Edward Lu
SWE at Noitom, Full Stack at Better Campus, Research at Princeton
Berkeley, CAedwardlu@berkeley.edugithub.com/edrluylucv.xyz
Summary
Hi! I am a second-year student at UC Berkeley, studying Computer Science and Statistics. I build systems around models: foundational robotics models at Noitom and LoRA models on the side. I am also a full-stack engineer at Better Campus, making school easier for the 2 million students on the platform. On the side I enjoy neuro research, which I am currently doing in the Murphy Lab at Princeton.
Experience
Jun 2026–present
Noitom Robotics
- Built and deployed LensGrid, a REST API video intake platform on AWS with Docker, Amazon S3, and PostgreSQL, so motion-capture footage lands, validates, and is delivered through one path instead of by hand.
- Built LensGrid Reviewer, a console in TypeScript, React, and Next.js over a Python FastAPI backend, letting non-engineers approve datasets directly with automated QA. It cut manual review time by 40%.
- Scaled AWS S3 batch exports to 4 regions for more clients while cutting delivery turnaround from 2 days to 10 hours.
- Built a Python/NumPy pipeline that normalizes and validates 6 incompatible motion-capture formats — SMPL and BVH among them — under a single schema.
Motion capture · AWS · Docker · PostgreSQL · NumPy · data QA
May–Sep 2026
Better Campus
- Built and deployed over 11 student-facing features in my first two weeks with TypeScript, React, and Vite over WXT, across a monorepo spanning a browser extension, iframe-embedded applications, and an edge API.
- Raised active trials by 42% in four weeks by shipping live rooms, class spaces, and collaborative study workflows.
- Fixed 25 production issues across our Cloudflare REST API and PostgreSQL queries, which serve over 2 million students.
- Cut 10.4k lines of unreachable code and hardened production by pinning verification to the Supabase issuer.
- Created end-to-end tests for React components with Browserbase, raising frontend test coverage by 13%.
- Built AgentTree, which orchestrates coding agents across Git worktrees to work Linear issues in parallel and hand context between them.
TypeScript · React · WXT · Vite · Cloudflare Workers · PostgreSQL · Supabase · Browserbase
Apr 2026
Murphy Lab, Princeton
- Built a merge-suggestion pipeline for FlyWire/FAFB neuron proofreading that detects severed neurite endpoints and ranks the nearby candidate segments, rather than leaving a proofreader to find them by eye.
- Ranked neurite continuations by PCA trajectory alignment and surface gap using NumPy, SciPy, and KD-trees.
- Piloted it with 4 proofreaders over 200 endpoints, cutting median resolution time from 6.5 to 4 minutes.
Python · PyTorch · CAVEclient · CloudVolume · NumPy · SciPy
Dec 2025–Apr 2026
Operis
- Built a LangGraph legal research agent that uses context graphs to derive relationships across 4.1M documents in False Claims Act investigations, turning filings, contracts, emails, and records into searchable case knowledge.
- Shipped a React/TypeScript review console that translates complex model outputs into evidence an attorney can read and act on.
- Built LLM evals and hybrid retrieval over BM25 and dense embeddings, with graph expansion and LightGBM reranking.
- Deployed a FastAPI server with PostgreSQL + pgvector on Google Cloud for retrieval across 100+ GB of data.
Python · LangGraph · FastAPI · React · TypeScript · BM25 · pgvector · LightGBM
Projects
2026
Percept
- Compressed and deployed the TRIBE v2 model behind a Redis-backed service, so a predicted population-average cortical response came back fast enough to sit in an interactive loop, cutting inference from around [9000ms] to [2400ms].
- Built a visualization for the brain-activation matrix, along with a video-context pipeline using visual embeddings and semantic retrieval to generate hyper-realistic videos.
- Increased brain activation per video clip by approximately [10%] after each run, and improved video generation quality: of 60 people polled via Terac, [93%] preferred our videos.
Model compression · Redis · cloud deployment[repository]
2026
Manga Net
- Trained a LoRA on Animagine XL 4.0 using 15K Qwen2.5-VL-captioned manga panels, increasing human preference for hand-drawn linework over the base model by [95%] (Vote).
- Maintained character identity across [10/13]-panel pages using dual IP-Adapters and a scene-breakdown context graph, improving cross-panel identity consistency by [0.38%] on a 13-panel benchmark, measured against the base model with no LoRA.
