Last login: Fri Jul 10 2026 · The Colony, TX
adrian@utd:~$ whoami

Adrian Hautea
AI systems · backend infra · LLM apps

# CS undergrad @ UT Dallas — B.S. Dec 2026, Fast Track M.S. May 2028, taking graduate coursework this fall. RL research for LLM code generation, production platforms for 800 students. Open to summer 2027 SWE / ML internships.

./resume.pdf github.com/AdrianHautea linkedin/adrianhautea
adrian@utd:~$

adrian@utd:~$ ls projects/ # click a row to expand

yearprojectstack
2026 Ascend · RL for LLM code generation DeepSeek Coder · GRPO · HuggingFace poster ↗ expand
› Improved DeepSeek Coder 1.3B from 5.9% → 14.0% pass@1 (6.0% → 16.0% pass@10) on Microsoft P3
› Prototyped the synthetic variant-generation and difficulty-filtration stage feeding the curriculum
› Implemented training-pipeline components in PyTorch + HuggingFace; primary presenter at the ACM Research competition
2026 FSA Platform · org infra for 800 students Next.js · TypeScript · Supabase live ↗ expand
› Architected a 10-table Postgres schema: membership, attendance, events, payments, officer permissions
› Built 21 REST endpoints across 17 route modules in Next.js/TypeScript with role-based auth
› Sole engineer on the live platform, coordinating 6-7 collaborators on design and onboarding
2026 Sensora · real-time study-focus monitor [best beginner project] OpenCV · CLIP · Flask · LangChain · ESP repo ↗ expand
› Built the CLIP-embedding emotion/engagement classifier + weighted, smoothed focus-scoring logic
› Wrote the LangChain + Gemini engine turning aggregated focus signals into concise study nudges
› Integrated the vision pipeline, Flask API, and serial/ESP output into one live loop
2025 Simmer · hands-free voice cooking assistant Flask · LangChain · Supabase · RN expand
› 10+ REST endpoints for recipes, sessions, and voice navigation
› TikTok / YouTube / blog import pipeline → structured recipe data via Gemini
› LangChain voice agent answers cooking questions mid-recipe
2024 ThreadConnect · multimodal clothing search CLIP · Qdrant · Flask · React expand
› CLIP embeddings for image/text similarity search
› Preprocessing pipeline: resize, normalize, tokenize before encoding
› Vector storage + fast retrieval in Qdrant, wired to a Flask-powered UI

adrian@utd:~$ git log --oneline roles/

a8f3c21 Webmaster / Technical Lead · UTD Filipino Student Association HEAD → mainled technical direction · coordinated 6–7 collaborators · platform for 700–800 members May 2026 – present
5d91e07 Research Team Member · ACM Research at UT DallasRL for LLM code generation · primary presenter at the ACM Research competition Feb–Apr 2026
c47ab92 Projects Participant · ACM at UT Dallasbackend on a five-person team · Agile, code review, iterative delivery Sep–Dec 2025
19f0d5e Projects Participant · AI Society at UT Dallasconcept to launch in 11 weeks · presented to faculty and mentors Sep–Dec 2024
0000001 init: enroll BS Computer Science · UT Dallas Aug 2024

adrian@utd:~$ cat about.md

headshot.jpg Adrian Hautea

# Why I build

I got into CS to build things people use, and stayed once I found the machine learning. Over two years that's looked like a few different things: on a four-person research team, I worked on the reinforcement learning that pushed a 1.3B-parameter code model from 5.9% to 14.0% pass@1. I built a voice assistant that walks you through cooking a meal hands-free. And I was the sole engineer on the web platform behind an 800-member student org.

interests[0] = "FSA community" interests[1] = "cooking w/ simmer" interests[2] = "ML research" interests[3] = "hackathons"

adrian@utd:~$ cat education.txt

The University of Texas at Dallas
M.S. Computer Science (Fast Track) · expected May 2028
B.S. Computer Science · Dec 2026
Currently taking graduate coursework toward the M.S.
# coursework: advanced algorithm design & analysis · database systems · operating systems ·
# intro to machine learning · computer architecture · UNIX systems programming · software engineering · linear algebra
3.94
GPA · dean's list

adrian@utd:~$ printenv SKILLS

languages
PythonJavaJavaScriptTypeScriptC/C++SQL
ml / ai
PyTorchOpenCVMediaPipeLangChainGeminiOpenAI CLIPHuggingFacescikit-learnGRPO / reinforcement learning
backend / infra
FlaskNode.jsReactReact NativeNext.jsSupabasePostgreSQLQdrantStripeAWS S3Vercel
tools
GitGitHubPostmanGoogle Colab

adrian@utd:~$ mail -s "internship" adrian.hautea@gmail.com

# Looking for summer 2027 SWE / ML internships. If you're building intelligent systems at scale, I'd love to hear about it!

send email (817) 513-0850
less projects/fsa.md

# FSA Platform (2026)

## Webmaster / Technical Lead · UTD Filipino Student Association

The FSA needed real infrastructure: membership, events, attendance, dues, and officer workflows for 700–800 students. I was the sole engineer — a 10-table Postgres schema, 21 REST endpoints, and 27 pages — coordinating 6–7 collaborators on design and onboarding. Live on a purchased domain, launching to members in August 2026.

