01 — About
The short version
I grew up in Abu Dhabi, moved to Cincinnati for a CS degree at the University of Cincinnati (3.99 GPA), and now work out of Seattle as a machine learning engineer at Oumi.
Along the way: a paper on polysemanticity in language models at the NeurIPS 2025 Mech Interp Workshop, a role running Bearcat Ventures (UC's $1M student-led fund), and a co-founded company, PhizzIO, that uses computer vision to guide physical therapy from a laptop camera.
The through-line, as best I can tell: I'm interested in what's inside things. Language models, clinical notes, drum recordings, real estate documents. Different surfaces, same question.
Mechanistic Interpretability
Circuits, features, superposition. When a language model answers a question, something specific is happening inside it. I want to understand what.
Open-Source AI Tooling
Most of my work at Oumi lives here: training loops, eval harnesses, synthetic data pipelines. Unglamorous to describe, important to get right.
Applied ML
Physical therapy video, clinical notes, mridangam recordings, real estate PDFs. Unfamiliar domains tend to teach you things clean benchmarks cannot.
02 — Experience
Work and research
Machine Learning Engineer Intern
OumiData synthesis and interpretability tooling on Oumi's open-source stack. Currently building an MCP server that exposes Oumi's ~500 training configs to AI coding assistants; I've been calling the direction VibeML: natural-language ML orchestration.
AI Research Intern
AlgoverseLooked at how polysemanticity shows up and stabilizes across training in language models. The paper got into the NeurIPS 2025 Mech Interp Workshop.
Biomedical Informatics Software Engineer
UC College of MedicineClinical software with the iCDCU lab. Designed a research project management system for a pediatric heart institute; two papers came out of it.
Managing Partner → Alumni
Bearcat VenturesJoined as an Analyst, moved to Director of Deal Flow, and eventually led the fund as Managing Partner. Ran UC's $1M student-led fund: sourced deals, led diligence, managed a team of analysts. Placed 2nd at the VCIC Midwest Regional at Carnegie Mellon. Transitioned to Alumni in April 2026.
Machine Learning Intern
Kinetic VisionTrained vision models to catch structural defects in surgical staples, plus the data pipelines that fed them.
Software Engineer Intern
Phillips Edison & CompanyTwo co-op rotations. Built a multimodal ETL pipeline that pulls structured fields out of the real estate documents nobody wants to parse by hand (leases, site plans, abstracts).
04 — Honors
Awards and recognition
05 — Projects
Selected side projects
linear_probes
A mech interp toolkit I keep reaching for. Pulls activations from any transformer layer with flexible selector syntax, runs linear probes with proper train/val/test hygiene, sweeps layers. Every probe experiment I run starts here.
Oumi Platform
My open-source work on Oumi: training configs, inference engines, eval tooling, synthetic data pipelines, and the MCP server I'm building on top of it all.
PhizzIO
CV-powered physical therapy app. Watches you do the exercises, flags when your form slips. Won Launch It: Cincy ($28K+) and placed 2nd at RevolutionUC.
Mridangam Transcription
A CNN with attention that transcribes stroke patterns from mridangam recordings. The mridangam is the South Indian drum I grew up around; I wanted a solver for a dataset that barely exists.
ctrlFind
Codebase search in plain English. Written in Rust. Embeds the code, runs cosine similarity, hands you back the thing you half-remember writing.
BuildGPT
A small GPT written from the ground up. Not novel work; a forcing function to make sure I could explain every line before trusting anything larger.
GraphRag
RAG with a knowledge graph underneath instead of flat chunks. Built it after losing one too many research ideas that lived in the connection between papers, not any one sentence.
MNIST from Scratch
Three implementations side by side: stock PyTorch, hand-coded from the math with no frameworks, and a custom reimplementation of the PyTorch API. 96% accuracy across all three. Built to understand backprop at the level of the math, not the library.
06 — Contact
Get in touch
I'm always happy to talk about interpretability, open-source ML infrastructure, startups, mridangam, or whatever you're working on. My inbox is open.
rameshad@mail.uc.edu