# Ethan Cai — Portfolio > Ethan Cai's portfolio. EECS student at UC Berkeley. Some things I care about: verified AI, applied AI (mostly in healthcare), model training, reinforcement learning, trying new food, and chatting it up with cool people :) Selected research, software tools, games, source repositories, papers, and live demos. - Location: Berkeley, CA - Availability: Looking for Summer 2027 internships in ML, software engineering, or product management. Also open to part-time roles during the school year. - Now: Researching verified AI for drones at UC Berkeley's Seshia Lab; Consulting for Freshworks on AI coding-assistant usage analytics, with Generative AI at Berkeley; Post-training LLMs to write more reliable clinical summaries at UCSF's Sohn Lab - Homepage: https://ethancai.com - Canonical project data: https://ethancai.com/projects.json - Resume (PDF): https://ethancai.com/Ethan_Cai_Resume.pdf - Resume (Markdown): https://ethancai.com/resume.md - Email: ethan_cai@berkeley.edu - GitHub: https://github.com/vitgruib - LinkedIn: https://www.linkedin.com/in/ethan-cai-25a775314/ ## Projects ### Metabolomics-Based ML Diagnostics of PASC Machine learning pipeline for diagnosing Post-Acute Sequelae of SARS-CoV-2 Infection (Long COVID) using metabolomic data. A proof of concept that molecular descriptors calculated using E-Dragon can work to build assay-agnostic diagnostic algorithms with disease differentiating power. Differentiates PASC from 4 of the 5 PASC-similar diseases tested (ME/CFS, Lyme disease, POTS, and IBS), with only fibromyalgia proving indistinguishable. First-author paper published in Metabolites. Category: research Technologies: Python, Machine Learning, Bioinformatics Highlight: First-author paper in Metabolites · 0.899 AUC-ROC Organization: UC San Diego / San Diego Supercomputer Center Dates: May 2024 – Dec 2025 Repository: https://github.com/Ethanc143/Metabolomics-Based-Machine-Learning-Diagnostics-of-Post-Acute-Sequelae-of-SARS-CoV-2-Infection Paper: https://www.mdpi.com/2218-1989/15/12/801 Image: https://ethancai.com/projects/metabolomics-pasc.png Page: https://ethancai.com/projects/metabolomics-pasc ### Prioritized Level Replay in RL Driving Environments Demonstrates that prioritized level replay, an automated curriculum learning method for RL, improves PPO performance in highway driving scenarios with reactive traffic. Scenes are sampled from Scenic scenario specifications into MetaDrive, scored by positive value loss, and resampled by rank. In a controlled 30-seed comparison, PLR-trained agents beat the unmodified PPO baseline in 21 of 30 seeds (sign test, p < 0.05). Category: research Technologies: Python, Reinforcement Learning, PPO, Autonomous Driving Organization: UC Santa Cruz Dates: May 2025 – Jun 2026 Repository: https://github.com/vitgruib/SIPACL Image: https://ethancai.com/projects/sipacl.png Page: https://ethancai.com/projects/sipacl ### Do Face Embeddings Predict Attractiveness? Benchmarked 11 face-embedding families (deep identity CNNs, classical methods, CLIP, DINOv2) against 3 human attractiveness-rating datasets totaling ~7,800 rated faces, using frozen embeddings and ridge regression. Found that embeddings trained to capture identity did consistently worse at predicting attractiveness than those that were not, pointing to a tradeoff between pose invariance and attractiveness predictivity. CLIP led every dataset (r = 0.94 on SCUT-FBP5500), while Fisherface—the most aggressively identity-separating method tested—finished last at a mean r = 0.32. Category: research Technologies: Python, PyTorch, scikit-learn, Computer Vision Organization: Independent research Dates: Jun 2026 – Aug 2026 Repository: https://github.com/vitgruib/AttrPred Image: https://ethancai.com/projects/attrpred.png Page: https://ethancai.com/projects/attrpred ### Vigil Grand prize (Bright Data track), Talk to Users Hackathon (SF Tech Week 2026), built in 12 hours. Checks an investment adviser's website, blog, and third-party pages for marketing that may break the SEC Marketing Rule, quotes each flag for human review, and drafts rewrites for approval. Led the project and built its two-pass, cross-referenced inference pipeline and an eval suite drawn partly from 12 real SEC enforcement orders. On 125 eval cases: F1 0.97 with 4 false flags, vs. 0.57 and 111 for a single-prompt LLM. Category: tools Technologies: Python, TypeScript, LLMs, Bright Data Highlight: Grand prize (Bright Data track) · F1 0.97 vs. 0.57 for a single-prompt LLM Organization: Talk to Users Hackathon, SF Tech Week Dates: Oct 2026 Repository: https://github.com/jadenvv/Vigil Live demo: https://marketing-rule-watch-a0747dd2.base44.app/ Image: https://ethancai.com/projects/vigil.png Page: https://ethancai.com/projects/vigil ### Grant Test Auto Built at Cal Hacks 13.0 for the Hyphenate track, after our PI talked