Gerard – AI agent development, Python, LLM

This project is a literary competition platform with an AI-powered cinematic audiobook production engine. Authors submit stories, readers complete comprehension checks before voting, and winners receive full audiobook productions with character voices, soundscapes, and original music. The backend is a Python-based AI pipeline that processes manuscripts, tags characters and emotions, synthesizes voices, generates sound and music, and renders the final audiobook. The frontend is a Next.js 14 app with PostgreSQL, and the system uses a Strategy Pattern for flexible, cost-optimized integration with external providers like LLMs, TTS, and music engines.

• Designed and built a 12-agent AI pipeline in Python/LangGraph that transforms manuscripts into cinematic audiobooks end-to-end Implemented the Strategy Pattern across four service interfaces (LLM, TTS, sound, music) with swappable provider implementations and a tier-based cost optimization system;
• Built a two-pass NLP character extraction system combining spaCy NER with Claude structured output for 95%+ accuracy on speaker attribution Architected a shared PostgreSQL database serving both a Python FastAPI backend and a Next.js 14 frontend using SQLAlchemy and Prisma respectively;
• Created a reading verification system with Claude-generated comprehension questions, paginated behavioral tracking, and engagement scoring;
• Implemented a ranked voting system with anti-gaming constraints - comprehension gates, minimum read counts, and 5-point ranked allocation per category;
• Built the Spectrum of Authorship disclosure framework with category-specific forms capturing AI involvement, prompts, and workflow metadata;
• Designed a 21-table database schema optimized for future ML training - authorship perceptions, reading transactions, and linguistic metrics as labeled training data
• Developed automated pipeline verification tooling - HTML screenplay review reports, fuzzy-match attribution checkers, and multi-fixture comparison testing;
• Integrated Azure Cognitive Speech and ElevenLabs TTS behind a unified provider interface, reducing per-book production cost from $28 to $5 Built Docker Compose infrastructure with PostgreSQL, Redis, and FFmpeg for local development and CI;
• Implemented structured LLM output with Pydantic v2 schema validation, automatic retry on validation failure, and exponential backoff on rate limits.

It is a desktop productivity app that helps users maintain focus during work sessions through app/website blocking and gamification. It offers three session modes (Zen, Flow, Legend) with escalating enforcement levels. The app tracks flow states, awards badges, and maintains streaks to motivate consistent deep work. Built for knowledge workers and developers who struggle with digital distractions. Currently available on macOS with Windows support planned.

• Designed full-stack architecture using Tauri (Rust) backend and React/TypeScript frontend;
• Built a cross-platform desktop app with native system integration for app monitoring;
• Implemented SQLite database schema with 7+ tables for sessions, analytics, and badges;
• Created 26+ Tauri commands bridging Rust backend to TypeScript frontend;
• Developed a flow state machine with grace periods and real-time tracking;
• Built a gamification system with 40+ badges, streaks, and social sharing;
• Designed dual-layer documentation strategy (architecture specs + implementation prompts);
• Handled macOS app signing, notarization, and distribution;
• Implemented telemetry, interventions, and penalty systems for session enforcement;
• Created agent-ready task files enabling autonomous AI-assisted development.

It is a voice-first clinical documentation platform for senior living communities - skilled nursing facilities, assisted living, and memory care. It replaces checkbox-based incident reporting with guided voice conversations powered by agentic AI, allowing nurses to speak naturally about what happened while the system analyzes their narrative against clinical Gold Standards frameworks and generates precisely the questions needed to complete the record.

The platform operates in two phases: Phase 1 is frontline-staff-led voice capture and AI gap-fill, and Phase 2 is leadership-led investigation with interdisciplinary team coordination, root cause analysis, and dual-signature lockdown.

• Designed and built the full multi-agent agentic pipeline using LangGraph.js;
• Architected the clinical Gap Analysis and Question Compression engine (Gold Standards framework);
• Built the stateful conversational session management system using Redis across serverless functions;
• Designed and implemented the multi-tenant data isolation layer with facility-scoped query enforcement;
• Created the full-stack application architecture across Next.js 14, MongoDB, and Vercel;
• Built the voice capture system with real-time transcription, Wake Lock, and iOS voice fallback;
• Designed the PHI-aware push notification system with device-type routing;
• Implemented the WAiK Intelligence natural language query layer with vector embeddings;
• Built the two-phase investigation workflow with electronic dual-signature and audit trail;
• Designed the complete 30-screen UI specification across all user personas and roles;
• Architected the Progressive Web App layer with offline queue and background sync;
• Integrated OpenAI (GPT-4o-mini, Whisper, text-embedding-3-small) across the agent pipeline;
• Built the MDS revenue intelligence layer, surfacing reimbursement opportunities from clinical data;
• Designed the role-based access control system across 10 distinct permission tiers;
• Produced the complete 18-task agentic build plan structured for autonomous AI agent execution.

An enterprise-grade conversational AI assistant designed to streamline the discovery and retrieval of approved corporate marketing and media assets from a large-scale centralized repository. The core challenge was eliminating the friction caused by manually navigating a massive media library - a pain point for both internal teams and external partners searching for legally approved assets for specific campaigns. The solution enabled users to find the exact assets they needed through natural language queries, dramatically reducing time-to-asset and minimizing human error in the selection process.

• Researched AI frameworks such as LangChain, Prompt Flow, and Graph through rapid proof-of-concepts;
• Architected the final chatbot solution using Microsoft Copilot Studio;
• Adapted project scope and strategy to meet changing enterprise constraints;
• Collaborated with the engineering team to leverage individual developer strengths;
• Delivered a client-focused product tailored to specific business requirements.