Samridhi Vaid, ML engineer building multimodal LLMs, agentic systems, RAG pipelines, and evaluations. 4+ years in ML/AI. Samridhi Vaid /work /projects /research /writing /reading
Samridhi Vaid is a machine learning engineer specializing in healthcare AI, agentic systems, and the practical economics of deploying language models. She is first author of research on Berta AI Scribe, an open-source clinical documentation system used by 198 emergency physicians across 105 Alberta facilities.
From software engineering to clinical AI
Vaid began her career at Samsung Research & Development Institute, working on television automation and conformance testing from 2019 to 2021. She subsequently completed a computer science master’s degree at the University of Alberta in 2023, researching Quality-Diversity Transfer Learning for medical prediction problems with limited training data. Earlier projects included retail footfall analysis and medical predictive modeling.
She has worked with the University of Alberta’s Faculty of Medicine & Dentistry since 2023 and Alberta Health Services since 2025, developing language-model applications and evaluation pipelines for clinical summaries. She has also advised startups through the Alberta Machine Intelligence Institute, mentored AI4Good Lab participants, and contributed to Towards AI as a machine learning engineer, consultant, and technical writer.
Her principal healthcare project, Berta, combines speech recognition and language models to convert physician-patient conversations into clinical notes while keeping sensitive data within secure health-system infrastructure. Research she led documented more than 22,000 clinical sessions between November 2024 and July 2025, with operating costs below $30 per physician each month.
Engineering ideas and contributions
Make research agents inspectable. Vaid’s deep research agent architecture separates an agent’s reasoning from MCP-hosted search, video-analysis, and report-compilation tools. Intermediate results remain available for verification and debugging, while progressive disclosure for agent skills loads detailed instructions only when a task requires them.
Test whether compaction actually reduces costs. Working on the Towards AI open-source tutor, Vaid found that prompt caching versus context compaction can reverse conventional optimization assumptions: retaining complete conversations preserved more information and sometimes cost less because cached tokens were discounted, while summarization introduced fresh tokens and discarded useful details.
Match retrieval to deployment constraints. Her context-engineering experiments distinguished cloud models, where cached conversation histories can remain economical, from local systems constrained by smaller context windows. She favors hybrid semantic and keyword retrieval when precise facts disappear in dense-search results, combining semantic matching with BM25.
Combine exploratory research, file-based handoffs, structured editorial feedback, and a calibrated judge to turn a human brief into technical content you can inspect and improve.
Why a fluent post still needs research and editing
An AI tutor’s compaction experiments show how cached history, retrieval, and hardware limits change what belongs in the context window—and when removing it becomes necessary.
The instruction was there. Why did the agent ignore it?
▶ Watch ↗