Principal Applied Scientist · Amazon AGI

Samridhi
Choudhary

I work on post-training for foundation models, with a focus on multi-turn reinforcement learning for software-engineering agents. Previously, I built multilingual NLU and conversational AI systems for Alexa, including compressed and quantized models for on-device inference; before that, I studied Language Technologies at Carnegie Mellon.

New York

What I’ve built

01Foundation models

Amazon Nova

I develop post-training and evaluation systems for Amazon Nova, spanning the original family, Nova 2, and Nova Sonic.

Nova ↗ Nova 2 ↗ Nova Sonic ↗

02Current research

Coding Agents

I build multi-turn reinforcement learning systems that train agents to navigate codebases, plan changes, use tools, and verify their work.

Current work

03Production AI

Alexa

I developed compressed and quantized neural NLU models under 5 MB and 60 ms for on-device inference.

Read the Alexa story ↗

04Edge intelligence

Efficient Multilingual AI

I built multilingual NLU and spoken-language understanding systems deployed across 35+ locales and five device families.

Read the paper ↗

Research threads

What I’m working on

Curriculum learning for agentic RL

Increase difficulty without losing the learning signal.

I design curricula that teach component skills first, then combine them across increasingly complex, longer-horizon SWE tasks. As rewards become sparser, I design intermediate signals that make credit assignment clearer without prescribing every step.

Autonomous RL environments

Build learning loops across a broad portfolio of software tasks.

I lead the development of 30+ SWE reinforcement-learning environments across languages, repositories, and task types. My own work turns developer traces into verified environments, so real failures can become the next generation of training tasks.

Evaluating useful autonomy

Measure whether an agent can finish the job.

I design evaluations for understanding large codebases, planning and executing multi-step changes, using tools, debugging, verifying work, and recovering when an approach fails.