Software Engineer · AI Platform
Building search at DevRev
Designing search as an agent tool
Search is a tool the agent can call more than once per question. I redesigned what agents pass into search and how we return snippets and results. That brought the average from 6.5 to 3.8 tool calls per query while maintaining task success. I also built Compass to group similar queries and find recurring failure patterns in agent traffic.
Retrieval, ranking, and knowledge access
I work on a two-stage pipeline that combines keyword and semantic retrieval, then reranks the merged results. I evaluated reranker models for relevance and latency and built versioned settings for staged rollouts. I also built a way to control which knowledge bases, tools, and retrieval strategies each customer’s agents can use.
Measuring search quality in production
I built an A/B testing framework that started with rerankers and expanded to retrieval settings, models, and agent tool changes. I also instrumented search so teams could see click-through rate, how high useful results ranked, and agent feedback, and catch regressions before a full rollout.
Other experience
Data pipelines · APT Portfolio
I built an alerting pipeline, from Kafka message routing and producer libraries in C++ and Python to a dashboard for monitoring alerts.
Model efficiency · NTU Singapore
In Prof. Weichen Liu’s lab, I benchmarked how much binarized neural networks could speed up image classification on AlexNet and VGGNet.
Solar forecasting · Tensor Dynamics
At this IIT Delhi startup, I used satellite imagery to track cloud motion and forecast solar irradiance up to three hours ahead.