Ganga Meghanath

Applied Scientist · Toronto

I build and evaluate machine learning systems.

Most recently I was an Applied Scientist at Amazon Security, working on multi-agent LLM systems that find vulnerabilities in application code. Before Amazon I spent four years at Microsoft. On the Windows Experimentation platform I worked on causal inference and large-scale telemetry analysis, and earlier, on Bing Ads, I built classifiers for advertising policy violations and fraud.

Much of my work has been in domains that were new to me, and I enjoy the early stage of a project, when the landscape is still unclear and the problem has to be defined before it can be solved. The data has usually been large, in the millions to billions of records, and the systems have had to work in production.

A lot of what I do is evaluation. The systems I work on tend to have scarce or missing ground truth: rare vulnerabilities, sparse fraud, or the output of an LLM agent that nobody has labelled. Designing the benchmark, and deciding what performance and failure even mean in that setting, is often a bigger piece of the work than the model. I would rather report an accurate story from the data than a flattering one, which means being explicit about assumptions, caveats and sampling, and calling an estimate an estimate.

I also write things down. Documentation, wiki pages and internal guides are part of how I finish a project, along with handing the work over so other teams can run it without me.

I hold a Master’s degree in Electrical and Computer Engineering from Purdue University, where I researched neural population dynamics and speech decoding with Prof. Joseph Makin, and a Bachelor’s in Electrical Engineering from IIT Madras.

Interests: LLM and agentic systems, causal analysis and experimentation, applied machine learning.

Outside work: dance, pottery, and painting.