Principal Research Scientist
Cognizant AI Lab
Lead Researcher driving the roadmap for the lab's automated AI research effort.
Autonomous Self-Improvement Systems
- Built a values-based self-improvement framework in which an agent edits its own solver and its own values document across long iterative runs, organised around three questions: what founding values are seeded, which anti-reward-hacking invariants the agent cannot overwrite, and how much of its own instructions it may rewrite. Written up across an research blog, figures and generating scripts included.
- On variable-radius circle packing the agent wrote its own solver, matching the best known value at N=26 for under $20 of compute to first tie with no web assistance. At N=27 the same solver beat a record standing since 2011 and is now the listed best known result, credited on Packomania. A later run produced a solver that beats 21 more of the published records, at sizes from 50 to 87 circles, each re-verified from its coordinates against the live table.
- On a real AtCoder heuristic contest, a run's solver outscored the top human leaderboard entry over the 150 official private seeds, confirmed on the official judge after the contest.
Caesar: Autonomous Reasoning & Knowledge Synthesis
- Architected Caesar, an autonomous Deep Research agent that outperforms the strongest of three frontier deep-research agents by 13–23% on a blinded, LLM-judged creative-synthesis benchmark (Cliff's δ ≥ 0.76). Scaled inference-time compute via an adversarial refinement loop that critiques internal drafts, generates orthogonal queries to attack weaknesses, and consolidates findings through generative merge.
- Replaced static RAG with a dynamic Perceive-Think-Act loop backed by a dynamic knowledge graph, enabling associative reasoning that surfaces non-obvious cross-disciplinary connections.
- Designed an active information-foraging policy that detects stagnation and autonomously executes strategic backtracking over long horizons. Ablation: cutting the exploration budget from 1000 to 250 iterations measurably degrades long-form answer quality.
Multi-Agent Systems & Large-Scale Evolutionary Optimization
- Developed an FSM-based multi-agent code-generation framework, evolving the composition of specialized LLM teams to achieve strong performance on SciCode and HumanEval+.
- Designed a hierarchical expert-agent pipeline that decomposes hard coding problems and routes sub-tasks to specialized LLM roles.
- Pioneered Evolutionary Population-Based Training (EPBT), evolving loss functions and hyperparameters jointly with weights. Beats the PBT baseline by 1.3pp on CIFAR-10 ResNet-32 (92.79% vs 91.53%) while exploring 520 candidate loss functions at the cost of 40 trainings, a 13× reduction (GECCO 2021).
- Pioneered evolutionary prompt optimization, improving LLM performance on challenging code-generation benchmarks.
- Created Code Archaeologist, an agentic system that reasons over Git/LFS repository histories to detect architectural patterns and technical debt.
- Scaled distributed ML systems to produce SOTA neural networks for vision and language tasks.