Building LLM-based and agentic AI systems.

• Designed and implemented Markov Chain Monte Carlo and Reinforcement Learning approaches for globally optimizing BPE tokenization (entropy + compression objectives) training on the MiniPile corpus.

• Curated a longitudinal imaging cohort of ~2k abdominal CT scans from ~80k patients with acute pancreatitis, linked to 3-year follow-up data; built a DICOM-to-NIfTI pipeline with automated PHI stripping.
• Prototyped a deep-survival model on 80k MRIs that fuses 3-D ResNet radiomics with clinical labs (CRP, BUN) in a Cox/DeepSurv head to predict time-to-chronic progression; pilot C-index 0.81 vs 0.72 for Cox on 5-fold patient splits.

• Developed two GNNs (GCN, GNN-DTI) for RNA metal-ion binding, gaining +6.2 pp ROC-AUC over a CNN on 6.4k PDB structures.
• Built a GPU-accelerated PyG stack on SLURM and a DGL graph-builder that cut preprocessing 5x and streamed 1.1M edges/s, enabling 128-config sweeps overnight.

• Built a LangGraph multi-agent workflow (Gemini Flash, DeepSeek-R1) with end-to-end reasoning and tool use, testing latency, usage, and system reliability.
• Extended and productionized the system with a Neo4j knowledge graph for drug-target-indication analytics, modeling evidence-weighted relationships (PageRank, community clustering).

• Built LangChain RAG document QA over 1,200 technical manuals; 92.0 token-level F1 on a held-out test set; p50 <200 ms E2E on a single NVIDIA L4 (24 GB) with INT8 inference; hybrid BM25 + vector + cross-encoder rerank.
• Raised Exact Match from 41% to 84% (+43 pp) on a 500-question benchmark via LoRA fine-tuning, RAFT, and hybrid retrieval with an embedding-based reranker.

• Prototyped a RISC-V vector extension and LLVM backend, implementing scatter/gather intrinsics to deliver 1.7x GEMM throughput and -34% ResNet-50 latency in cycle-accurate sims.

Proposed a novel method using autoencoders to optimize distributed learning and presented at the workshop.

Benchmarked 7 hematology-based prognostic models on 195k patient records across Brazil, Italy, and Western Europe; uncovered ~20% AUROC drop in cross-continental transfer and released an open-source validation toolkit.

• 2nd position - AI4Purpose Hackathon
• Research Mentor - Roaring Cubs Collective