Roxana dT
Builder in San Francisco

Agent tools, AI evals, and small software that has to survive contact with users.

I spent a long time at FAANGs, did a stint in VC, and now build apps with my interns, coding agents, and a lot of verification.

See LoopGate Harness Read the conference writeup
Black and white pixel portrait of Roxana dT

Selected Work

Current public projects around coding agents, evaluation, and the practical mess of using AI to build software.

AI Deployment Calculator
Web app

AI Deployment Calculator

A browser-based AI deployment calculator that estimates GPU VRAM requirements and recommends hardware for AI workloads — text generation, embeddings, vision, image diffusion, and LoRA/QLoRA/full fine-tuning — accounting for model size, quantization, batching, and framework overhead. Runs fully client-side, no backend.

Launch calculator Repository
Loop of absurdity diagram from the inference conference project
AI eval case study

Inference Conference

Three AI agents built spaCy model recommenders. A fresh AI peer-review panel ran the repos, caught leakage and product-path failures, and rejected all three.

Repository Writeup

What I Care About

The theme is not bigger demos. It is tighter loops between intent, implementation, evidence, and judgment.

Agent loops Fresh context, clear specs, commits, and gates so coding agents can be useful without becoming fire-and-forget.
Evaluation that bites Measure the product path, audit provenance, and treat confident wrongness as worse than refusing to guess.
Practical software Small tools, clear interfaces, and enough rigor that the next human can understand what happened.

Find Me

The best trail is public work: repos, writeups, and the artifacts left by agent runs.

GitHub Dev.to X / Twitter