10
Postgres tables
21
API endpoints
27
pages shipped
## what i did
adrian@utd:~$ cat fsa/architecture.txt
client · browser
Next.js 16 · App Router React 19 TypeScript Tailwind CSS v4 FullCalendar · html5-qrcode
RSC / server actions · Zod-validated requests
application · vercel
Route Handlers & Server Actions Zod validation qrcode gen
data & auth
Supabase · Postgres OAuth
server-side integrations
external services
Stripe
Checkout + webhooks
AWS S3
media storage
Resend
transactional email
Vercel Analytics
+ Speed Insights · CI/CD
stack: Next.js · TypeScript · Supabase
./live-site
less projects/ascend.md

# Ascend (2026)

## Research Participant · ACM UTD

Five-person team testing whether curriculum learning — proven on mathematical reasoning — transfers to code generation. We had DeepSeek Coder 1.3B generate its own difficulty ladder from the Microsoft P3 puzzle dataset, then trained up it with GRPO using the Python interpreter as ground-truth reward. Pass@1 went from 5.9% to 14.0%, pass@10 from 6.0% to 16.0%.

5.9% → 14.0%
pass@1 after GRPO
2.4×
improvement over baseline
1.3B
DeepSeek Coder params
## what i did
adrian@utd:~$ cat ascend/pipeline.txt
01 · dataset: selecting & generating problems
P3 Dataset
Problem 1 … Problem N
take one problem
Problem
generate synthetic variants
Candidate Problems
Synthetic Problem 1 … N
Difficulty Estimator
bucket by pass-rate
30% group
easiest · RL entry
70% group
mid difficulty
100% group
original problem
02 · GRPO: the reinforcement-learning loop
30% Difficulty Group
take one problem
Input Problem
Model
DeepSeek Coder 1.3B
Generated Solution
Python Interpreter
assert sat(sol)
reward signal
GRPO Trainer
model updates
stack: DeepSeek Coder · GRPO · HuggingFace
./research-poster.pdf
less projects/sensora.md

# Sensora (2026)

## Backend / ML · Hackathon, team of 4 · Best Beginner Project

A webcam watches you study, a computer-vision pipeline scores how engaged you are moment to moment, and an LLM turns those signals into short, actionable study nudges pushed to an ESP hardware device over serial. I owned the backend intelligence layer: vision-to-focus pipeline, the LangChain + Gemini recommendation engine, and the integration between vision, Flask, and the serial/ESP output.

6
engagement states
≤12
word LLM nudges
~2,100
labeled training images
## what i did
adrian@utd:~$ cat sensora/how-it-works.txt
Webcam
OpenCV pulls frames
Vision pipeline
MediaPipe face + landmarks · CLIP engagement · gaze + head pose
6 states → ENGAGED / NOT
Focus score
weighted · smoothed over a rolling window
POST per second
Flask backend
rolling sample history
summarize recent samples
LLM recommendation
Gemini via LangChain · ≤12-word nudge
over serial
ESP device
physical display / feedback
adrian@utd:~$ cat sensora/limitations.txt

The engagement classifier trained on a ~2,100-image dataset — too small. It overfit, and held-out accuracy wasn't where I'd want it for real use. With more time: augmentation, proper cross-validation, and fine-tuning the CLIP encoder rather than freezing it.

stack: OpenCV · CLIP · Flask · LangChain · ESP
./github.com/AdrianHautea/Sensora
less projects/simmer.md

# Simmer (2025)

## Backend · ACM UTD Projects, team of 5

A mobile app that guides you through cooking without touching your phone: reads recipes aloud, listens for commands, answers questions mid-recipe. I owned the backend: endpoints, database, media import pipelines, and the voice agent.

10+
REST endpoints
-80%
manual recipe input
3
import platforms
## what i did
adrian@utd:~$ cat simmer/architecture.txt
React Native app
voice commands · Expo Go
REST requests
Flask API
10+ endpoints · recipes · sessions · voice nav
Import pipeline
TikTok / YouTube / blog → Gemini → structured recipe
Voice agent
LangChain · answers cooking questions mid-recipe
persist
Supabase
relational schema · SQL
stack: Flask · LangChain · Supabase · RN
less projects/threadconnect.md

# ThreadConnect (2024)

## Backend / ML · AI Society @ UTD, team of 5

A search tool for ethically-sourced clothing alternatives: drop in a photo or a description, get visually similar items back. Built the CLIP embedding pipeline and vector search; concept to demo in 11 weeks.

CLIP
ViT-B/32 embeddings
11
weeks to launch
5
person team
## what i did
adrian@utd:~$ cat threadconnect/clip.txt
Images
product photos
Image Encoder
CLIP → I₁…I₄
Descriptions
text captions
Text Encoder
CLIP → T₁…T₄
T₁
T₂
T₃
T₄
I₁
I₁T₁
I₁T₂
I₁T₃
I₁T₄
I₂
I₂T₁
I₂T₂
I₂T₃
I₂T₄
I₃
I₃T₁
I₃T₂
I₃T₃
I₃T₄
I₄
I₄T₁
I₄T₂
I₄T₃
I₄T₄
Diagonal = matched image/text pairs. Embeddings stored in Qdrant; a query vector retrieves nearest matches.
stack: CLIP · Qdrant · Flask · React