Skip to content

David O.

Austin, TX

I build evaluated AI systems, agentic developer tools, and interactive engineering software.

UT Austin Computer Science
'28

UT Austin Computer Science

About

UT Austin computer science student minoring in Statistics & Data Science. Recent work spans QLoRA model routing, deterministic verification, Rust engine extensions, and deployed TypeScript products. I care about explicit limits, measurable gates, and useful interfaces.

Education
UT Austin, BS Computer Science
Minor
Statistics & Data Science
Based in
Austin, TX
Focus
SWE · Applied AI · ML/LLM Systems
Languages
English · Spanish (fluent)

Experience

Alpha AI

Jun 2026 to Aug 2026

AI Engineering Intern

Built evaluated applied-AI and learning systems across model routing, native study tooling, deterministic learner modeling, and a public audio course.

  • GT100K: served as primary engineer on a two-person, synthetic, pre-production platform spanning 28 packages, 16 adapters, and 9 apps, with deterministic Beta-Bernoulli interest inference and human-controlled transitions.
  • Small Learning Model: fine-tuned Qwen3-1.7B with QLoRA around a calculate/defer/abstain/explain contract and an 80-function Formula Core.
  • Speedrun: extended Anki's Rust engine with six GRE Math RPCs and shared Svelte experiences for early Windows and Android builds.
  • Blazing Audio: shipped 13 lessons, 50 graded problems, 24 interaction types, Leitner review, and five authenticated OpenAI Cloud Functions.
  • Short-video MVP: built a React/Firebase prototype with transactional handles, a real-time scroll-snap feed, social actions, and telemetry.
TypeScriptPythonRustSvelteFirebaseQLoRA

Featured Work

  1. 2026

    Live browser enclosure-engineering workstation used by 1,000+ users and growing, connecting acoustic simulation, physical geometry, Three.js visualization, warnings, charts, cut sheets, nesting, persistence, and sharing.

    • Next.js
    • TypeScript
    • Three.js
    • Vitest
    Technical details
    • One normalized design model drives sealed, ported, passive-radiator, and release-gated bandpass calculations plus fabrication geometry.
    • WebGL and geometry work use independent computation fingerprints, demand rendering, a persistent canvas, and adaptive worker offload.
    • Its current Vitest unit/parity gate passes 5,068 cases across 422 files.

  2. Small Learning Model

    development-only

    2026

    Fine-tuned Qwen3-1.7B with QLoRA to route audio questions across calculate, defer, abstain, and explain, with accepted calls executed by an 80-function Formula Core.

    Dataset
    • Qwen3
    • QLoRA
    • PyTorch
    • Python
    Technical details
    • A historical V7.2 tuned-model-plus-executor evaluation scored 94.4% versus 30.7% for base on 319 frozen value-and-unit items.
    • V7.4 reached 91.9% numeric accuracy but remained development-only after failing strict family and behavioral release gates.
    • As of August 2026, the public V7.4 development checkpoint had 434 all-time model downloads (400+ downloads) and its companion dataset had 54 all-time dataset downloads (50+ downloads).
    • The pipeline includes deterministic data generation, fail-closed tool compilation, hash-pinned training, Gradio inference, and a 387-row evaluation manifest.

  3. GT100K

    pre-production

    2026

    Primary engineer on a two-person, synthetic education platform organized as 28 packages, 16 adapters, and 9 apps behind deterministic ports.

    • TypeScript
    • React
    • PostgreSQL
    • PGlite
    Technical details
    • Implemented Beta-Bernoulli learner-interest inference with evidence decay, normalized negative evidence, uncertainty gates, and distinct-day requirements.
    • Kept consequential hypothesis transitions under human authority and excluded duration and performance from interest-belief updates through a type-enforced signal firewall.
    • Encoded deny-by-default consent, child-safety constraints, evidence-class governance, and opt-in LLM adapters while preserving deterministic offline behavior.

  4. Agentic Software Factory v1

    development-only

    2026

    Cross-project agentic software factory built around bounded Claude and Codex loop harnesses, self-QA agents, deterministic gates, recovery, browser QA, and pull-request delivery.

    • Python
    • Bash
    • Claude
    • Codex
    Technical details
    • Runs repeated builder-review loops in isolated worktrees with file-backed cross-turn state, time and no-progress caps, and stranded-run recovery.
    • Self-QA agents combine semantic browser walks, screenshot and pixel checks, model and deterministic graders, and specialized review panels.
    • Used across GT100K and Subwoofer Central with host-side Git controls and pull-request delivery; the repository remains private.
  5. 2026

    Public interactive audio course with 13 lessons, 50 graded problems, 24 interaction types, concept-level Leitner review, and real-time learning tools.

