Tools

What it is: A GitHub App that performs document-grounded compliance reviews on pull requests. Instead of generic linting, CodeReview AI ingests a company's actual documentation (security policies, style guides, architecture standards, legal requirements) and reviews every PR against that ground truth.

What it demonstrates: Fail-Closed Execution and Artifact-Based Authority patents applied to code review. Every finding is tagged DOCUMENT_BACKED or LLM_OPINION, so teams know exactly which findings come from their own rules and which are the model's suggestion. A two-pass pipeline (analysis + hallucination guard) ensures the system never fabricates policy violations.

Technical implementation: Multi-gate review pipeline (Security, Architecture, Style, Legal, Onboarding) with per-gate document retrieval, structured finding generation with citation references, hallucination verification pass, and traffic-light output (RED/YELLOW/GREEN). Deployed on Railway with GitHub webhook integration. Discovery module auto-classifies repository documentation into review gates during setup.

V2 Roadmap: V2 will have 2 tiers, Tier 1 is everything above. Tier 2 adds full codebase analysis on setup, scoring your repo across all five gates with a baseline strength rating for each. A historical dashboard tracks how each gate trends over time so you can see whether your codebase is improving or declining since install. Tier 1 tells you what's wrong with this PR. Tier 2 tells you what's happening to your codebase.

What it is: An autonomous AI trading agent built on a two-phase trigger system. Deterministic technical analysis runs continuously as passive radar, and when indicators converge on a high-probability setup, AI analysis comes online to verify and execute. Every trade passes through fail-closed gates. Every decision is fully auditable.

What it demonstrates: Fail-Closed Execution, Zero-Loop Control, State-Locked AI Orchestration, Bidirectional Decision Quality Analysis, and Four-Quadrant Diagnostic Regime Detection patents applied to autonomous financial execution. The deterministic layer computes technical indicators locally in microseconds with zero hallucination risk. The AI interprets deterministic ground truth. It does not originate signals from raw data. A hallucination verification gate compares AI output against deterministic values before any capital moves.

Technical implementation: Two-phase "gun-cocking" trigger system with deterministic passive radar, multi-dimensional signal ranking, dynamic strategy confidence weighting (30/70 to 70/30 AI-to-deterministic ratios depending on trade type), TTL management with cached analysis, constrained self-correction schema preventing permanent bad lessons, bidirectional learning from all four decision outcomes (correct trades, incorrect trades, correct rejections, incorrect rejections) with full data parity between real and phantom trades, four-quadrant diagnostic regime detection (STABLE/REGIME_SHIFT/GATE_ISSUE/EXECUTION_ISSUE) that pinpoints where in the pipeline to fix rather than just flagging that something is wrong, five independent circuit breakers, and API cost optimization gating. The agent learns through structured context injection, not fine-tuning, and is architecturally incapable of blowing up the account.

Available after paper trading validation

What it is: A production stock analysis system with a 3-stage LLM pipeline and dual compliance gates. Ask it about any publicly traded company and it returns confidence-scored analysis grounded in live market data and current news.

What it demonstrates: Fail-Closed Execution, State-Locked AI Orchestration, Artifact-Based Authority, and Intent-Contract Enforcement patents in a production system. Most AI systems fail open, producing output regardless of data quality. This system fails closed, gating all responses behind explicit confidence scores (HIGH/MEDIUM/LOW) derived from data freshness and source validation.

Technical implementation: Gate A intent classification, Stage 1 Draft Builder (Sonnet), Stage 2 Delivery Filter (Haiku), Stage 3 Policy Scanner (Haiku), and Gate B output validation (PASS/REWRITE/FAIL_CLOSED). Uses Alpha Vantage for market data, Brave Search for news, and Anthropic Claude for analysis, all wrapped in fail-closed validation. Response time optimized from 32s to 20s through model-tier allocation.

Launch Finance Assistant →

Note: Educational use only. Not financial advice. Demonstrates patent architecture, not investment recommendations.

What it is: An automated daily email system that delivers a personalized morning briefing with weather, tech news, market data, and curated intellectual content at 6 AM every day via cloud execution.

What it demonstrates: State-Locked AI Orchestration applied to recurring automated workflows. A two-LLM pipeline with structured history prevents content repetition across days. Each run is fully stateless with all context reconstructed fresh per execution through structured prompt injection.

Technical implementation: Claude Sonnet with web search tools generates full HTML email content, sent via SendGrid API. Deployed on Railway with cron-based scheduling. Inline CSS only for Gmail compatibility. Structured summarization of each sent email prevents repetition without persistent memory. Runs under $1/month in API costs.

Automated, no public interface