Qiran Hu

IBM Research

Research Intern (Advisor: Dr. Dhaval Patel), 2026.09 - Present, New York, NY

• Lead Model Context Protocol (MCP) server architecture for AssetOpsBench that serves failure-mode knowledge graphs, root-cause ranking, and intervention estimation to multi-turn tool-use large language model (LLM) agents via deterministic Neo4j queries with provenance, raising top-1 tool accuracy by 12.9% compared to per-server sub-agents and cutting standing tool context by 98.5%.
• Design fault-diagnosis reinforcement learning (RL) environments on a hydraulic test rig with verifiable cost-aware rewards, priced observations, and reward-hacking checks for multi-step LLM agents, raising top-1 diagnosis accuracy by 42.6 percentage points compared to a Plan-Execute agent and eliminating 87.7% of model calls per episode.
• Streamline the diagnosis environment into standard Gymnasium and OpenEnv training interfaces for RLVR post-training with GRPO on verl, minimizing per-episode inference latency by 52.7% across 47 recorded fault combinations and saving 47.5% in maintenance cost compared to a do-nothing policy.

Links: AssetOpsBench

Memoria

Technical Lead, 2026.01 - 2026.08, Champaign, IL

• Led end-to-end agentic workflow deployments across MCP servers, sub-agents, and agent skills with multi-agent orchestration, permission-aware tool calling, and policy guardrails, accelerating product delivery by 4.0x and boosting overall task success by 17.9% on CL-bench.
• Managed production operations for agentic workflows with logging, tracing, evaluation harnesses, and monthly iteration cycles for client stakeholders, reducing serving cost by 90.5% compared to GPT-4o and improving rule-application task success by 31.2% on CL-bench.
• Integrated recursive self-evolving memory architecture across sessions with cost-aware model routing into enterprise agent workflows, saving 96.0% in token cost compared to GPT-5.1 and increasing procedural task success by 71.4% on CL-bench.

Links: Website

Two by Two Learning Center

Full Stack Developer, 2025.08 - 2025.12, Champaign, IL

• Launched NOODEIA, a multi-agent tutoring platform for K-12 students below grade level, to plan, critique, and monitor each user with long-horizon memory, boosting user confidence by 2.44x and increasing independence by 2.35x in counterbalanced within-subjects studies.
• Implemented recursive self-evolving long-term memory on Neo4j to retrieve contextually relevant interactions, accelerating query speed by 3.6x compared to recency-biased FIFO memory on PostgreSQL and minimizing cross-session retrieval to 2.1 s.
• Deployed complexity-aware model selection across planner, retrieval, solver, and critic stages to score each request and route it to the smallest capable model, eliminating 89.9% of monthly serving cost compared to GPT-4o and routing 75% of calls to lighter model tiers.

Links: Website

WRC

Data Analyst, 2024.08 - 2024.12, Champaign, IL

• Built paired pre/post analytics pipelines over survey data from 9,935 incoming students, raising correct responses by 14.0% and decreasing ambiguous responses by 61.9%.
• Conducted A/B tests on two consent scenarios to locate where misconceptions persisted after each workshop, increasing accuracy on consent comprehension by 19.5% and cutting persistent misconceptions by 41.6%.
• Designed scenario-based learning modules for underrepresented subgroups, raising correct responses on bystander scenarios by 36.8% and narrowing the gap in correct-response rates between subgroups 5.6x.