AI Engineer · AI Systems Architect · Detroit, MI
My breakthrough: I asked what an mp3 is. An array of numbers that pushes air, every second. Then I asked what the eye sees. Only 2D. The brain guesses the shadows and the shapes.
Telemetry is the world read as numbers. Sound, images and text all arrive as arrays, and I build the systems that turn them into decisions. Music and OCR are my passions, so most of this started as a challenge I set myself, and the best of it went to production.
audio, voice and music as feature spaces
in productionpixels into text, text into structure
challenges · projects nextraw text into retrieval and agents
in productionA song, a voice or a full room is one long array. I cut it into frames, extract what the ear and the room care about, compress it and let the structure show itself.
the raw array, live or recorded
short windows, noise filtered
100+ DSP features per frame
the few directions that matter
moments that sound alike
similarity scored, acted on, evaluated
Real-time crowd detection in a live venue, by filtering for voice frequencies, then 100+ features through PCA and KMeans into similarity scoring. It adapts while people are in the room, and a continuous evaluation loop checks it in production.
Python systems that turn tracks into structured feature spaces for real-time, context-aware music decisions, with cover art alongside the signal for a richer picture of a song. Behind the decks it runs live: sequencing, energy progression, and the floor as the feedback loop.
An image is a grid of numbers, the same way a song is a line of them. OCR is where vision meets language: find the text, read it, and give it back a shape a database can hold. It is a passion of mine, and the projects are coming to this page.
the grid, as the camera gave it
light, angle and noise fixed
where the text lives
characters into words
words into rows and fields
Scraped pages, spreadsheets and documents come in broken. I clean them, cut them into pieces a model can quote, and build retrieval and agents on top, with telemetry on every step so I can see what it did.
the web and the files
one shape, no junk
passages worth quoting
RAG over their own records
do the work, logged
evaluated, every step
In production in the financial decision-support RAG and the M&A Training Hub (see Portfolio).
Where sound goes once you can read it. I built each of these to learn the edge of it, not for a client. Press play anywhere.