at lab meeting about how much of his time goes to keeping funding instead of research. Since 2025, a grant can be canceled, restored by a court, and still not pay a cent, and the official records don't agree on which is true. On top of that, every purchase has to clear obscure funding rules, like no supplies bought in a grant's last 90 days, and someone has to check all of it by hand, constantly. Grant Test Auto automates that grunt work. It sits between a lab's researchers and its grant office: researchers ask for things in plain words, agents check each request and invoice against the grant's real balance and every spending rule, and one human grant manager approves each one with the reasons in front of them. "Awarded," "restored" and "paid" are three different facts, so a court order alone doesn't unfreeze a grant; only money actually arriving does. Claude reads and explains but never decides, since the rules live in the database where no model can talk its way past them. Worked on the backend, feature ideation and market research. Tested only on real cases: caught 38 of 38 freezes on 37 NIH grants and flagged 28 of 28 unallowable items from real audits. Category: tools Technologies: TypeScript, SpacetimeDB, React, Claude, Apify Organization: Cal Hacks 13.0 Dates: Oct 2026 Repository: https://github.com/vitgruib/CalHacks13.0 Live demo: https://calhacks130.vercel.app Image: https://ethancai.com/projects/grant-test-auto.png Page: https://ethancai.com/projects/grant-test-auto ### Anchor Lines Google AI-powered document intelligence app that transforms dense financial-aid documents and syllabi into clear, decision-ready summaries. Every insight stays anchored to the original source, with direct citations that make claims easy to trace and verify through a sleek, modern interface designed to support an expanding range of document types. Category: tools Technologies: TypeScript, Next.js, Gemini Dates: Jul 2026 – Aug 2026 Repository: https://github.com/vitgruib/Anchor-Line Image: https://ethancai.com/projects/anchor-line.png Page: https://ethancai.com/projects/anchor-lines ### Human Benchmark 3D Virtual reality edition of classic block-tapping memory tests utilizing blocks of various heights. Uses Firebase to record user performance and provide comparative feedback to other users of the same age bracket. Downloaded over 1,800 times on SideQuest. Category: tools Technologies: C#, Unity, VR, Firebase Highlight: 1,800+ downloads on SideQuest Dates: May 2025 – May 2026 Repository: https://github.com/Ethanc143/Human-Benchmark-3D Live demo: https://sidequestvr.com/app/45060/vr-memory-test Image: https://ethancai.com/projects/human-benchmark-3d.png Page: https://ethancai.com/projects/human-benchmark-3d ### Shambo Featured five-star extension on the Chrome Web Store. Blocks distracting sites and auto-closes tabs, with support for custom subpaths. Utilizes psychology (optionally) to fight procrastination by leaving snarky comments each time users try to open unproductive sites on their custom blocklist. Category: tools Technologies: JavaScript, Chrome Extensions Dates: May 2026 – Jul 2026 Repository: https://github.com/vitgruib/shambo Live demo: https://chromewebstore.google.com/detail/shambo-%E2%80%94-website-blocker/nnjocfponpkmpkldmkfalljmacbbjkao Image: https://ethancai.com/projects/shambo.png Page: https://ethancai.com/projects/shambo ### Toto Simple Five-star Chrome Web Store extension designed around minimizing friction, with proactive reminders, one-click snoozing, urgent-task flags, and cloud sync. Originally built around AI task organization, but it made everything cluttered. Kept as a dear lesson that AI doesn't belong in everything. Category: tools Technologies: TypeScript, React, Chrome Extensions Dates: May 2026 – Aug 2026 Repository: https://github.com/vitgruib/todo Live demo: https://chromewebstore.google.com/detail/toto-simple-todo-due-remi/glnmgmemmaedcgmnbbipllpfikpnmbjl Image: https://ethancai.com/projects/toto-simple.png Page: https://ethancai.com/projects/todo ### DADS: Distraction-Aware Driving Simulator Winning submission to the Games for Change Game Jam in the 'Own the Road' category. A driving game that warns about the dangers of distracted driving by making users contend with simulated distractions while they try to drive safely on the road. Features summaries of risky driving caused by distracted driving. Category: funsies Technologies: C#, Unity Dates: 2024 Repository: https://github.com/Ethanc143/gamejam Live demo: https://kyozen1.itch.io/dads-distraction-aware-driving-simulator Image: https://ethancai.com/projects/dads.png Page: https://ethancai.com/projects/dads ### Party Box Most party game apps are ruined by forced ads and paywalled categories. Party Box bundles party games such as charades and imposter into one app without any content restrictions. It runs