Ankit Bhardwaj

AI Engineer and Architect specializing in Multi-Agent, Multimodal RAG with LLM orchestration, building GPU-accelerated, cloud-native platforms achieving ultra-low latency and high accuracy for scalable, production AI solutions.

• Engineered the core backend for an agentic AI platform, building low-latency MCP server pipelines, asynchronous RPC channels, and a multi-agent execution layer for scalable autonomous workflows.

• Designed and deployed cloud-native agent runtimes using LangGraph on AWS/GCP, with isolated agent state, secure tool bridges, and parallel task runners to support concurrent reasoning and execution.

• Developed a graph-based Knowledge Memory Layer, including an ontology builder, natural language → Cypher translation, and hybrid graph + vector retrieval, enabling long-horizon reasoning across agents.

• Built standardized agent runtime infrastructure—unified tool schemas, message formats, routing logic, and capability negotiation → improving observability, scalability, and fault isolation for distributed agent workflows.

• Computational Science & Programming: Data Structures Algorithms, Object-Oriented Design.

• High-Performance Computing (HPC): Parallel and GPU-accelerated computing, Linux, Shell script, Git.

• Quantitative Methods: Statistical modeling, probabilistic methods, R.

• Master’s Thesis: ML-Driven (GA-GNN) Mechanism Reduction for High-Temperature Processes

– Genetic Algorithm (GA): Initialized a population of 50 boolean species masks and evolved them over 100 generations using tournament selection (size 5), crossover rate 0.8, mutation rate 0.1, and elitism (top 5 retained per generation).

– Graph Neural Network (GNN): Designed a 3-layer GNN (hidden dim 64, learning rate 0.001, 100 epochs, batch size 32) to surrogate the simulation-based fitness function. Encoded detailed mechanisms as graphs using message passing with ReLU activations and global mean pooling.

– Fitness Evaluation: Combined validation pass/fail constraints with a species-count penalty. Evaluated reduced mechanisms against full simulations using errors in KPI’s performance.

– Achieved a 75% reduction in species count while retaining 98% simulation accuracy, and improved CO2/syngas operational efficiency by 25%.