Projects

An internet-born character and the short-form video workflow behind it. Episodes are rendered, lip-synced, and reviewed by hand before anything is published, then pushed to the social accounts from a single approved queue. The public page collects every short in order, from waking up in a box to building his own neon sign.

A public MCP server for the Open Science Framework. Three read-only tools over the OSF API — project metadata, abstracts, and component trees — reachable by any MCP client with no credentials and no account. Rate limited per IP, upstream responses cached, and constrained by a single rule: public data in, public data out.

Production MCP server with safety-gated tool execution, ToolContext metadata system, JSONL audit logging, and structured error envelopes with self-healing breadcrumbs. stdio transport for Claude Desktop and VS Code. Includes lab, git, docs, knowledge, devloop, and script tools.

Multi-strategy ML pipeline for classifying Gmail at scale. Regex pattern matching with confidence scoring, LDA/NMF topic modeling, TF-IDF keyword extraction, sender family clustering, and confidence-based LLM escalation for uncertain cases. Processed 16M+ emails.

CLIP-first media triage pipeline for Immich. Zero-shot semantic classification across 14 topics using CLIP embeddings (ViT-B/32), logistic regression keep/discard classifier on 512-dim vectors, NMF unsupervised clustering, NSFW detection, and a human-in-the-loop feedback cycle for continuous retraining.

YAML-driven paper trading framework with 27+ technical indicators, composite signal scoring across 9 weighted categories, and a Random Forest crash predictor (200 trees, 18 features). Runs 24/7 on crypto via Alpaca paper API. Includes backtester, Flask dashboard, and systemd service management.

Interactive H3 hex grid scraper for Google Places data. Click hexagons on a Leaflet map to select them, scrape nearby businesses through a Google Apps Script proxy, and export enriched data as CSV. Auto drill-down when a hex hits the 60-result API cap — recurses into child hexes at higher resolution (up to 3 levels).

Hybrid MT system for the Kaggle Deep Past Challenge (Akkadian to English). TF-IDF char n-gram retrieval baseline with SequenceMatcher reranking, name-aware top-2 mixing, and ByT5 fine-tuning pipeline on aligned pairs. Includes BLEU/chrF evaluation and systematic alignment sweep framework.

Self-hosted job discovery and fit-scoring system. Pulls postings from Greenhouse, Lever, RemoteOK, and HN Who's Hiring, then scores each against a weighted skill profile. Includes a resume tailor that builds an LLM prompt constrained to the master profile — no fabricated experience.

Self-hosted email intelligence pipeline that scans Gmail via API, classifies messages using regex pattern matching and LLM reclassification (Ollama), extracts subscriptions and keywords, builds sender profiles, and automates inbox cleanup. 8 MCP tools, tested on 16,000+ emails with 200+ subscriptions extracted.

Knowledge management system built on SQLite FTS5 with multi-model coordination, epistemic decision capture, and a memory recall/confirm protocol. Hot/warm/cold document tiering with automatic promotion and decay. Multiple AI models (Claude, GPT, Codex) collaborate through a shared devloop database exposed over SSE.

Real-time League of Legends coaching tool that polls the Live Client API, recommends items based on enemy composition, analyzes death patterns, and displays advice via an in-game overlay. Built from 89 matches and 640 death snapshots — 44% of deaths identified as overstays, the top coaching signal.

WordPress publishing automation pipeline built on MCP. Create and update posts with confirm/dry_run safety gates, auto-generate branded OG images (1200×630) using Pillow — no AI APIs, fully deterministic. Category and tag management, content file support, and SEO-ready featured images with category pills.

Automated LinkedIn posting system with JSON-based queue scheduling. Define posts with scheduled dates, times, content, and hashtags — a cron job handles publishing via the LinkedIn REST API. One-post-per-run throttling, dry-run preview, force-post, and queue status reporting. Supports opinion posts and article shares.