CarlosBueso

What I take off your plate

You have a client project with an LLM component and nobody free to build it. I ship production AI systems end to end — I don't need an architect, a data engineer and a frontend dev to be three separate people.

A production LLM system serving over a hundred active users: WhatsApp voice notes through Whisper transcription and GPT-4o prosody analysis, into a RAG pipeline that writes the client's reports for them.

The retrieval behind it runs over 4,800 document chunks — text-embedding-3-small, exact cosine in NumPy against SQLite-backed vectors. I chose exact over an approximate index because at that size recall mattered more than the milliseconds an ANN would have saved.

An art valuation platform for Artvia in Next.js and Flask, with algorithmic pricing models for high-value works, built from spec to production. And AI-generated video for BBVA, directed and produced on generative media tooling.

Before consulting, two years at AllianceBernstein — gradient-boosted models on twelve million rows of financial time series in Snowpark and Snowflake Cortex, pipelines traders used daily, and internal libraries and REST APIs opening forty datasets across ten databases. Transaction latency down about ninety percent, sixty seconds to three.

Where I come from

Mexico City. Before I wrote software I rode ambulances here: nearly two years as a volunteer paramedic with SUUMA, weekly training, EMT coursework, nineteen patients in the field. My first lesson in systems that cannot fail quietly.

Then computer science at Loyola University Chicago with a minor in finance — natural language processing, cryptography, the student investment fund, and four winters colder than any deadline since. Two years building quant tooling in Nashville tied the threads together, and in 2024 I came home.

I also run a small consultancy here, Itzam.ai. Its clients are Mexican and Latin American commercial teams — yours are not on that list and will not be. Subcontracting is how I fill the space between those projects. Running my own shop is also why I understand delivery pressure and client politics: I won't make you look bad in front of your client.

Things made for joy

Cecil — a multi-agent financial analysis platform in Next.js and Python, coordinating agents over live market data through Groq and LLaMA, streaming their reasoning as it happens.

Matra — an app that listens to family conversations and grows them into a living ancestry tree. Pre-beta, headed for the app stores.

Archie — an MCP server that exposes SketchUp's Ruby API as structured tools, so a model can build 3D geometry from a sentence. I ran evaluations across the geometry operations and triaged twenty-odd reproducible failures into schema fixes, then put it in front of practising architects driving it from Claude Desktop.

voxcentra — AI films. Generative media pipelines built on Runway and Higgsfield, where cinema and code overlap.

And the earliest joy: streaming my own brainwaves from a Muse 2 headset into a Raspberry Pi, the first time a machine read my mind.

How this usually starts

A small scoped piece first — a spike, a prototype, one module — so you can see how I work before you put me on anything load-bearing. Two weeks is usually enough for both of us to know.

I'm not pitching a partnership or trying to get onto a vendor list. If a project lands that you'd otherwise turn down or scramble to staff, I'm a known quantity you can call. Email is the fastest route.

charlesbueso@gmail.comgithublinkedin