Looking for a job

Devendra Sai M

@devendrasaim

I build reliable production AI agents through full-stack engineering, rigorous evaluation, and LLM reliability research.

United States

Message

What I'm looking for

I'm looking to build and harden production AI-agent systems where rigorous evaluation, reliability engineering, and full-stack ownership turn unpredictable LLM behavior into dependable user experiences.

At VelocitiPM, I own the Velo AI drawer end to end, building agent experiences across React, TypeScript, FastAPI, LangGraph, Supabase, and Postgres.

I’ve added durable task persistence, corrected agent state-tracking behavior, rebuilt onboarding and the app shell, and wired completion-driven status automation with immediate dashboard updates. I ship automated tests with nearly every feature and maintain local development workflows with Docker Compose and the Supabase CLI.

My Iowa State research made CodeLLM output more reliable without retraining: activation steering reduced invalid, non-compiling output on CodeLlama-7B from 4.38% to 0.00% while improving accuracy from 48.12% to 56.25%. I bring that evaluation-first approach to production agents, using evidence and error analysis rather than guesswork.

Experience

Work history, roles, and key accomplishments

Current

Software Engineer

Current

VelocitiPM

May 2026 - Present (4 months)

Owns the Velo AI drawer end to end, adding task persistence and fixing state-tracking bugs. Works full stack with React/TypeScript, FastAPI, LangGraph, and Supabase, and ships features with automated test coverage.

fastAPI LangGraph React TypeScript Supabase PostgreSQL Docker Compose GitHub Actions Cloudflare Turnstile Claude Code

Graduate Research Assistant

Iowa State University

Aug 2025 - Dec 2025 (4 months)

Designed a training-free activation steering method in PyTorch to make CodeLLM output reliably correct. Built an evaluation framework across 40+ configurations, eliminating invalid output entirely on CodeLlama-7B while improving accuracy.

PyTorch Slurm CodeLlama Qwen Human Eval Defects4J Evaluation Activation Steering

Education

Degrees, certifications, and relevant coursework

Google Cloud

Prompt Design in Vertex AI, Prompt Design

Completed the Prompt Design in Vertex AI certification from Google Cloud in February 2026.

Iowa State University

Master of Science, Computer Science

2023 - 2025

Master of Science in Computer Science with a thesis on automatic program repair through activation steering, focusing on artificial intelligence, machine learning, and system architecture.