Portfolio
AI systems built and operated
Each system below was architected, coded, deployed, and is operated by Owen personally. The stacks listed are what actually runs in the code.
Ailura AI Receptionist
In production · paying customers
Multi-tenant AI front office answering voice, SMS, chat, and email for businesses. A tool-using agent books appointments against live calendars, quotes prices, relays messages, and escalates to humans, across all channels from one engine. Emotion AI (53-signal prosody detection) automatically escalates distressed callers. Tiered voice caching controls LLM and TTS cost; per-tenant OAuth calendar integrations; TCPA and A2P 10DLC compliance built in as code; automated nightly activity reporting.
StackOpenAI GPT-4o and Realtime API, agent tool-calling, Hume emotion AI, ElevenLabs, embeddings with RAG; Python, FastAPI, PostgreSQL, Redis, Twilio and Telnyx, Google and Microsoft OAuth; GCP Cloud Run, Fly.io.
Ailura Home Services
In production · live HVAC client
Multi-tenant field-service and e-commerce SaaS running a real HVAC company: missed call to AI text-back, conversational SMS booking agent, route-optimized dispatch, voice-dictated notes, AI-drafted Good/Better/Best quotes, text-to-pay, QuickBooks sync, and nightly KPI reports. Claude vision reads receipts and equipment nameplates into strict schemas with confidence gating. Explainable customer-opportunity scoring carries mandatory reasons and confidence on every row, with fairness guardrails coded in. A database exclusion constraint makes double-booking structurally impossible; the schema, not the prompt, is the final referee over the AI.
StackClaude agents and vision, Whisper transcription, explainable scoring; Next.js 15, React 19, Expo/React Native, Supabase Postgres with row-level security, 23 edge functions; Stripe Connect and Tap to Pay, QuickBooks OAuth sync, Twilio A2P 10DLC; Vercel, EAS.
Cheqer
Live · cheqer.vercel.app
AI-powered Bible word-study platform and graduate research artifact. A multi-turn Claude agent with four retrieval tools answers questions over a corpus of 425,000+ word tokens from eight ancient-text sources, using hybrid retrieval: exact lemma SQL where scholarly tagging exists, pgvector semantic search where it does not, with a deterministic citation-verification layer that structurally prevents hallucinated references. Retrieval quality is measured, not assumed: a reproducible evaluation harness reports precision@10 of 0.772 against a keyword baseline, with published per-case failure analysis.
StackClaude Opus 5 tool-use agent, OpenAI embeddings, pgvector hybrid RAG, IR evaluation (precision@k, recall), anti-hallucination guardrails; TypeScript, Expo, Supabase with 20 SQL RPC functions, Python ETL over eight corpora; Vercel.
Ailura Quant
Deployed · live forward test
Autonomous options-research and paper-trading engine running around the clock in the cloud. Black-Scholes pricing, greeks, and implied-volatility solving implemented from first principles; intraday regime classification; multi-factor candidate scoring; vectorized backtesting; and a read-only brokerage integration that coaches real trades against codified risk rules. Engineered with production reliability discipline: dead-man switches, crash-safe state recovery, audit logging, and 146 automated tests.
StackQuantitative modeling, backtesting and analytics, natural-language command interface; Python, FastAPI, pandas, numpy, SQLite with raw SQL, Alpaca and Schwab APIs; Docker, Fly.io.
NeuroDriver
Research · classroom-tested
Browser-based 3D "glass-box" driving simulator that teaches students how autonomous vehicles perceive and decide. Physically simulated LiDAR, thermal, camera, and audio sensors feed a hybrid AI, combining a TensorFlow.js deep Q-network, a safety behavior tree, and transparent utility scoring, whose every decision is inspectable live. Deployed with 97 middle-school students; the mixed-methods study is under review at ACM SIGCSE 2026.
StackTensorFlow.js DQN, reinforcement learning, explainable-AI interface, sensor simulation; React, Three.js, react-three-fiber, Vite.
Ailura WordPress product line
Shipped commercial products
Two commercial WordPress plugin families demonstrating end-to-end productization. Ailura Guru is a site chatbot with a five-provider LLM abstraction (Claude, GPT-4o, Gemini, Grok, and custom endpoints), encrypted key storage, knowledge-base retrieval, and CRM lead capture. Ailura Blog Writer (eight tagged releases, in production for a nationally distributed counseling-education platform) does deep-research generation with enforced JSON schemas, DALL·E 3 imagery, and automatic SEO integration. Licensing, free and pro editions, marketplace compliance, and release engineering included.
StackFive-provider LLM routing, structured outputs, DALL·E 3; PHP, WordPress, MySQL, Freemius licensing, release automation.
STRONGHOLD and Vigilant Spirit
Mobile · explainable AI
STRONGHOLD (Android, Google Play pipeline) is a recovery-aware AI strength coach that fuses wearable biometrics, including HRV baselines and workload ratios, with training data. Claude adjusts the plan through schema-constrained structured outputs that server-side guardrails re-validate: the model can adapt but never violate the safety envelope. Includes hand reverse-engineering of a rowing machine's Bluetooth protocol. Vigilant Spirit is an AI dream journal with genuine explainability, showing SHAP and LIME attributions to end users to demonstrate transparent consumer AI.
StackClaude structured outputs, Zod-constrained decoding, SHAP, LIME, biometric feature engineering; TypeScript, Expo, SQLite with Drizzle, Health Connect, BLE; Express, PostgreSQL.