Selected Projects

Building real value from AI across the full stack

I help teams get results from AI by building the full stack of solutions: simple software tools that boost productivity, AI-driven automations that connect existing systems, and AI applications like RAG and agents that search, summarize, and generate outputs with traceability.

Six Projects

01

Stomps

Self-hosted AI customer support assistant

GitHub

Problem

Small businesses want reliable 24/7 website support automation without recurring SaaS subscriptions and vendor lock-in.

What I Built

A self-hosted RAG-based support chatbot that trains on website content/files, answers questions using OpenAI models, and retrieves context via Pinecone.

  • Self-hosted deployment: runs on a VPS with your own domain/subdomain (not a hosted SaaS dependency)
  • RAG-style knowledge base: trains on files + website links using Pinecone for vector indexing and retrieval
  • LLM-powered responses: uses your OpenAI API key (GPT-3.5/GPT-4) to generate context-aware answers grounded in retrieved content
  • Admin + observability: admin panel for API keys, knowledge import, conversation history, and basic analytics
  • Multi-site + white-label: embed on multiple websites with brand-customizable UI and no "powered by" branding

02

Matter Pack Assistant

RAG-powered legal document analysis

Watch Demo

Problem

Legal teams waste hours manually reviewing lengthy matter packs to extract key facts, parties, dates, and risks for case preparation.

What I Built

A RAG-powered legal document analysis system that ingests matter packs, enables semantic search with access control, and generates structured case summaries with grounded citations.

  • Multi-format parsing: extracts text from PDFs (with form fields), DOCX (with heading hierarchy), and TXT with precise source locators
  • Smart chunking: hybrid strategy merges small blocks and splits large ones with 100-token overlap, preserving document structure
  • Access-controlled retrieval: per-user ACL enforcement ensures users only retrieve documents they're authorized to access
  • Grounded summaries: LLM-generated summaries with validated citations that link back to specific document locations
  • Eval harness: 25-case test suite measuring retrieval recall and party/date extraction F1 scores against ground truth

03

BizPlan Model-in-a-Box

AI business plan + 5-year financial model generator

Live Demo

Problem

Early-stage ecommerce founders struggle to produce investor-ready financial models and business plans without expensive consultants or spreadsheet expertise.

What I Built

An AI-powered web app that turns a 10-minute questionnaire into a complete 5-year financial model (Excel) and narrative business plan (Word doc).

  • Hybrid research architecture: curated benchmark database (11 categories, 5 regions) with live web search fallback via Claude Agent SDK
  • Equity optimization loop: iteratively adjusts funding to ensure positive cash balance across 5-year forecast
  • Direct Excel manipulation: XML-level editing + xlcalculator for formula recalc without Excel dependency
  • Type-safe pipeline: JSON schemas generate both TypeScript types (frontend) and Pydantic models (backend)
  • Production-ready API: async job queue with progress polling, 24-hour retention, background cleanup

04

WriteSharp

Chrome extension for clearer, more professional writing

GitHub

Problem

People waste time rewriting messages for clarity and tone across web editors, especially in professional support and ops workflows.

What I Built

A Chrome extension that rewrites selected text using the user's OpenAI API key, with a review-first popup and editor-aware insertion.

  • Client-side rewrite workflow: popup-based, side-by-side review of original vs improved text
  • BYO OpenAI key + endpoint: uses user-provided API key with option to use a proxy endpoint
  • Output fidelity controls: preserves structure (lists/paragraphs) and factual details; supports custom prompts + model selection
  • Editor-aware insertion: direct insert where allowed; copy-to-clipboard flow for strict editors; detects editor type
  • Minimal-permission + synced settings: uses activeTab/scripting/storage; no telemetry

05

TicketPeek

Zendesk ticket previewer

Internal @ Shopify

Problem

Zendesk agents lose time and focus opening tickets just to read recent comments while triaging from list/search views.

What I Built

A Chrome (MV3) extension that previews ticket comments on hover inside Zendesk list/search views, with Markdown rendering and one-click open.

  • Zendesk in-page preview: shows ticket comments from list/search views without opening the ticket
  • Secure, instance-local fetch: uses the logged-in session on *.zendesk.com (no third-party services; minimal permissions)
  • Rich rendering: displays internal notes + public replies with Markdown, author names, and timestamps
  • Configurable triggers: hover preview with optional "hold P" / "hold Q" modes; preference saved via storage
  • Fast navigation + guardrails: one-click open to full ticket; respects rate-limit/permission constraints

06

Carrier Claim Automation

Zendesk + Zapier + OpenAI workflow

Internal @ Shopify

Problem

Carrier claim processing is repetitive: agents manually extract order details and reformat them into carrier-specific templates before forwarding.

What I Built

A Zapier-orchestrated workflow that detects the carrier, extracts claim details from incoming Zendesk tickets, and generates/forwards a carrier-compliant template via OpenAI.

  • Carrier detection + routing: inspects incoming claim tickets (e.g., subject pattern) to identify the carrier and route to the correct workflow
  • Structured data capture: uses order-page-generated claim details already present in Zendesk as the source payload
  • Carrier-specific templating via OpenAI: sends extracted fields with a carrier-specific prompt to match each carrier's required format
  • Automated handoff: creates a new outbound Zendesk ticket and forwards it to the carrier automatically
  • End-to-end automation: handles thousands of tickets at scale with consistent formatting across carriers