offline on Android, and in any browser on a phone or laptop with no signup or install: just a link. Category: funsies Technologies: Kotlin, Android, JavaScript Dates: Nov 2025 – Feb 2026 Repository: https://github.com/vitgruib/spider-party Live demo: https://vitgruib.github.io/spider-party/ Google Play: https://play.google.com/store/apps/details?id=com.vitgrub.passandplaypartygames Image: https://ethancai.com/projects/party-box.png Page: https://ethancai.com/projects/party-box ## Resume ### Ethan Cai Email: ethan_cai@berkeley.edu Portfolio: https://ethancai.com/ LinkedIn: https://www.linkedin.com/in/ethan-cai-25a775314/ GitHub: https://github.com/vitgruib #### Education ##### University of California, Berkeley (Expected May 2028) B.S. Electrical Engineering and Computer Sciences (EECS) Prior coursework: SDSU Calculus III, Linear Algebra (4.0); UCSD Extended Studies: Deep Neural Networks (4.0). #### Technical Skills - Languages: Python, TypeScript, JavaScript, Java, C++, C#, HTML/CSS - Frameworks & Tools: PyTorch, scikit-learn, pandas, NumPy, CleanRL, React, Node.js, Firebase, Unity, Azure, GitHub Actions - Foundations: Machine Learning, Reinforcement Learning (PPO, curriculum learning), Large Language Models (LLMs), LLM Evaluation, Web Scraping, Statistical Analysis, Data Structures & Algorithms #### Publication Cai, E.; Kouznetsova, V.L.; Tsigelny, I.F. Metabolomics-Based Machine Learning Diagnostics of Post-Acute Sequelae of SARS-CoV-2 Infection. Metabolites 2025, 15(12), 801. doi.org/10.3390/metabo15120801 #### Research Experience ##### UCSF Sohn Lab (UC Berkeley URAP) (September 2026 – Present) Research Apprentice - Post-training an open-weight LLM with PPO for clinical history summarization: adapting and tuning a reliability-monitoring framework into an LLM-as-a-judge whose scores are the reward, and validating the final results against UCSF radiologists' assessments. ##### University of California, Santa Cruz (May 2025 – June 2026) Research Intern | Code: https://github.com/vitgruib/SIPACL - Implemented prioritized level replay on CleanRL's PPO in Scenic/MetaDrive highway simulations, with a disk-backed replay buffer and rank-based resampling; replay-trained agents beat baseline PPO in 21 of 30 seeds (sign test, p < 0.05). ##### University of California, San Diego / San Diego Supercomputer Center (May 2024 – December 2025) Research Intern | Code: https://github.com/Ethanc143/Metabolomics-Based-Machine-Learning-Diagnostics-of-Post-Acute-Sequelae-of-SARS-CoV-2-Infection - Built an assay-agnostic Long COVID diagnostic from metabolomic profiles; a multi-layer perceptron classifier reached 0.899 AUC-ROC and separated Long COVID from four of five clinically similar conditions (ME/CFS, Lyme disease, POTS, IBS). #### Industry Experience ##### RapDevPro (January 2025 – May 2025) Machine Learning Engineer - Built a golf-score prediction model for FanSlices, a fantasy sports betting platform, and an internal HTML/CSS/JS dashboard to test, compare, and visualize it. #### Projects ##### Vigil — AI Compliance Scanner (October 2026) Python, TypeScript | Code: https://github.com/jadenvv/Vigil - Won the grand prize (Bright Data track) at the Talk to Users Hackathon (SF Tech Week, 20 selected teams) with a 12-hour build of an LLM scanner that flags misleading SEC Marketing Rule claims in a firm's marketing. - Led the project: defined the problem and pitch, ran live interviews with target users and turned their feedback into product changes, and ramped up on SEC marketing compliance mid-event. - Took over and rebuilt from scratch a two-pass, cross-referenced inference pipeline and a harness testing it against a one-prompt LLM wrapper (F1 0.97 vs. 0.57, 4 vs. 111 wrong flags, 125 cases); added 24 cases from 12 SEC orders fining advisers. ##### Human Benchmark 3D — VR Memory Benchmark (May 2025 – May 2026) Unity, C#, Firebase | Code: https://github.com/Ethanc143/Human-Benchmark-3D - Built and published a VR block-tapping memory test on SideQuest (1,800+ downloads), with a Firebase backend that logs performance and benchmarks players by age bracket. Founded and led a 20+ member VR Club. ##### Grant Test Auto — AI Grant Office (October 2026) TypeScript, React, SpacetimeDB, Claude, Azure | Code: https://github.com/vitgruib/CalHacks13.0 - Designed and built an AI grant office (Cal Hacks 13.0) automating a lab's tedious checks against obscure funding rules (no supplies in a grant's last 90 days) and conflicting official records; built its database backend, React front end, evaluation and GitHub Actions CI/CD to Azure; caught 38 of 38 freezes on 37 NIH grants and 28 of 28 audited unallowable items. #### Honors & Awards - Talk to Users Hackathon, Grand Prize (Bright Data Track, 2026) - USACO Gold (2024) - AIME Qualifier, 3 years (2023-2025) - Games for Change Winner (2024) - National Merit Scholar (2026)