MSE Computer Science @ Penn  ·  asheoran@seas.upenn.edu  ·  LinkedIn  ·  Personal Website

I'm drawn to problems with real constraints: enterprise systems that have to scale, research questions that need tooling before they can be answered, agents that have to be trustworthy not just capable. Background in bioinformatics. I run CBC, Penn's 900-member community for students building with AI.

Thinking about AI safety

There's a spectrum from intrinsic safety (model weights, training) to extrinsic safety (environment constraints, scaffolding, governance) and most practical work lives in the middle without a clear theory of why. I want to understand what good technical solutions look like at each layer, beyond system prompt interventions.

• Safety interventions for clinical agents — layered evaluation harness testing whether structured inference-time interventions close the performance gap between frontier and domain-specific medical models, no fine-tuning
• SWE-bench agent evaluation — baseline + intervention experiments on Claude 3.5 Sonnet; found interventions can shift failure distributions even when top-line resolve rate holds steady

Healthcare + bioinfo tools

• Agentic pipelines for pharma drug launch readiness — built at Clariem for senior exec teams; trust and auditability mattered as much as capability
• Single-cell RNA tooling — exploring where agentic workflows can accelerate the manual reasoning steps in scRNA-seq interpretation

Other things I've built

• Google Search, 1998 (available upon request) — full search engine from scratch: distributed crawler, TF-IDF indexer, PageRank, and search server on AWS EC2
• Image2GPS — predicts GPS coordinates from campus images using a ConvNeXt + k-NN hybrid; cut baseline localization error by ~70%
• Market forecasting tool — deterministic, reproducible market forecasts for consulting teams through a 5-stage gated process
• CBC demos — tooling and demos built for Penn's AI builder community

Research

• Data reporting quality and semantic interoperability increase with community-based data elements — Experimental Neurology, FAIR Data Informatics Lab, UCSD
• Enhancing Claude 3.5 Sonnet Reliability on SWE-bench Verified — Penn CIS 7000, available on request