Makes model tradeoffs visceral instead of theoretical.
Three AI workflows for legal practice: matter activity → client update emails, time entries → billing narratives, court documents → structured calendar events. Built to show what 'AI workflows, not AI features' looks like in the legal domain.
Systematic LLM prompt evaluation tool. Define a prompt template with variables, add test cases with expected outputs, run them all against Claude, and get scored results with a 'Fix My Prompt' AI assistant that streams improvement suggestions based on failures.
Side-by-side comparison of four prompting techniques — zero-shot, few-shot, chain-of-thought, system-prompt tuning — all streaming simultaneously on the same question. Shows TTFT, token counts, and quality differences in real time. Temperature slider and model selector included.
Live, visible multi-agent coordination system: user gives a task → Planner decomposes into subtasks → three Specialist agents (Researcher, Writer, Critic) execute in parallel → Synthesizer assembles the final output. All five agents stream in real time. Shows what production AI orchestration looks like.
Real-time voice AI assistant in the browser. Speak → Claude hears → Claude responds → Claude speaks back. Under 1 second from end of speech to start of audio response. Built using browser-native Web Speech APIs — no Deepgram, no Whisper, zero external audio dependencies.
Six context strategies applied to the same task, streaming in parallel. Baseline, Role+Persona, Grounding, Few-Shot, Constraints, Full Stack — all fire simultaneously. Watch quality emerge from context engineering in real time. Based on Andrej Karpathy's 'context engineering' framework — the discipline replacing 'prompt engineering' in production AI.
An AI that actually remembers you. Chat on the left, live memory graph on the right. Every message: Claude responds using injected memory, while a parallel lightweight call extracts new facts and adds them to the store. Memory persists across sessions. Makes the extract-store-inject architecture that powers production AI memory tangible and interactive.