machine learning / reinforcement learning / AI engineer
IIT Roorkee B.Tech + M.Tech, Geophysics Minor in Economics
I build models that learn, and the systems that keep them alive in production.
Reinforcement learning, agentic AI, and applied ML, taken all the way from a research idea to a monitored, secured, deployed service. The modeling is the work; shipping it is the proof.
Four roles across RL research, applied AI, and cloud engineering. Each one connects model behavior to the production system around it.
Architected the task distribution for 4 long-horizon coding-agent RL environments, specifying tool-use observation/action spaces, state-transition and reset semantics, and parameterized scenarios. Designed verifier-derived reward functions for long-horizon credit assignment, then ran distributed GRPO policy optimization with ablations over reward normalization, KL control, rollout sampling, and curriculum difficulty.
Built a memory-aware RAG system for meeting, reminder, and profile recall across temporal records, semantic profiles, and episodic chat using metadata-filtered BM25/dense retrieval and RRF, with cross-encoder reranking and CRAG correction specified for ambiguous queries. Improved resilience by 60% with reverse-proxy security, rate limiting, circuit breakers, and message queues for crash reporting; fail-closed Stripe verification maintained accurate tier-gating for 100% of users.
Cut manual compliance work by 80% by building agentic RAG over 2,400 IPR pages, reaching >95% Recall@10 with MPNet, Pinecone HNSW, cross-encoder reranking, and 91.3% cache hits. Automated 70% of verification at 99.1% accuracy, then scaled retrieval and CNN inference independently on EKS using CPU demand and Redpanda consumer lag.
Consolidated NAT Gateways via Terraform for a 96% infra cost reduction, saving $55K+/yr. Moved Jenkins to ArgoCD GitOps with Trivy/SonarQube quality gates, cutting release cycles by 40%.
Every project below is a trained model or agent wrapped in the serving, evaluation, and safety machinery it needs to run for real.
Autonomous drone-fleet platform: real-time telemetry, GPU vision inference, stream analytics, and LangGraph mission coordination, with an MCP policy gate on the LLM-to-actuator path so the model can never drive the hardware outside its envelope.
Reduced held-out world-model error by 28% with contrastive features and matched GRU-RSSMs over seven public-state channels. Cut live-trajectory needs by 35% using BC/IQL with two-step MOPO/MOReL/COMBO rollouts, then sustained 99.5% valid actions at sub-100ms p95 with multi-GPU recurrent PPO and four runtime gates.
Models Terminal-Bench tasks as an OpenEnv-compliant RL environment behind a FastAPI + WebSocket orchestration server. Cut manual evaluation overhead 4× and lifted throughput 3×.
A containerized OCR ensemble with selective SLM/VLM adjudication. Holds 93%+ accuracy at a sub-$0.01/document cost target across isolated Docker microservices.
A 3-tier confidence cascade that spends the LLM only when it has to: PII scrubbing, HyDE, BM25 + FAISS, cross-encoder reranking, abstention gates, and enforced citations.
Early work, self-published. An exploration method I built and documented while learning to take RL research end to end.
An exploration method that adapts intrinsic curiosity rewards over time and uses episodic memory to stop agents re-exploring already-visited states. Built for sparse-reward environments where agents need a stronger signal for novelty, memory, and long-horizon discovery.
Accepted, AAAI-26 Student Abstract Program.
Semi-Finalist, Pan-IIT AI/ML Hackathon (IDFC First Bank), 2026
National Rank 4, Launch Pad 2025, Pan-India Case Competition, SRCC