My name is Barath. I am a Master's student at MIT and previously studied at the University of Waterloo.
Here is my personal portfolio: web.mit.edu/barathv/cv/ and LinkedIn.
Feel free to checkout my writing: Substack and Medium.
- kaetram-arena: structured game-agent distillation, distilling frontier-LLM gameplay reasoning into a small open student model through a typed MCP tool API, with SFT + KTO data collection across four LLM harnesses (with Niral Patel).
- subliminal-learning: replication and extension of subliminal learning, an MLP that learns MNIST digits from pure noise via a shared initialization, plus token entanglement in Llama prompting, with a mechanism story the data supports.
- crucible: a multi-agent evaluation harness on Google ADK, a ReAct researcher, a self-reflection reviser, and an LLM-as-judge, with groundedness scoring, LLM-native metrics, and per-call tracing; runs offline with no API key.
- clinical-corpus-pipeline: turns messy multi-provider EHR and claims data into grounded clinical-reasoning training examples, with privacy-preserving linkage and quality gates that reject hallucinations.
- screenmem: local-first macOS screen memory, capture, OCR, and hybrid keyword and semantic search over your screen history, with a local-LLM grounded ask layer.
- attractor-lens: reusable tooling for mapping the recurring attractor states LLMs fall into when left in a self-loop.
- deterministic-voice-intelligence: a tap-to-talk fintech voice assistant, Whisper transcription, deterministic LangGraph intent routing, and Edge TTS.