Bryce Kan
Los Angeles, CA
Graduate Researcher @ USC ML Center | LLM Post-training · Agentic AI · NLP
I’m Bryce Kan, an ML researcher and CS Master’s student at the University of Southern California. I work at the intersection of LLM post-training, agentic AI, and NLP, currently focused on long-term memory for LLM agents and LLM-powered recommendation systems that reason about both preference and qualification.
Outside of research, I’ve hiked Half Dome, run a marathon, and competed internationally in junior badminton.
Experience
- Built agentic evaluation judge framework across task completion, context retention, and error recovery scenarios
- Routed LLM inference across Anthropic and OpenAI via OpenAI-compatible layer; built FastAPI-backed data visualizer
- Dockerized and deployed eval pipeline; integrated Databricks model-serving endpoints, tracking per-call cost and latency
RaMem: Contextual Reinstatement for Long-Term Agentic Memory
- Built RAG-based agentic memory pipeline enabling LLM agents to recall evidence across multi-session conversations
- Engineered temporal evidence filtering with session metadata to remove semantically related but invalid memories
- Implemented hybrid retrieval with Reciprocal Rank Fusion (RRF) over semantic search and BM25, improving F1 by 11.6 points across GPT-4o, GPT-4.1-mini, and Qwen3-8B
- Reduced prompt context length by 50% while preserving performance against state-of-the-art agent memory baselines
Student-Job Recommendation System
- Led end-to-end development as 1st author, mentoring 2 Master’s students and collaborating with PhD candidates
- Tackled cold-start scenarios using synthetic explanation generation via GPT-4o-mini distillation into Llama-3-8B-Instruct for student-teacher learning
- Implemented multi-step fine-tuning via dual optimization framework with LoRA, jointly training 2 objectives for ranking (InfoNCE contrastive loss)
- Achieved up to 2.6× Recall@5 and over 3× NDCG@5 improvement over LLM baselines
- Developed custom Python pipelines to automate experimental data analysis, adapting to evolving scientific demands
- Designed interactive Plotly dashboards to visualize complex biological datasets, enabling faster data-driven insights
- Launched Lab Inventory Database as final product, delivering full workflow traceability and reducing processing time by 80%
- Built automated data visualization pipelines using Pandas and Plotly to extract insights from terabyte-sized datasets
Education
news
| Dec 15, 2026 | Graduating from USC with an MS in Computer Science — looking forward to what comes next! |
|---|---|
| Jun 23, 2026 | New preprint out — RaMem: Contextual Reinstatement for Long-Term Agentic Memory. Addresses context collapse in LLM agent memory by anchoring retrieved memories to their original episodic conditions, achieving 10%+ average F1 gains across multiple LLM backbones. |
| May 01, 2026 | Starting as a SWE-Agentic AI Master’s Intern at Honeywell in Atlanta, GA — working on LLM evaluation workflows. |
| Feb 01, 2026 | New preprint out — De-conflating Preference and Qualification: Constrained Dual-Perspective Reasoning for Job Recommendation with LLMs. 2.6× Recall@5 and 3× NDCG@5 over LLM baselines. |
| Mar 01, 2025 | Joined the USC Machine Learning Center as a Graduate Research Assistant, working on long-term agentic memory and LLM-based recommendation systems. |
