Projects
Selected work in LLM systems, agents, retrieval, medical AI, and AI for Science.
Multi-Agent Medical Evidence & Triage System
A task-oriented medical AI agent system for health consultation, risk triage and evidence-based research, combining multi-agent orchestration, hybrid retrieval, dual-layer memory and safety constraints.
Key Contributions
- Built task-oriented routing for consultation, risk-triage and medical-research workflows.
- Introduced agent-level capability policies to control accessible medical and memory tools.
- Implemented hybrid medical retrieval using dense search, BM25 and reciprocal-rank fusion.
- Integrated Redis short-term memory with Mem0-based long-term memory.
- Added pre-routing risk screening and final-response safety validation.
Medical Deep Search Agent with Agentic Post-training
A medical Deep Search agent post-training project focused on long-horizon tool use, evidence acquisition, recovery behavior and agent policy optimization through Agentic SFT, Recovery SFT and Agentic RL.
Key Contributions
- Designed the runtime contract for multi-step medical evidence acquisition.
- Defined structured tool observations, evidence metadata and fatal-error boundaries.
- Built data contracts and training pipelines for Agentic SFT, Recovery SFT and Agentic RL experiments.
- Designed a training path covering tool-trajectory SFT, grouped rollouts and Fatal-Aware GRPO.
- Developed the project under a local-first architecture with scalable training components.
Protein Language Models for Antimicrobial Peptide Discovery
A computational antimicrobial peptide discovery project organized around a Generate → Filter → Rank framework, covering conditional peptide generation, AMP/non-AMP classification and MIC activity ranking with protein language models and graph-based representations.
Key Contributions
- Built a conditional peptide generation module using a GPT-2-style protein language model, soft prompts and explicit function embeddings.
- Explored LoRA, supervised contrastive learning and margin-based classification for AMP discrimination.
- Combined sequence representations, residue graphs and protein embeddings for MIC prediction.
- Organized generation, classification and activity ranking as separate stages in a Generate → Filter → Rank computational framework.