LoRA · fine tuning · Stable Diffusion · eval
2026
LLM Royale
- Built a real-time AI agent that played Clash Royale on a physical iPhone, using dual YOLO models and Tesseract to convert live frames into structured game state and Quartz events to execute model actions.
- Developed a provider-agnostic LLM control loop with structured outputs, action validation, card tracking, and latency tracking, to benchmark OpenAI and Anthropic models under identical conditions.
- Presented synchronized gameplay and telemetry showing that model inference and tokenization consumed ~75% of runtime, with the more efficient agent winning both completed matches.
YOLO · vision · OpenCV · low-latency loops · benchmarking[game 1][game 2]
2025
cpvc
- Wrote one cross-platform wrapper over the native volume APIs on macOS, Windows, and Linux, because controlling system audio should not be three code paths in every project that needs it.
- It has 5,217 downloads across 9 releases, which is the only honest measure I have that the abstraction was the right one.
Rust · systems · cross-platform · audio[crates.io][repository]
2025
Savor
- Fine-tuned YOLO through transfer learning on a synthetic food-waste dataset generated with multimodal models, achieving [38% mAP] for food classification (with limited time to tune the model).
- Deployed the model as an on-device computer-vision pipeline on a Raspberry Pi camera, achieving [2800 ms] inference latency and enabling real-time waste detection without network connectivity.
YOLO · fine tuning · transfer learning · edge AI[repository]
2025
Beryl
- Launched a full-stack productivity app built around calendar timers, recurring sessions, Google OAuth, and persistent user state on Prisma over PostgreSQL.
- It grew to ~500 logins and 2,000+ monthly page views while I kept reworking the UI flows around accessibility and user feedback.
Next.js · TypeScript · Prisma · PostgreSQL · Google OAuth
2024–2025
Smart Racing Drones
- Built a racing drone from the ground up, carbon frame through motors, ESCs, and flight controller. Connected it to a Raspberry Pi for vision projects.
- Deployed a lightweight event-based vision model onboard for low-latency obstacle detection at racing speed, where anything that round trips to a server is already too late.
Wiring · airframe build · drone software · event-based vision
Research
More coming from Princeton…
2026
Can a Fly Brain Play Doom?
- Mapped 139,255 Drosophila neurons and 3.7M neurotransmitter-signed synapses into both a reservoir-computing model and a leaky integrate-and-fire (LIF) spiking network, then asked it to balance a cartpole.
- Ran matched controls (shuffled sign, degree-preserving rewire, random graph), because the interesting claim is not that it works, it is whether the specific wiring is doing the work or any graph of that shape would.
Connectome · LIF spiking networks · Dale's law · reinforcement learning[simulator]
2025
Hydrogen sulfide ameliorates peritoneal fibrosis
- I contributed [analysis / experiments: replace with the real contribution] on the HMGB1 and TGF-β/Smad3 signalling pathways in peritoneal fibrosis.
[paper]
2023
ATR-binding lncRNA ScaRNA2 promotes cancer resistance
- I contributed [analysis / experiments: replace with the real contribution] on lncRNA interactions in DNA repair and resistance.
[paper]
Open Source Contributions
2026
Markdown Preview Enhanced
- Fixed issue #2147: the preview was retained at registration time and then mutated after creation, so Preview Only lost its scroll position every time you left the tab and came back.
- 21 lines across 4 files, with regression coverage, because the fix is only worth having if the next refactor cannot quietly undo it.
TypeScript · VS Code extension API[pull request]
2023
MustangLib
- Rewrote the LED subsystem to set colour from either RGB or HSV, and replaced the season's hard-coded constants with a named set, so a match-day colour change stopped being a code change in three places.
Java · FRC · WPILib[pull request]
Awards
2026
YC Expo CTGT Challenge, Winner — 1 of 98
Reverse-engineering five hidden linear probes over Gemma 2 2B.
2026
Cal Hacks 12.5, Winner — 6 of 400 teams
Percept, neuro-guided video optimization.
2026
TreeHacks Honorable Mention
Savor, edge vision for food waste.
2025
Cal Hacks 11.5, Terac Challenge Winner — 3 of 150 teams
An optimization program for hiring.
before 2025