Joshua Izzard

tmux in your browser, but it knows what your AI agents are doing. Real PTYs in 6 grid layouts, per-terminal theming, output analyzer that detects Claude Code / Gemini CLI / Python servers. Status dots show active, thinking, idle, and errored at a glance. MIT, vanilla JS client, no build step.

MCP server that gives any AI coding tool persistent memory across sessions. Hybrid search with tiered recency decay, cross-project recall, and automatic dedup. Works with Claude Code, Cursor, Windsurf, and any MCP-compatible client.

Continues reasoning over your work after you close your laptop. Runs as a Supabase Edge Function on a 15-minute cron, finding cross-project patterns, synthesizing insights from weeks of accumulated data, and surfacing them proactively in your next session. The LLM is stateless. Rumen isn’t.

Enterprise scheduling engine for multi-day, multi-venue events. OR-Tools CP-SAT solver handles room conflicts, faculty availability, participant preferences, and practice allocation across 23 constraints with a 100/100 fairness score. 12-step pipeline with real-time SSE streaming, NL-first setup wizard, multi-tenant architecture, Zapier/Make connectors, and 461 tests. 28 frontend routes, zero TypeScript errors.

End-to-end pipeline: Wix CMS sync, AI-powered application cleaning (name normalization, repertoire parsing), server-rendered tearsheet PDFs, review catalog with scoring and keyboard navigation, multi-event support. Live in production with real applicants.

Routes AI agent requests across multiple LLM providers with automatic circuit breaker failover. When one provider goes down, traffic cascades to the next healthy provider. Streaming SSE passthrough, per-task routing tiers, health monitoring daemon.

Full-stack marketplace where patients find and compare healthcare providers through an AI-driven price discovery mechanism. 194K data points, 12K locations, pgvector-powered service matching, Stripe Connect for provider payouts. Built solo in 6 weeks.

Local-only macOS web app that reads Apple’s chat.db directly (read-only VFS), renders conversations with virtual scrolling, and uses Claude for streaming AI analysis with vision support. Google Drive backup with OAuth2 and resumable uploads.

Monitors customer service tickets in real-time, classifies severity using multi-tier AI analysis, and flags conversations likely to escalate to claims or legal action. Multi-tenant architecture with per-organization model configuration.

Chrome extension that integrates Claude AI into Gmail for intelligent email composition, summarization, and response drafting directly in the inbox.

Bulk email tool with template personalization, merge fields, and send scheduling. Handles large recipient lists with rate limiting and delivery tracking.

Building small language models using KAN architectures that produce interpretable outputs — including novel solutions to partial differential equations. Applications in quantitative finance, trading models, and scientific computing.

Deep learning survival framework that predicts when a subscriber will churn, enabling dynamic pricing and retention interventions. Extends the classical survival analysis architecture with novel contributions for subscription economics.

Statistical arbitrage models exploiting pricing inefficiencies in a firm’s capital structure — equity, credit, and derivatives. Connecting the hedge fund experience to modern ML infrastructure.