    • React
    • TypeScript
    • Firebase
    • OpenAI
    Technical details
    • A typed content registry drives deterministic local grading, progress, prerequisite warm-ups, and interleaved review.
    • Five authenticated OpenAI Cloud Functions handle generated review, semantic judgment, tutoring, and capstone evaluation through strict schemas and private rubrics.
    • AI grading falls back to deterministic multiple choice when unavailable; the live product does not call the custom audio SLM.
  6. 2026

    Co-developed Firebase and Stripe marketplace for browsing and purchasing verified college-application content.

    • Next.js
    • Firebase
    • Stripe
    Technical details
    • Hardened purchase-gated content delivery, administration workflows, API rate limits, and deployment reliability across the Firebase and Stripe marketplace.

  7. NeuroBaseline

    prototype

    2025

    Co-developed a non-diagnostic prototype using Isolation Forest and PELT to model baseline changes in simulated typing-behavior data.

    • Python
    • scikit-learn
    • FastAPI
    • Next.js
    Technical details
    • Combined anomaly detection and change-point analysis into a personalized Neuro Variability Index.
    • Won the Synovate Hackathon first-place Grand Prize among 61 participants.

  8. Speedrun

    early release

    2026

    Extended Anki's Rust engine with six GRE Math RPCs, exam-aware queue behavior, shared Svelte experiences, and early Windows and Android builds.

    • Rust
    • Svelte
    • Kotlin
    • FastAPI
    • SymPy
    Technical details
    • Custom layers add uncertainty-aware Memory, Performance, Readiness, calibration, and scheduling hooks while preserving inherited FSRS due dates and Anki sync.
    • The optional FastAPI/LangGraph generator grounds proposals with hybrid retrieval and requires SymPy verification before emission.
    • Offline checks reverified 50/50 gold answers, rejected 6/6 invalid specifications, and measured a 0.900 Recall@10 tie across BM25, dense, and hybrid retrieval.

More Projects

05 entries
  1. 2026

    Founded Virgilio Acoustics, an Austin audio installation business that has served 10+ clients with $14,000 in revenue ($5,000 in net profit), and built its public site and administration workflow.

    • Next.js
    • TypeScript
    • Neon
    • Vercel Blob
    Technical details
    • Served 10+ Austin clients across custom stereo, home-theater, and high-output audio installations, generating $14,000 in revenue ($5,000 in net profit on $9,000 in costs).
    • Combined a Next.js and TypeScript marketing site with Vercel Analytics and a password-protected Neon and Vercel Blob administration workflow, reaching 500+ site visitors.

  2. UTMAX

    live

    2026

    Ongoing contributor to the team-built UT Austin schedule-planning platform, used by 200+ UT Austin students before the 2026-27 school year began, including an 18-file UI revamp.

    • Frontend
    • Responsive UI
    Technical details
    • The 18-file contribution tightened frontend presentation and interaction quality across the shared product.

  3. UTMap

    live

    2026

    Ongoing co-development on the team-built UT Austin campus map, used by 200+ UT Austin students before the 2026-27 school year began, with a focus on mobile interaction and map readability.

    • Frontend
    • Geospatial UI
    Technical details
    • Worked on mobile UX, label clustering, overlay handling, and shuttle styling.
  4. 2024

    Co-developed a searchable Next.js, React, and Supabase product spanning roughly 400 curated extracurricular opportunities.

    • Next.js
    • React
    • Supabase
    Technical details
    • Contributed responsive search and discovery flows around the shared opportunity catalog.
  5. 2026

    React and Firebase prototype with transactional unique-handle signup, a real-time scroll-snap feed, social actions, and watch telemetry.

    • React
    • TypeScript
    • Firebase
    Technical details
    • Implemented the central consumption loop, optimistic likes and favorites, live comments, and append-only engagement events.
    • The prototype did not include uploads, personalized recommendations, Cloud Functions, or a moderation interface.

Toolkit

Languages

  • Python
  • TypeScript/JavaScript
  • Rust
  • SQL
  • C/C++
  • Go
  • Java

AI and ML

  • PyTorch
  • Transformers
  • QLoRA/PEFT
  • Hugging Face
  • scikit-learn
  • LangGraph
  • OpenAI API
  • SymPy
  • NumPy/SciPy

Web and Systems

  • React
  • Next.js
  • Svelte
  • FastAPI
  • Three.js
  • Firebase
  • PostgreSQL/Supabase

Testing and Tools

  • Vitest
  • Pytest
  • Playwright
  • Git
  • Linux/WSL
  • Vercel

Benchmark

Beat 200 WPM

Type the passage below. Accuracy counts, so only correct characters score.

WPM

0

Accuracy

100%

Best

··

Type the following text: program follow general child present then year right through just also take

Click to start

0/75 characters