Elizabeth Ho

• Two-stage GenAI pipeline: an LLM “prompt compiler” turns fragmentary child speech into style-constrained image prompts, then a diffusion model generates against a fixed emoji template for consistent output.
• Layered content safety for a child audience: input pre-filtering, LLM semantic rewriting so scary ideas become kid-friendly instead of hitting a “denied” screen, Bedrock Guardrails, and constrained templating. An adversarial test suite checks all of it.
• Voice-first UI designed for pre-literate users: Web Speech API input, giant touch targets, emoji-based feedback, zero required reading.
• Pluggable inference backend. One env var swaps Bedrock for a fully local stack (Ollama/Llama 3.2 + ComfyUI/SDXL Turbo) so it runs offline at no cost.
• Physical output pipeline: 300 DPI print composition matched to 2-inch round-label sheets with slot tracking, CUPS printing, and a fallback cache so no kid leaves empty-handed.

• Fully event-driven pipeline: S3 upload → Lambda → Transcribe → DynamoDB, so recordings up to 2GB process in the background while the UI polls.
• Custom speaker-diarization parser that walks Transcribe's raw word-level output and reconstructs a labeled, readable transcript.
• Prompt-engineered GPT-4o-mini call returns strict JSON scores for communication, clarity, pace, confidence, and professionalism, with fallback parsing and typed error handling.
• Self-deploying infrastructure. An API route creates the CloudFormation stack, zips and pushes the Lambda, and wires up the S3 trigger.

• The rulebook cannot leak private info by construction. A deterministic redactor runs twice and errors on anything left over, and the rule type has no field for a name, snippet, or file path.
• Detection is a pure AST engine, not an LLM: the flagship detector finds newly added public methods with no test referencing them, so a lesson minted from one codebase catches the same mistake in a completely different one.
• One write path, always human-gated. Accepting a lesson is the only way the rulebook changes, and each rule's id IS the new version.
• An optional AI advisory layer flags semantic resemblances, labels itself honestly, and degrades to a deterministic heuristic if anything fails.

Built with Tiffany Yung, Trianna Nguyentu, Elizabeth Ho, William Mar