Shawn Liu

The checks run before the user sees the answer.

I'm Shawn Liu, a first-year M.S. student in computer science at Columbia. I build software around language models, and the part I care about most is what checks the model's output before a person sees it. Two of those systems, Loop and Foothold, are live. Four accepted papers from two years of research at UC Irvine. I'm looking for a Summer 2027 internship.

GitHub LinkedIn

Work

Loop · study plans onto Google Calendar

Give Loop a career goal and your free hours and it puts a study plan on your Google Calendar. Language-model nodes draft the plan; ordinary code validates and schedules it, and nothing lands on the calendar until you approve. Product prompts are pinned by hash, so editing one fails the build until it is re-pinned. Each push to main re-grades saved model outputs against fixed floors for schema validity and citation coverage. The retrieval test behind the plan's citations still runs by hand.

Foothold · which courses transfer

A California community college student enters their courses and a target UC or CSU and learns which will transfer. The verdict is deterministic over 352,024 rows of official ASSIST transfer agreements, and findings cite the agreement behind them. One Claude node drafts the appeal letter for at-risk courses, and every course code in it is checked against the findings before the letter reaches the student. I built it solo in three weeks for Stellic Pathfinders 2026, where it was featured (top 20% of 169 submissions). The agreements are 2025-26, with no refresh yet.

Crash Anticipation · dashcam collision warnings

A model that watches dashcam video and warns before a crash. I distilled a VideoMAE teacher into a 2.95M-parameter MobileNetV2-GRU that streams at 2.45 ms per frame (408 fps, RTX 5070 Ti). On average it warns 3.05 seconds before impact, with 3 false alarms in 301 no-crash videos. My first benchmark held only crashes, so a model that always predicted a crash scored perfectly, and I rebuilt it with real no-crash videos.

Arrhythmia Classifier · MIT-BIH leakage study

A course project with two classmates: we showed that a beat-wise train/test split on MIT-BIH leaks patient identity. In one run, the same 64,966-parameter CNN scores 98.36% beat-wise and 83.88% on a random patient-wise split. My search over 50,000 sampled patient assignments found a patient-wise split that keeps all six beat classes in the test set, but nothing has been trained on it yet. Many published MIT-BIH results use a beat-wise split. I'd rather report the 83.88.

Stack

Languages
Python, TypeScript / JavaScript, SQL
LLM systems and evaluation
Anthropic API, RAG, BM25, structured outputs, output validation and repair, LLM-as-judge evals, offline eval harnesses, CI eval gates, recall@k / MRR / nDCG
ML
PyTorch, scikit-learn, NumPy / pandas, CLIP, VideoMAE, YOLO11n, graph neural networks, knowledge distillation, hyperdimensional computing
Backend and web
FastAPI, Node / Express, REST APIs, PostgreSQL, SQLite, React, JWT, Firebase Auth, AWS S3
Deployment
Docker, Fly.io, GitHub Actions CI/CD, Git

Papers

All four accepted papers are from BiasLab at UC Irvine.

  1. Cross-Modal Event Encoder: Bridging Image-Text Knowledge to Event Streams

    WACV 2026 · Co-author. Built the CLIP-event contrastive integration and the attention-visualization pipeline.

  2. Integrating Symbolic and Neural Mechanisms for Adversarially Robust Hyperdimensional Computing

    ISLPED 2026 · Co-author. Ran the robustness evaluation and the symbolic-neural encoding experiments.

  3. Geometric Priors for Generalizable World Models via Vector Symbolic Architecture

    NeurIPS 2025 NeurReps Workshop · Co-author. Ran the VSA representation experiments and ablations.

  4. Optimal Hyperdimensional Representation for Learning and Cognitive Computation

    Frontiers in AI 2026 · Third author. Ran the encoding-dimension analysis and made the decoding-accuracy heatmaps.

About

Shawn Liu holding Coconut, a black-and-white cat, and Kumquat, an orange-and-white cat

At Columbia (M.S. expected December 2027) I'm also a graduate research assistant on a project affiliated with DAPLab and the ARiSE Lab. It uses SEC-bench, a benchmark of real-world security bugs for coding agents, and I check whether its developer patches fix the root cause or only stop the program from crashing. At UC Irvine (B.S. 2026) I worked in Prof. Mohsen Imani's BiasLab, where I trained a UAV carrier-landing model for the U.S. Navy. In Dr. Haleh Alimohamadi's BioIntelligence Lab I worked on structure-aware peptide prediction. Most of what I read right now is about how machine learning systems fail in production. I live with two cats, Coconut and Kumquat, who are named after fruit and are not otherwise alike.