Selected projects spanning agentic LLMs, neural architecture search, and evolutionary optimization.
Autonomous web-exploration agent that constructs a graph-based knowledge base from its discoveries. Performs creative Q&A synthesis through iterative knowledge-graph querying and adversarial refinement, scaling inference-time compute toward state-of-the-art creative outputs.
Agents Knowledge Graph Inference-Time ComputeAn FSM-based framework for orchestrating multi-agent software-development workflows: automated code generation, testing, and execution against benchmarks like HumanEval+. Includes hierarchical expert teams that decompose hard problems via specialized roles, plus evolutionary optimization of prompts and team composition.
FSM LLM Multi-Agent EvolutionA novel algorithm that automatically discovers optimal deep architectures by coevolving both macro and micro structure. Scaled to large distributed environments with hundreds of GPUs and applied across vision and language tasks.
NAS Neuroevolution DistributedMeta-evolutionary bilevel algorithm for complex control problems, applied to challenging tasks including helicopter hovering and sparse-reward RL, optimizing topology and weights simultaneously where gradient-based RL struggles to converge.
Evolution RL ControlEvolves neural-network hyperparameters and loss functions during training. Backed by scalable distributed infrastructure for large-scale evolutionary optimization in parallel with gradient-based learning.
PBT AutoML DistributedProduction AutoML system built on CoDeepNEAT for end-to-end architecture discovery, training, and deployment. Applied at Cognizant across vision, language, and tabular domains, including medical imaging and Wikipedia toxicity classification, demonstrating evolutionary NAS at industry scale.
AutoML NAS ProductionPosition paper proposing a framework for artificial general intelligence grounded in non-equilibrium thermodynamics and chaotic dynamics. Argues that thermodynamic attractors and self-organizing dynamics — not pure gradient descent — are the substrate for open-ended, recursively self-improving cognition.
AGI Theory ThermodynamicsAgentic system that reasons over repository histories (Git/LFS) to detect architectural patterns, surface technical debt, and support large-scale code understanding.
Agents Code AnalysisInference-Time Compute Scaling · Agentic Reasoning (System 2) · Associative Synthesis · Population-Based Optimization · Neural Architecture Search · Neuroevolution · Recursive Self-Improvement · Meta-Learning
Graph-Augmented RAG · Stateful Agentic Memory · Multi-Agent Orchestration · FSM-Based Coordination · LLM-as-a-Judge Evaluation · Evolutionary Prompt Optimization
Python (Advanced) · Keras / TensorFlow · Frontier LLM APIs · Multi-Agent Frameworks (AutoGen, MetaGPT) · Automated Web Parsing