Projects

Financial projection software for a specific industry. Users model long-term scenarios, plan capital improvements, and generate stakeholder presentations.

• Multi-year financial modeling
• Scenario comparison
• Stakeholder presentation generation
• Persona-tested with 3 user profiles

Automated prompt engineering using Stanford's DSPy framework. Optimizes Claude Code skill prompts through bootstrap, copro, and iterative algorithms with cross-validation.

• 11 of 13 optimization targets deployed
• Bootstrap, CoPro, and iterative algorithms
• Cross-validation with dropout regularization
• Background optimization with progress tracking

Docker Compose setup for running AI agents. Best-effort Docker hardening for broadly trusted agents, not a containment boundary for compromised code.

• Read-only root filesystem, dropped capabilities
• Ordinary HTTP(S) traffic routed through a Squid allowlist; DNS and documented engine-specific bypasses remain
• Best-effort output checks with secret detection; they do not gate consumption or scan all writes
• Optional agent metadata in YAML

Three-phase autonomous revenue discovery system. Discovers niches, validates with AI-realizability gates, develops MVPs, and deploys.

• AI-realizability gate prevents unbuildable projects
• Autonomous niche discovery and validation
• 26+ shared skills across pipeline

Pipeline-based game development exploring AI-generated games. Active projects include Slime Survivor, WW2 Gacha, AFK Gacha, and autonomous game development experiments.

• AI-generated .tscn scene files
• Autonomous game development experiments
• Multiple genres: survival, gacha, idle

Two Discord bots for workspace automation. Evolution bot handles revenue and capability updates. Orchestrator bot manages workspace status, game updates, and cross-project awareness.

AI-assisted family history research using multiple search APIs, document analysis, and systematic brick-wall breaking strategies.

• Multi-source search (Exa, Brave, Codex)
• Document analysis with Gemini OCR
• Brick-wall breaker agent for stuck research
• Family tree builder with relationship integrity

This blog. Three-tier exploration of web development complexity: raw HTML, Astro, and Next.js. Features agent-accessible API, comment system, and embedded AI chat.

Custom Model Context Protocol server integrations: Codex CLI wrapper, search framework selection, mgrep semantic search, and tool search optimization.

• Codex (GPT-5) MCP for cross-validation
• 85% token reduction via Tool Search
• mgrep semantic search replacing grep
• Multi-model orchestration framework

Time-locked language models trained on pre-cutoff historical texts using Karpathy's nanochat pipeline. Exploring whether small models trained exclusively on period texts can reproduce the linguistic patterns of their era.

• 65GB historical text corpus across multiple eras
• Time-locked training methodology (a publication-date filter; a 2026 audit removed misdated volumes, 0.28% of the training cache)
• RTX 3090 local training pipeline
• Parquet-based shard management

Local OCR pipeline for digitizing physical documents. Scans paper records through a local LLM (OLMo 2) to produce structured spreadsheet data, with human review for accuracy.

• Local LLM inference (no cloud PII exposure)
• TIFF/PDF to structured data conversion
• Human-in-the-loop validation workflow

Personal conversation analysis system. Extracts structured data (claims, decisions, emotional arcs, negotiation patterns) from SMS and chat history, with a benchmarked extraction schema refined through multi-metric field screening.

• Benchmarked V3 extraction schema with weighted field scoring
• Four-metric candidate screening (consistency, information gain, discrimination, redundancy)
• Calibration measurement for extraction confidence

Private voice and speech research. Details withheld by design.

Collaborative fiction project: first-person fantasy narrative from an anterograde amnesiac protagonist. Multi-agent workflow with world-librarian, story-writer, story-curator, and story-editor agents.

• Multi-agent creative pipeline
• World-librarian for factual consistency
• Story-curator for post-chapter canonization
• Journal entry format with unreliable narrator

Fork-based prompt-cache warmer for Claude Code. Keeps long sessions warm in the Anthropic prompt cache without modifying them: idle sessions nearing cache expiry are warmed through disposable interactive forks, with per-warm cache_read receipts proving each refresh worked.

• Zero session impact — live transcripts verified byte-identical before/after warming
• Empirical TTL measurement from session history (measure-ttl.py, zero API spend)
• Self-verifying: every warm logs the measured cache_read token count
• Disproved two naive approaches empirically (print-mode pings, local heartbeat files)
• ~20x cheaper resume on large sessions vs paying a cold cache re-write

Personal portfolio falsification system: every investment thesis decomposes into pre-registered falsifiers with named public observables, watched twice daily by an AI monitor with escalating alerts. Regime-tolerant exit rules and state-triggered re-entry replace judgment at extremes — past me decides, present me executes.

• Pre-registered falsifiers with named public observables (CLEAR / WATCH / FIRED)
• Twice-daily AI monitor with web search checks each observable against current evidence
• Escalating decision-class alerts that repeat until acknowledged
• Regime-tolerant exit rules with state-triggered re-entry
• Born from a model-led audit of five years of instinct trading

A buying guide that ranks 32 espresso machines sold in Canada for the drink you order at Starbucks. Every score names its source (a written review, a moment in a YouTube test, or the maker's specification), prices are in Canadian dollars and dated, and the page says where the evidence is thin.

• 576 scores (32 machines, 18 factors), each citing its source: 387 in all, including 87 YouTube tests cited to the second
• Your drink order sets the weights; the result is an estimated fit out of 100 that says how much of it rests on thin evidence
• Prices in Canadian dollars as each store showed them on 1 October 2026, each dated and marked not live; stores whose terms forbid using their prices show none
• Researched and built by Claude Opus 5.5 with parallel research, checking and calibration agents, then corrected after an independent review