Meron Oumer
Computer Science & Applied Mathematics student at Smith College building ML systems and asking what neural networks are actually doing when they learn.
Undergraduate Researcher in Smith's HPC Lab, studying optimization dynamics and loss landscape interpretability
Experience
-
Undergraduate Researcher 2025 – present
Smith HPC Lab
Studying optimization framework behavior and neural network training dynamics; extending toward loss landscape interpretability.
-
AI Engineering Intern Aug 2025 -- Dec 2025
Chambers Capital Ventures
Built a Random Forest classifier (F1 = 0.835) on 5,000+ workflow events and a Streamlit dashboard that cut manual monitoring time by 50%.
-
Data Science Intern 2024 – 2025
Conway Innovation & Entrepreneurship Center
Built Ubuntu FieldOps, an offline-first Next.js app for community program tracking in low-connectivity environments.
-
AI & ML Fellow 2025
Break Through Tech × Cornell Tech
End-to-end ML pipelines, LangGraph agents, and deployment-oriented ML systems.
-
Machine Learning Researcher 2024
AI4ALL
Clinical audio ML pipeline across 864 recordings; ensemble models reaching F1 = 0.71 with feature attribution investigation.
Projects
-
Induction Head Replication Study
Replicated key results from In-context Learning and Induction Heads (Olsson et al.) using TransformerLens on GPT-2 Small. Traced induction head formation through attention patterns and activation analysis.
-
TaskPilot
AI workflow automation agent built with LangGraph and LLM APIs. Investigated failure modes in tool-use chains and what they reveal about multi-agent safety.
-
Ubuntu FieldOps
Offline-first sync architecture with Supabase and local persistence. Built for real deployment in low-connectivity environments.
Writing
Notes on interpretability, alignment, and the odd corners of ML systems — written to think, not to summarize.
All writing →About
I'm a sophomore at Smith College (graduating May 2028) most interested in optimization, mechanistic interpretability, multi-agent safety,assistive tech and what it would mean to understand a model well enough to trust it.
More about me →