Nate Velasco

Portfolio · V1

Hi, I'm Nate Velasco.

I build small, working software by directing AI coding assistants, making the product and scope calls, and checking what comes back. Below are two early projects you can try right here in your browser, along with an honest account of how they were made.

See the projects Get in touch

Projects

Both are early, changing builds — not finished products. Each demo runs the project's real logic on fictional sample data, entirely inside this page. Nothing you type leaves your browser, and a refresh clears it.

ProductivityEarly buildDeterministic, no AI calls

To Do List: Prioritization + Auto-Assist

Capturing tasks is easy. Deciding what deserves attention next is the hard part. This project reads a quick one-line task, proposes structured details, and ranks everything with a stated reason for each position.

The story

Problem
A long list with no clear "do this next". Due dates, effort, priority and dependencies all matter, and weighing them by hand every day is tiring.
What it does now
Parses quick-add text (dates in Central time, effort, priority words, tags, "hard deadline", "waiting on…"). Suggests a priority only when you didn't give one. Ranks open tasks: near hard deadlines first, then P0 → P4, with tasks that unblock important work moved up, and explains each placement.
This demo
The real parsing and ranking code, transplanted unchanged, running against fictional starter tasks in page memory. The full app's sign-in, database, calendar and other pages are not included.
Limitations
Early and changing. Rules are hand-written heuristics, not learned. The "Auto-Assist" AI direction is a plan, not something this demo does. The parser understands common English phrasings only.
My role
Project owner: I defined the problem and make the product and scope decisions. Exact implementation contributions are still being documented.
AI assistance
Much of the code was written with AI coding assistants under my direction. I'm not claiming sole authorship of the source.
Learning & next
The ranking rules live in plain, testable code with fixed example cases, so every placement can be explained. Next: an AI assist step, kept separate from the deterministic reasons so they stay visible.

Multi-agent prototypeEarly buildSimulated AI

Mastermind Group

A local prototype that explores structured discussion between several AI "panelists": a router picks who joins and which discussion format to use, a critic challenges claims, selected claims are checked, and a final answer keeps disagreements and caveats visible.

Simulated AI — no live model calls. This demo runs in the project's existing test mode.

The story

Problem
A single AI answer can sound confident while hiding weak claims or real disagreement. I wanted to explore whether a structured panel could surface that.
What runs here
The prototype's real routing, Critic Pass and Round Table protocols, cancellation, claim/status handling, verification and synthesis assembly, and trace recording — called directly in your browser.
What is simulated
Every panelist, claim extractor and verifier is the project's mock adapter. Responses, claims and verification verdicts come from templates, not real reasoning or fact-checking. Token counts are estimated from character length; cost is zero. The final answer is not independently verified.
Limitations
An early local prototype. It is not a trained neural network, not a hosted service, and not shown to beat a single model. Its security and escalation hooks are placeholders.
My role
I set the goals, scope and product decisions and coordinated AI-assisted work. Historically, Codex handled integration and tests and Claude handled backend and cancellation work. Further detail is still being documented.
Learning & next
Every run, including a stopped one, keeps a step-by-step trace you can inspect. Not yet done: a comparison against a single model, so no benefit is claimed.

Behind the Scenes

How I build: small steps, written down, checked before they are claimed.

  1. Name the problem. Say plainly what is hard before building anything.
  2. Write the brief. Audience, scope, what's out, and what "done" means — before code.
  3. Direct the build. AI coding assistants implement; I decide scope, review results and push back.
  4. Test against concrete cases. Fixed example inputs with expected outputs, plus trying it by hand.
  5. Record what's true. What works, what doesn't, who did what, and what's still unknown.

This site follows the same process: I set the direction, Claude (an AI coding assistant) implemented it, and Codex coordinated and reviewed. The demos reuse my projects' existing code rather than mock-ups.

Let Them Tell It

Coming soon

Short video references from people I've worked with, in their own words. None are posted yet, so rather than fill this with placeholders, it stays empty until they're real.

Contact

I'm looking for technical and AI-adjacent roles. Reach out any of these ways.

Seen enough?

Thanks for exploring. You lovin it?

You lovin it?

Thanks for taking a look. If you'd like to talk, my contact details are right behind this.