dhruva bhatbuilding backend and infra
no light mode (why would you try this...)
work
Oak Ridge National Lab
AI platform engineering intern · May 2026 – July 2026
An asynchronous AI pipeline that generates scientific candidates, built to stay up across long multistage workflows.
- Engineered an asynchronous pipeline in FastAPI that orchestrates Google CoScientist, multi-agent tournament reasoning, retrieval-augmented generation, and molecular diffusion models for candidate generation.
- Hardened the modular API services with asynchronous task execution, retries, fallback strategies, caching layers, and service orchestration.
- Containerized every service with Docker to standardize dependencies and run the same stack across environments.
FastAPI · Python · async I/O · RAG · Docker
Blueprint at Berkeley
Full-stack and algorithms developer · Feb 2026 – present
A cloud volunteer platform and a routing algorithm for Amigos de Los Rios.
- Built and deployed a cloud-based volunteer application with React, Next.js, Supabase, and AWS Lambda, including authentication workflows and live operational dashboards for admins.
- Designed a geospatial routing algorithm that joins REST API data from PlanItGeo with PostgreSQL to plan volunteer routes around time, location, and resource constraints.
- Ordered the routes with time-based waterfall prioritization.
React · Next.js · Supabase · PostgreSQL · AWS Lambda · REST APIs
Carnegie Mellon University
AI safety research assistant · Mar 2025 – present
Backend infrastructure for LLM experiments, and research into how rationally LLMs behave inside institutions.
- Developed a modular backend pipeline that orchestrates LLM experiments across multiple model environments, tying together inference APIs, experiment configuration, and automated analysis.
- Researching the bounded rationality of LLMs in institutional settings with Professor Sarah H. Cen, benchmarking model behavior against game-theory baselines.
Python · inference APIs · experiment orchestration
UC San Francisco
Machine learning / bioinformatics research assistant · Aug 2025 – present
Unsupervised machine learning on genomic data to find the best lung cancer gene markers.
- Implemented an unsupervised pipeline in R covering data cleaning, feature selection, hierarchical clustering, PCA, network-based outlier detection, and batch-effect correction.
- Ran it on 125,000+ patient data rows, comparing gene marker quantity with cell type prevalence across tissue samples, with Professor Oldham.
- Building a data scraper and pipeline for GRE data, and a neural network that compiles analysis across studies to surface the best markers for each cell type.
R · PCA · hierarchical clustering · neural networks
Garcia Center for Polymers
Researcher, Stony Brook University · Jun 2024 – Dec 2024
Machine learning for materials science, across thin films and hydrogels.
- Thin films: used 4D manifold learning in PyTorch and built a Gaussian process regression model relating molecular weight to film thickness, predicting commercial polystyrene semiconductor wafer thickness.
- Hydrogels: ran rheological analysis of F88DMA cross-linked with APS and TEMED, and found a 97% correlation between concentration and crosslinking time across conditions.
- Wrote and presented “Machine Learning for Assessing Impact of Polydispersity on Thin Film Thickness.”
PyTorch · Gaussian process regression · manifold learning · rheology
EEG Vector Database
Project
Vector embeddings that tell Parkinson’s EEGs apart from healthy ones.
- Used the Wilson–Cowan model to generate excitatory–inhibitory patterns that replicate the EEGs of Parkinson’s and healthy patients.
- Transformed the EEGs into vector embeddings with CEBRA and stored them in a vector database.
- Trained an offset-10 model to identify differences between Parkinson’s and healthy patient EEGs.
Python · CEBRA · ChromaDB
Opioid-Overdose Prevention Device
Provisional patent, U.S. 63/657,693
A modular device that dispenses naloxone before an overdose turns fatal.
- Dispenses naloxone based on a patient’s respiratory rate, heartbeat, and other physiological signals.
- Delivers it with a proprietary microneedle dispensing method.
- Currently developing the companion full-stack app.
Hardware · physiological sensing · full-stack app
about
hi, i’m dhruva, an eecs student at uc berkeley. i’m a swe specifically interested in backend and infra development, particularly integration with ai systems (check out my work section!).
in my free time, i like to fiddle with my espresso machine, try new food spots, and read the nyt!