My Work
On a mission to apply ML to impactful domains. Every place I've worked at has challenged me more than the last.
Octane Security
Co-founder and Head of ML - 2022-2023
I founded a blockchain security startup that uses ML to find vulnerabilities in smart contracts in 2022. We raised $1.7 million from Alchemy, OrangeDAO, Hyperithm, Duke Capital Partners, Druid Ventures, Symbolic Capital, and Builder Capital, with angels and advisors from Ledger, Apple, Quantstamp, Blowfish, and Meta.
- Developed a system for traversing and classifying control-flow graphs to determine control-flow vulnerabilities in smart contracts such as reentrancy and arbitrary external calls.
- Created and tested a Transformer-based fuzzer that filtered transactions based on the likelihood of breaking a predefined invariant or bug oracle. Used this method to augment an existing greybox concolic tester and a symbolic execution engine.
- Used SFT and RLHF to finetune and steer language models into identifying vulnerable code that evaded all traditional static analyzers.
- Setup hosting and infra for all ML-based workflows for easy integration into the Octane platform.
- Left the company after helping build the core tech; the company continues to thrive.
- EthereumWorkOSOpenAIPyTorchTerraformHuggingFaceQdrantMistralAWS API GatewayAWS EC2AWS ECSAWS LambdaLinear
Advising
Machine Learning Advisor - 2023-2024
I am working as an advisor for a few companies looking to integrate AI into their existing business workflow.
- Sagiliti: Built a PoV AI-OCR system to automate extraction of data from bill copies. Using OpenAI models and clever type-checking, I am able to automate 90% of their manual extraction process.
- Project Imagine: Helping create AI agents that intelligently extract financial data to help executives make better decisions and get analysis on unfamiliar domains for clients.
Duke Neurotoolbox Lab
Machine Learning Researcher - 2021-2024
I am working in the Neurotoolbox Lab on improving segmentation of neurons in 2-photon calcium imaging videos. I am advised by Dr. Yiyang Gong and Dr. Yijun Bao.
- Developed an active learning architecture that reduced the number of labeled neurons in videos by 50x to reach SOTA accuracy and 10x for baseline accuracy.
- Created a technique to use the data from active learning to identify neurons that impact convergence; these neurons happened to be low SNR and low frequency neurons. Isolating these led to faster convergence to the baseline F1.
- Ran experiments across multiple GPUs efficiently to reduce a process that would've taken days to hours.
Paper coming soon!
Tech StackSuperbAI
Machine Learning Engineer - 2022
I worked as a machine learning engineer for SuperbAI, a company that automates data labeling for enterprise ML applications.
- Implemented an approach to estimate training data influence by tracing gradient descent.
- Extended approach to object detection for detection of false negatives in human-labeled datasets. Treated mislabeled instances as its own class based on mixed confidences from its original labeled classes.
- Detected over 50% of false-negatives and reported them to human labelers as a form of feedback.
- Worked mostly with self-driving datasets with varying environments (weather, cities, time of day, etc).
AiFi
Computer Vision Team, Summer Intern - 2021
I worked at AiFi, a company that uses a camera-based computer vision system to automate retail stores. They have clients such as Zabka, Microsoft, Verizon, and various NFL, NBA, and Premier League stadiums.
- Spearheaded an effort to use domain-randomized synthetic data to train product recognition algorithms.
- Developed a product auto-labeling pipeline based on instance segmentations from simulation data.
- Generated photorealistic data from unpaired real and simulation images using cycle-consistent GANs.
- Used synthetic data to pretrain YOLOv5 and a small set of real examples for SFT; this reduced store deployment time from 2 weeks to 2 days and increased new SKU detection accuracy by 80%
Stanford University School of Medicine
Research and Development Intern - 2019-2021
I interned at the Bacchetta Lab in the Stanford School of Medicine Department of Pediatrics. I worked with PhD student Esmond Lee.
- Designed experiments to increase the NGFR+ percentage (editing rate) for FOXP3 gene editing in hematopoietic stem progenitor cells
- Used a design of experiments (DoE) approach to optimize CRISPR-Cas9 editing of HSPCs. Analyzed flow cytometry data using FlowJo.
- Created a cost analysis to show a reduction in cost of reagents for applications in clinical settings.
- Published a manuscript in the Cytotherapy Journal