Hi, my name is Richard Li

Video, world-model, and vision-language-action foundation models each represent the same robot actions in their own way, and it isn’t clear which of those representations actually helps a robot finish a task. I built a Python/PyTorch harness that feeds frozen V-JEPA, generative world-model, and vision-language-action representations into the same imitation-learning policy heads across 30 LIBERO tasks, so the representation is the only thing that changes. Each model is set up for 900 rollouts, with ablations over four representation depths, 64- vs. 256-token budgets, and language conditioning, plus blind and random baselines. I measure closed-loop task success and a permutation-calibrated representation-similarity score, using five-fold episode-grouped validation and a 10,000-resample paired bootstrap to quantify uncertainty.

Vertical and overhanging surfaces show up in infrastructure inspection, planetary exploration, and rescue. A successful locomotive robot must move both horizontally and vertically. Climbing involves choosing which holds to use, in what order, without ever falling. I built a Python pipeline that turns phone video of a climbing wall into a metric, gravity-aligned MuJoCo scene, and a linear-programming solver that checks whether the forces on a set of holds can keep the body on the wall, the test every stance in a climb has to pass. The solver matched the analytical load capacity within 2% in a synthetic single-contact test, by modeling force and moment balance, friction constraints, and limb limits, with diagnostics for why a stance fails. Across five synthetic holds, the scenes came in under 0.5% scale error, under 0.5° gravity-alignment error, and under 3 mm dimensional error, using photogrammetry, ArUco-marker calibration, hold segmentation, and convex decomposition. The project passes 24 tests covering calibration, contact physics, segmentation, and MuJoCo geometry.

Robot Wushu is about understanding how robots can take part in arts and performance, starting with Wushu, the Chinese martial art I practice with Columbia Wushu. I’m working on getting a humanoid robot to perform the basics of a Wushu routine. It’s early days, so more soon.

In RAG, retrieval decides which passages an LLM reads before it answers, in everything from support assistants to company search and coding tools. Energy distance has been theorized to beat cosine similarity at retrieval in early tests, but it needs a distance for every pair of embeddings, so it’s much slower. For my team project in Columbia’s High-Performance Machine Learning course, I’m working with IBM Research, mentored by Mr. Bhattacharjee, a Senior Technical Staff Member and IBM Master Inventor, to accelerate energy distance on GPUs with CUDA so it’s fast enough to use at scale. Results coming soon.

When a satellite spots a complex disaster like a flood triggering a landslide, its images usually wait for a ground station pass, ML pass, and human review before anyone responds. A fast-moving disaster doesn’t leave that time. As a machine learning intern at JPL in Pasadena, advised by Dr. Steve Chien and Jake Lee, I built a LangGraph multi-agent system of onboard Gemma VLMs that detects, triages, and plans a response to disasters on the satellites themselves, so a constellation can act without waiting on the ground. Onboard coordination cut simulated disaster-response time from 72 to 10 minutes (7.2×), measured in a Python constellation simulator I built with inter-satellite links, downlink scheduling, and Contact Graph Routing. The system reached 85% triage accuracy across a simulated 10-satellite constellation. I improved Gemma’s CLIPScore on aerial imagery by 40% by fine-tuning it with two-stage LoRA and FSDP, and the adapter and its inference harness were handed off to JPL’s flight software team for a satellite mission. When BLEU/CIDEr scores proved misleading, I traced them to caption-length confounding and built an LLM-as-a-judge evaluation pipeline (Claude on AWS Bedrock) to score model outputs instead.

Scientific figures made by an LLM are only useful if they’re valid and deterministic, which a model drawing pictures directly can’t promise. As a software engineer at AstraSomnia, I defined a typed Pydantic v2 schema that Gemini writes each figure in and built a deterministic matplotlib renderer for it, so every figure is checked before it’s drawn and the same spec always draws the same image. 98% of specs passed the schema on the first attempt (n=100), and re-renders were byte-identical. I cached rendered figures in Google Cloud Storage under a hash of each spec, so a figure is only drawn once, and shipped the renderer behind a feature flag with a regression test confirming nothing changed while the flag was off. I also built concurrency-safe session management for LLM agents in FastAPI/asyncio, deployed on Google Cloud Run with Firestore and Cloud Storage, and it sustained 100+ concurrent sessions in load testing.

Bite is a Columbia startup that delivers food from local restaurants to campus with no service fees, and its app has to keep each student’s account and data private. As a full-stack software engineer intern, I built Supabase Auth sign-in, Postgres row-level security, and trigger-based audit logging of every profile change for its React Native/Expo app, so each student can reach only their own data and every change is on record. It protects the accounts of thousands of graduate students that use the app.

Radiologists read CT scans in 3D, but most medical vision-language models that could draft their reports flatten the scan into 2D slices and lose that structure. As a machine learning intern at Microsoft Research in Redmond, advised by Dr. Ying Jin, I extended Gemma-3 with learned 3D pooling and an MLP projection so it reads the whole CT volume and writes the radiology report from it. Trained with an 8×A100 FSDP pipeline I built, it reached 0.86 BERTScore F1 on about 150 held-out volumes. I cut peak disk use 420× (21 TB to 50 GB) with a Python streaming pipeline (aria2c, PyTorch DataLoader) that downloads, preprocesses, and evicts fixed-size chunks, which made single-node training on all 21,000 CT-RATE volumes possible.

Toonsutra is a comics app serving 1M+ monthly active users, and in a comic the art changes what the dialogue means, so translating the words alone gets lines wrong that people then have to fix by hand. As an AI research intern, I helped develop image-conditioned RAG for comic translation in four South Asian languages, which gives the model each panel’s art along with its words.

Single-cell RNA sequencing shows what each cell in a tissue is doing, but a study’s conclusions are only as good as the data handling behind them, and technical noise in tens of thousands of cells can easily be read as biology. In Dr. Kristina Adams Waldorf’s lab at UW Medicine, I built a single-cell RNA-seq pipeline that removes that noise and labels cell types automatically. It became the backbone of our placenta atlas in BMC Genomics, where I’m co-first author. I filtered out empty droplets, dying cells, and doublets; normalized with SCTransform so differences in sequencing depth didn’t pass for differences in gene expression; and integrated the samples with Harmony so cells grouped by type rather than by the chip they were sequenced on. To compare cell-type proportions between groups, I used sccomp, a Bayesian compositional model that treats each sample, not each cell, as a replicate.