Open to opportunities

Satya Sheel Shekhar

@satyasheelshekhar

Computer engineering student building multi-agent, LLM-driven systems and full-stack platforms with production-ready focus.

India

Message

What I'm looking for

I want to build production AI/agentic systems and full-stack products—shipping reliable features with modern tools, owning backend/ML pipelines end-to-end, and growing through fast iteration and strong engineering practices.

I’m a B.Tech in Computer Engineering student at Thapar Institute of Engineering and Technology, where I’ve been focused on building end-to-end systems—from distributed multi-agent environments to production-grade web apps.

In CrisisCompute, I built an OpenEnv-compliant multi-agent distributed negotiation system where agents (loader, cleaner, trainer) manage shared cluster resources under real-time crisis events. I designed a composable reward rubric and trained a hybrid RL + LLM pipeline that improved average reward by +10.3% and on-time rate by +7.5% on held-out compound-crisis scenarios.

I also ship full-stack and ranking systems: StreamX is a Google OAuth-secured creator streaming platform with a Gemini-powered chat and a decoupled React + Fastify backend; AlgoForge streams real-time algorithm execution states for step-by-step visualization. On the ML side, Caliber implements a two-stage bi-encoder (FAISS) retrieval + cross-encoder reranking candidate ranking system, plus fraud detection layers—something I’m excited to bring to impactful AI products.

Experience

Work history, roles, and key accomplishments

CrisisCompute

Meta x OpenEnv Hackathon

May 2026 - Present (2 months)

Built an OpenEnv-compliant multi-agent distributed negotiation system where three agents negotiate shared cluster resources under deadline pressure and real-time crisis events. Implemented a composable reward rubric trained via a hybrid RL + LLM pipeline and deployed live on Hugging Face Spaces using Docker.

Python PyTorch Reinforcement Learning UnSloth Hugging Face TRL Reward Shaping Multi Agent Systems Docker Hugging Face Spaces

Caliber

Redrob Hackathon

Jan 2026 - Present (6 months)

Designed a two-stage candidate ranking system that retrieves a relevant shortlist with a bi-encoder + FAISS, then reranks top candidates with a cross-encoder against weighted job requirements. Built additional fraud detection and disqualification layers plus an ensemble scoring approach using semantic fit and trust signals.

Python Faiss Sentence Transformers Information Retrieval Candidate Ranking Fraud Detection Ensemble Scoring Bi Encoder