UIUC Computer Science
Ask my portfolio
Client-side hybrid search (BM25 + dense rank fusion) across my portfolio.
Press / to search.
Two retrievers score every chunk independently. BM25 is sparse keyword matching over an inverted index — good at exact terms, blind to wording it has not seen. Dense is latent semantic: the corpus TF-IDF matrix is factored by truncated SVD at build time, and the right singular vectors ship as a term-to-latent projection, so your query is embedded here by a matrix multiply rather than a model download or an API call. That is why it answers in a fraction of a millisecond and works offline.
The two rankings are merged with reciprocal rank fusion,
score = Σ 1/(60 + rank), which is the same fusion strategy
behind the sub-50ms clause retrieval in the DIGIIN work below. Query
terms are expanded before scoring, so kubernetes reaches the
Docker line; expansion matches are the dotted underlines.
I’m Duy Anh, a CS student at UIUC. I build software and explore what it takes for AI agents to do useful work reliably. Away from the keyboard, I’m a national-record swimmer and a classical-piano nerd working through Chopin and Beethoven.
B.S. Computer Science
GPA 3.95 / 4.0 · Dean's List (all semesters)
Activities: ACM SIGRobotics · ACM SIGICPC
Selected Coursework: Applied Parallel Computing, System Programming, Database Systems, Numerical Methods
Informatics (CS specialization)
SAT 1560 / 1600
All the tools (still counting) I use to build fast, intelligent software.
Software Engineer InternDIGIIN – Software Engineering Lab, VNU-HCM
Engineered hybrid Qdrant dense & BM25 clause retrieval (<50ms across 680+ policies) using the same rank fusion architecture that powers this portfolio's search bar. Accelerated an insurtech assistant on vLLM (Qwen3-14B) by 40% via async tool calling and closed-loop error recovery.
Research AssistantSTAPL · Parasol Lab @ UIUC
Optimized container runtime performance for the STAPL parallel C++ framework in HPC environments. Profiled proxy-layer indirection with GDB & Valgrind/Callgrind, cutting indirection overhead to achieve a 5x benchmark speedup on subscript access.
Software Engineer InternThe Booker
Architected the event-driven backend for a multi-tenant AI receptionist, offloading Twilio & Meta webhooks to async workers via Redis Streams with tenant session isolation, message redelivery, and a strict <2s SLA.
Take a look at some of the passion projects I've built.
Personal Context Engine · Screen Telemetry · Retrieval · Multimodal Ranking
Built an ambient personal context assistant that observes active screen workflow, surfaces relevant notes, past conversations, and collaborator context, and floats draggable cards for one-click prompt augmentation in ChatGPT, Claude, and Gemini.
College Admissions & Decision Analytics · React · TypeScript · FastAPI · PostgreSQL
Built a full-stack college admissions analytics platform with React, TypeScript, FastAPI, and normalized PostgreSQL featuring filterable decision dashboards and user-scoped data isolation.
Text-to-Sign Generation · PyTorch · Seq2Seq · VQ-VAE · SMPL-X
Engineered a text-to-sign language generation system at the Software Engineering Lab, VNU-HCM, adapting pretrained seq2seq models with a motion tokenization pipeline that discretizes continuous SMPL-X pose sequences into body- and hand-specific token streams, decoded autoregressively with retrieval-augmented sign-dictionary lookups.
3D mesh reconstruction · original (blue) vs. tokenizer reconstruction (red)
Agentic Machinery Inspection · Google ADK · FastAPI · WebSockets · PyTorch
Built an agentic heavy machinery inspection assistant via Google ADK and FastAPI using plan-and-act workflows and closed-loop tool reflection to heal malformed payloads.
Conversational Ordering Assistant · Python · AsyncIO · REST APIs · JSON Schema · pytest
Built an asynchronous REST API and state layer for an automated food ordering assistant, enforcing idempotency and strict JSON Schema validation across webhooks with <200ms latency and 95%+ test coverage.
Low-Resource Translation Pipeline · Python · PyTorch · BERT · Transformers · TF-IDF
Developed a low-resource translation pipeline in PyTorch using a frozen BERT encoder and Transformer decoder, boosting BLEU from 12.0 to 63.48 over the TF-IDF baseline.
Semantic Video Retrieval · CLIP · FAISS · TransNetV2
Delivered sub-second semantic video search across 200+ hours of footage by extracting keyframes with TransNetV2 and indexing CLIP embeddings with FAISS.
More than side projects — two organizations I co-founded to change how gifted students in Vietnam learn, compete, and lift each other up.
Built a competitive programming community grown to a 100-member organization, coordinating a 30-person team across 8 contests, an in-person finale for ~100 students (now in its 3rd edition), and live streams with 7,000+ views.
Jun 2023 — Oct 2024 · Student Org · 300-person
Scaled a 300-person student organization across Saigon, raising $8K+ in sponsorships and organizing academic workshops and mock exams with 300+ attendees per event.
I'm always open to high-impact collaborations.
Or directly email me at dahong2@illinois.edu
