Exploring how intelligent systems work.
Making a little music along the way.
Portable Neural Programs: Cross-Host Reuse and Native Execution
Can a learned program work beyond the system that created it? This paper tests unchanged artifacts across two language-model hosts and then compiles acquired behavior into native model weights, separating portability from native execution.
Read the paperThe same artifact runs through a shared executor behind different hosts.
Acquired behavior is compiled into a host's own weights.
Turn Aware Session Retrieval on LongMemEval S
Instead of embedding an entire conversation session as one block, Atlas scores individual turns and ranks each session by its strongest match. It retrieves all required sessions for 436 of 470 questions under the released evaluator; the direct annotated-ID check passes 433 of 470.
Discrete Representation Protocols for VRAM-Constrained Local LLM Pipelines
A study of how two locally run language models pass information between stages on a 16 GB GPU. Across six message formats, concise prose summaries offer a strong balance of factual fidelity and lower token and memory cost.
Sensor based Low-cost Emissions Management for Small and Medium Farms
Co-authored at Castleknock College, this project explored affordable sensors, predictive models, and a cloud interface to help smaller farms track agricultural emissions. A very different starting point from my current work on how AI systems represent and transfer behavior.
Original music · Open in Spotify ↗
I'm Sachin, studying EECS in UC Berkeley’s Management, Entrepreneurship, & Technology program. My interests span intelligent systems, engineering, and music.
This is a home for the work I’m thinking through and the things I’m making.
Let’s connect on LinkedIn