taotdm
works & undertakings

Shi-TaoChang

mapping beyond the treeline
austin, texas · in the morning dew
cross the cloudline
I About at first light

breaking bread with the stars.

interested in spatial datascience, neuro ai, & human-centered geography.

currently doing geo ai research @ GISense Lab

II Experience ground covered
01 research assistant · gisense lab mar 2026 —
gisense lab · austin, tx · nº 01
  • Reproduced and fine-tuned baseline models to generate landscapes from sound.
  • Developed and evaluated raw video & EEG/MEG/fMRI datasets.
  • Currently engineering multimodal brain-to-image models for human-centered applications.
02 backend/ml engineer · tx luminescence sep 2025 —
tx luminescence · austin, tx · nº 02
  • Developed back-end architecture for an AI-powered educational app for K–12 learning.
  • Integrated AI model with RAG to enhance app functionality and learning outcomes.
  • Engineered authentication & api endpoints for student organization listing website.
III Projects selected work
01 ecoscan mar 2026
python · yolov11 · nº 01
  • On-device waste detection app with FastAPI backend to promote recycling habits.
  • Fine-tuned YOLOv11 for waste classification — 24 recyclable and hazardous waste classes.
  • Advanced to the Woz Hackathon final round, placing in the top 30 of 500+ teams.
02 contextosolver apr 2026
word2vec · glove · fasttext · svd · nº 02
  • Compared Word2Vec, GloVe, FastText, and SVD under the same training setup to test whether algorithm choice alone changes embedding geometry.
  • Built the FastText pipeline and cosine-similarity beam-search navigation benchmark.
  • Validated navigation results across five seeded runs.
03 takemeter jun 2026
python · distilbert · huggingface · groq · nº 03
  • Built a balanced four-class dataset from 202 human-reviewed UT Austin Discord posts, then fine-tuned DistilBERT to classify academic, admissions, campus-life, and social discourse.
  • Benchmarked the fine-tuned model against zero-shot Llama 3.3 70B; the baseline won 90.3% to 71.0%, a 19.4-point difference.
  • Used confusion-matrix and error analysis to identify a topical-keyword shortcut in the fine-tuned model; validated annotation quality with Cohen’s κ = 0.869.
04 mypersona mar 2026
python · asyncio · discord.py · nº 04
  • Built a configurable real-time Discord bot that samples 30 real message examples for few-shot persona prompting and generates replies with Llama 3.3 70B through Groq.
  • Added optional Kagi translation across 250+ languages and styles, including controllable formality.
  • Packaged tokens, channel selection, persona instructions, and translation behavior through environment-based configuration.
05 utfoods jun 2026
python · sentence-transformers · groq · gradio · nº 05
  • Curated 60 review-level chunks from 11 source documents spanning Reddit, Discord, Yelp, and local reporting on UT Austin dining.
  • Built local semantic retrieval with all-MiniLM-L6-v2 embeddings and cosine distance, preserving each review as a complete, non-overlapping chunk.
  • Grounded Llama 3.3 70B responses exclusively in retrieved context, with inline citations, explicit out-of-scope refusal, and a Gradio query interface.
IV Uni base camp
ut austin · cs · co'29
01 campus involvement dirp · uls · lu
organizations
  • DiRP — Member · Coordinate sponsorship and fundraising outreach to corporate partners for a student computer-science research organization; member since August 2025 and administrator since April 2026.
  • ULS — Member
  • Longhorn Urbanists — Member
02 relevant coursework ut
classes taken at ut
  • Machine Learning for Text Analysis
  • Data Structures
  • Discrete Math
V Honors flags planted
01 ut fall urap fall 2026
research program
02 ut summer research scholar summer 2026
research program
03 research assistantship summer 2026
scholarship
  • Received $ for summer 2026 research assistantship at GISense Lab.
university honors 2025
VI Travels waypoints
01 travels
places i've been
02 recent activity via github
latest from github @taotdm
loading…
address copied