A few things I've built with AI.

A React Native prototype that analyzes one or more job-site photos and drafts hazards, mitigations, confidence ratings, and compliant conditions.

The model prepares a first pass. A worker reviews every finding, records corrections or dismissals, signs, and owns the final assessment.

The current evidence is the prototype's end-to-end review flow, which preserves the original model result alongside each human correction or dismissal.

Policy references are not yet grounded in an approved document store, and the model has not been evaluated against a reviewed set of job-site images.

A tested Python reference implementation. A repeat question is answered from a certified, versioned query asset with full provenance, no LLM in the answer path. A new question still goes through the model, but the result has to clear a deterministic safety check before it can run.

The LLM only touches new questions, and every query it generates has to pass a safety gate before it runs. Only a certified, reviewed asset can answer a repeat question on its own.

A working pipeline covering parsing, the safety gate, the registry, and a promotion workflow, all covered by an offline test suite and runnable locally against a stand-in warehouse.

It doesn't have a real UI yet, and it hasn't been connected to a live warehouse. Those pieces are designed for but not built.

A Windows desktop app that transcribes a course locally, captures slide changes, runs OCR on them, and uses Azure OpenAI to generate notes and flashcards for spaced-repetition review.

Raw audio is never written to disk. It's transcribed in memory and dropped after each chunk. Only the finalized transcript and what's built from it get saved.

A working capture-to-transcript pipeline built on local Whisper, automatic slide-change detection, and an AI note-generation flow, with a test suite covering the core logic.

It's Windows-only, and the review and export experience is still rough. It hasn't been run through a full course start to finish yet.

A React and Supabase app where trainees run realistic customer calls against an AI persona using OpenAI's Realtime API, then review an AI-scored transcript.

The AI runs the practice call and produces a first-pass score. A trainee or admin still decides what the score means and what happens next.

A working role-based platform: admins build scenarios and personas, trainees take calls, and every session is transcribed and scored automatically.

It hasn't been used with a real training team yet, and the AI scoring hasn't been checked against a human grader to see how much they agree.