Ethan Varghese
I’m a CS major at UC Berkeley and love building cool things!
I’m interested in RL post-training, AI safety, and startups. Reach out at ethanv@berkeley.edu if you’re doing something cool and need my help, or want to hear about me.

Experience
Software Engineer Intern, Starshield
Software Engineer Contractor
- Built a full-stack analytics platform (React, FastAPI, Supabase) that tracks how 100+ Berkeley students build agents.
- Surfaced ML-driven insights on the student experience and the pitfalls students hit most.
- Ran two live activation sessions and delivered the findings to Replit’s education leadership, informing product decisions on the agent-building experience.

Software Engineer Intern, Subsurface and Wells Data Modernization
- Built a GitHub Copilot-like agentic platform for 200+ engineers, deployed across Chevron’s full petrophysics department.
- It lets engineers index petabytes of subsurface data (PDFs, LAS, DLIS, PPTX), then query well log structure, draw conclusions and generate visualizations in natural language.
- Cut the time spent understanding well data from 80+ hours to 1 to 2 hours.


Software Engineer, contract
- Developed an NLP pipeline (TF-IDF, clustering, sentiment analysis) to analyze 100+ employee reviews from Amazon Fulfillment Centers.
- Uncovered the key drivers of burnout and attrition.
- Built Python and SQL pipelines to surface ML-driven insights on workforce sentiment and retention.


Machine Learning Engineer, contract
- Engineered an AI-powered multilingual research agent for 1,000+ UN journalists across 20 languages.
- Built the full stack and deployment pipeline (Next.js, TypeScript, Tailwind CSS, FastAPI, Supabase, Vercel), delivering a production-ready tool for global newsroom use.


Artificial Intelligence Researcher
- Contributed to RECITE, a RAG pipeline for automated clinical trial eligibility revision.
- Built an 11,913-instance benchmark from ClinicalTrials.gov and benchmarked 8 LLMs, reaching 85.8% binary equivalence.
- Created an agentic PubMed crawler that identified 44 paper-to-trial matches, with accrual gains of up to 246%.
Software Engineer Contractor
- Built a full-stack invoice automation platform (React, TypeScript, FastAPI) for content creator workflows, with two-way Airtable sync.
- Automated PDF contract generation with ReportLab, cutting processing time 35% across 10K+ contracts a year.


Engineering Intern
- Built a real-time outage detection system under Dr. Steven Low that watches campus eGauge meters for voltage deviations and data dropout.
- Automated SMS alerts for faculty and validated the system against 8 historical events, with a 95% detection rate.


Software Engineer Intern
- 2022: Deployed an AI picking robot (YOLO, LiDAR, RGB-D, semantic segmentation) for real-time object detection, demoed at the Intel Partner Showcase.
- 2023: Engineered an autonomous sanitizing robot (OpenCV, PyTorch) with 94% object classification accuracy across 10K+ sq ft.
Projects
Research in progress
Alignment-preserving post-training recipes
Can RL fine-tuning keep its gains without the alignment tax?
Research in progress. A harness that fine-tunes an aligned model with GRPO on a task built to tempt misbehavior, then measures fabrication, grader gaming, truthfulness, bias and refusals at every checkpoint, so nine training recipes can be compared at matched capability. No results reported yet.
Paper replication
Scaling laws, replicated on a laptop
Do the scaling laws for language models hold on a laptop?
A small-scale replication of Kaplan et al. (2020), Scaling Laws for Neural Language Models, on an Apple M4. Thirteen GPT-style models, from 25K to 0.8M non-embedding parameters, were trained on C4. The paper’s qualitative findings reproduce, with exponents in the same range but mostly lower. It is a sanity check of the general picture, not an independent proof.
Stanford CS336
Transformer from scratch
A Transformer language model written layer by layer in PyTorch
Built for Assignment 1 of Stanford CS336. A decoder-only Transformer language model written from basic tensor operations instead of prebuilt layers: linear and embedding layers, RMSNorm, a SwiGLU feed-forward network, rotary position embeddings, scaled dot-product attention, causal multi-head self-attention, and the pre-norm blocks that stack into the full model.
Hackathon finalist
Ramp Model Labs
One prompt in, a cheap specialist model out
An automated ML lab that benchmarks classical, fine-tuned (LoRA/QLoRA), and frontier models side by side for financial tasks, then deploys the cheapest model that meets the accuracy threshold. A planner, curator, trainer, judge and deployer run the whole lab from one plain-English prompt.
Prototype
Gitty
A hiring platform where real pull requests replace the interview
A prototype built with a collaborator. Engineers practice on open issues from real repositories and interview by shipping pull requests. A FastAPI evaluator pulls the issue and pull request context from GitHub and asks Claude for a structured score and feedback.
Front end on GitHub (opens in a new tab)Evaluator on GitHub (opens in a new tab)
Active startup
CollegeStory.ai
Agents that build a student’s story and find opportunities that fit
An agentic B2B platform for college counseling firms that autonomously builds a student’s narrative and discovers personalized opportunities. Launched across three counseling firms that serve 10K+ students.
Startup, 2022 to 2025
LingoSage
AI dubbing that turns a video into a lesson in another language
An AI language acquisition tool for students, built on a patent-pending multilingual dubbing pipeline that syncs transcription, translation, and audio-video re-encoding. It reached 15K users and was deployed in ESL classrooms.
Business, 2023 to 2024
Origins Customs
Custom apparel and merch, built to order
A custom-merch business run on a full-stack ordering platform. It generated $70K+ in annual revenue and worked with Wingstop and LegalZoom.
Awards
- USACO GoldGold division of the USA Computing Olympiad, the national competitive programming contest. Second highest of four divisions, reached by roughly the top 10% of competitors.
- Hackswift1st place out of 924 participants in a global online hackathon with a $40K prize pool.
- AAE Hackathon1st place out of 158 participants.
- Worthy Hacks2nd place out of 177 participants in a virtual hackathon run by the nonprofit WorthyWorks.
- YC Startup SchoolSelected for Y Combinator’s Startup School in San Francisco, about 6,000 attendees chosen from 30,000+ applicants.
- Stanford TreeHacksAccepted to Stanford’s flagship 36-hour hackathon, which takes about 6% of applicants.
- International Biology BowlSemi-finalist, top 16 of 300+ teams from around the world.
- National Quiz Bowl ChampionshipQualified for the high school national championship tournament.
- President’s Volunteer Service AwardGold, the highest annual tier, for 100+ hours of volunteer service in a year.
Toolkit
Python
C++
Java
JavaScript
TypeScript
React
Next.js
Flask
FastAPI
Firebase
Supabase
MongoDB
SQL
AWS
Azure
GCP
Docker
Git
PyTorch
LangChain
Databricks
Interested in chatting?
Reach out about agents, research, or a hard engineering problem.
