Jason
Steltman

Python, Next.js 16, Supabase, Vercel, Railway, multi-LLM integration

• AI platforms confidently misrepresent businesses — wrong cities, wrong industries, wrong services. Sourcepull measures it. Runs structured queries across ChatGPT, Perplexity, Gemini, and Claude, classifies each response for accuracy and citation quality, and delivers a fix plan.
• Designed a six-tier classification methodology with an accuracy-first flow — hallucinated citations score negatively, not positively. This widens the score distribution and catches the false confidence that other tools miss.
• Full measurement methodology (v3.0) covering stochastic variance, sample size disclosure, data provenance, and confidence calibration. The honesty layer is the product differentiator.
• Built and operate "Frank," an AI CEO agent with persistent memory, advisory board dispatch, and multi-agent meeting orchestration across 90+ working sessions. Frank chose the niche, argued for it, and won.
• Multi-model pipeline matching cost to task: lightweight models for high-volume classification, capable models for analysis and strategic reasoning
• Accepted into Innovation Factory's Concept Catalyst startup program (April 2026)

Vanilla HTML/CSS/JS, Cloudflare Workers, Python, Claude API

• Government websites are terrible. Finding your representative shouldn't require knowing your ward, your riding, and three different websites. Civic Engagement takes one address and returns every elected official from city councillor to Prime Minister. No cookies, no data collection, no tracking.
• Ward boundary mapping, representative profiles with local and federal voting records, nearby public services, and one-click email drafting
• Burlington city council wasn't transcribing their meetings despite compliance obligations. Had to build a meeting transcriber before we could build a meeting scraper — the full pipeline from scratch using Playwright and AI summarization
• AI chatbot routing residents to the correct city department across 9 Ontario municipalities
• Single HTML file, 5,100 lines, no framework, no backend. Open source and free to use. Two confirmed civic impact stories from real users.

Node.js, Claude API (@anthropic-ai/sdk), pdf-parse, Claude Vision

• Takes a project description (typed, uploaded PDF, or screenshot via Claude Vision), extracts a structured profile, and scores it against a curated grant database. Designed to scale toward discovering every applicable grant for a given project type and region.
• Multi-model architecture: Haiku for high-volume scoring passes, Sonnet for profile extraction and application draft generation
• Pre-filter eliminates obvious mismatches before spending tokens on AI scoring
• Grant Tracker companion tool turns GrantMatch results into a managed application pipeline — track status, deadlines, requirements, and next steps for each grant you're pursuing

Python, numpy, scipy, pytest, Alpaca API

• Quantitative trading system: LMSR pricing engine (logsumexp-stable), sequential Bayesian inference in log-space, fractional Kelly criterion for position sizing
• 66 tests across all modules; math cross-verified against known LMSR values
• Full risk management layer with hard position limits, daily loss caps, and drawdown stops
• Rebuilt the core math for traditional stock markets (MarketBot) via the Alpaca API when prediction markets proved inaccessible from Ontario