Manasvi S.

I'm a Senior AI, Control Systems & Autonomous Agents Specialist with 14+ years of R&D experience at the intersection of artificial intelligence, control theory, CAE simulation, and agentic automation. What sets me apart: I don't just build AI systems — I build systems where reasoning is traceable and outputs are verifiable, making me equally effective as an architect, evaluator, and adversarial domain expert. 𝗔𝗜 & 𝗔𝗴𝗲𝗻𝘁𝘀 — Multi-agent pipelines (LangGraph, MCP, RAG, knowledge graphs), document intelligence (LayoutLM, OCR, LLMs), and RL/GRPO-based control strategies. Recent builds include a 19-tool ReAct simulation orchestrator where the agent writes all SQL, Python, and Cypher itself — tools only execute — and an autonomous GUI agent with iterative learning and human-in-the-loop verification. 𝗖𝗔𝗘 & 𝗖𝗼𝗻𝘁𝗿𝗼𝗹 — High-fidelity FEM and multibody simulation (SIMPACK, ANSA, NASTRAN, ABAQUS) for NVH, driveline dynamics, and planetary gear systems. Hierarchical MPC (CasADi) with CLF/CBF safety constraints and Physics-Informed Neural Networks for real-time parameter estimation. Three granted patents in energy-efficient MPC for electric vehicles. ── 𝗦𝗵𝗶𝗽𝗽𝗲𝗱 𝗦𝘆𝘀𝘁𝗲𝗺𝘀, 𝗡𝗼𝘁 𝗖𝗼𝗻𝗰𝗲𝗽𝘁𝘀 ── ⬡ 𝗦𝗶𝗺𝘂𝗹𝗮𝘁𝗶𝗼𝗻 𝗢𝗿𝗰𝗵𝗲𝘀𝘁𝗿𝗮𝘁𝗼𝗿 — A single-orchestrator platform where a ReAct agent drives 19 MCP tools across discovery, data, configuration, and simulation. The agent writes all SQL, Python, and Cypher; tools only execute. Session persistence passes artifact paths between tools. A full digital twin runtime supports configurable controllers, fault injection, and performance metrics. Every session produces traceable, inspectable artifacts — all reasoning lives in the orchestrator, never buried inside tools. ⬡ 𝗔𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀 𝗚𝗨𝗜 𝗔𝗴𝗲𝗻𝘁 — Converts natural language tasks into executable GUI sequences via a 6-node LangGraph state machine. ChromaDB retrieval with KB-source attribution traces failures to the exact pattern that caused them. Trust scores decay on unreliable patterns. Parameterized tasks separate operations from file paths, making verified skills reusable across contexts. Deliberately iterative — no automatic retries. Users rerun, the system delivers progressively better plans using error learnings and LLM-generated recovery strategies. First-run success: 70–80% → after learning: 85–95% → verified reuse: 95%+. ⬡ 𝗗𝗼𝗺𝗮𝗶𝗻-𝗔𝗴𝗻𝗼𝘀𝘁𝗶𝗰 𝗔𝗻𝗮𝗹𝘆𝘀𝗶𝘀 𝗣𝗹𝗮𝘁𝗳𝗼𝗿𝗺 — Plugin architecture where Discovery, Correlation, and Root Cause agents work across any domain — thermal, motor, aerospace — with zero code changes. Swap one import line and the full pipeline runs on new data. NLQ intent classification routes queries to the right agent workflow. Typed memory (causal, episodic, pattern, rejection) with confidence tracking ensures the system learns from experience. Negative memory prevents repeated dead-end investigations. A cross-domain abstraction engine promotes recurring patterns into general principles automatically. ── 𝗖𝗼𝗿𝗲 𝗖𝗼𝗺𝗽𝗲𝘁𝗲𝗻𝗰𝗶𝗲𝘀 ── ✦ Autonomous Agent Development — ReAct/LangGraph orchestrators with reasoning traceability & MCP integration ✦ AI Agent Evaluation — Adversarial trace-level review; root-cause failure attribution ✦ AI Simulation Pipelines — Agents that configure, execute, and analyze simulations autonomously ✦ MPC & RL Controllers — Safe, adaptive controllers with CBF/CLF constraints ✦ Document Intelligence — 90%+ automation across 30,000+ PDFs in production ✦ Digital Twins — Physics-informed twins with live data and AI decision layers ✦ Automation & DevOps — Docker, Kubernetes, ROS2, Kafka, Azure DevOps ── 𝗠𝗲𝗮𝘀𝘂𝗿𝗮𝗯𝗹𝗲 𝗜𝗺𝗽𝗮𝗰𝘁 ── → 90% reduction in manual effort across 30,000+ PDF pipelines → Sub-200ms real-time MPC for GPU server racks and EV battery trailers → 8% real-world energy savings in adaptive cruise control — 3 patents granted → 95%+ task success rate on verified skill reuse in autonomous GUI automation ── 𝗧𝗼𝗼𝗹𝘀 ── LangGraph · MCP · RAG · ChromaDB · Neo4j · PyTorch · PINNs · CasADi · XGBoost · Abaqus · Nastran · ANSA · MATLAB/Simulink · Simpack · Docker · Kubernetes · Azure DevOps · ROS2 · Kafka · PostgreSQL · Redis · FastAPI · React Whether you need a multi-agent system built, an existing pipeline hardened for production, or a complex control + AI system taken from concept to deployment — I bring domain depth, structured reasoning, and a track record of shipping systems that work in the real world.