Kushal Regmi

I build the production multi-agent system on the longitudinal health memory, the durable record of a child’s history that the agents reason over across sessions.

• Engineered the production multi-agent system: stateful LangGraph agents with human-in-the-loop control, evidence-grounded RAG, and clinical safety guardrails, so answers come from the child’s own record and never drift into medical advice.
• Own the agent evaluation harness that gates every release: it drives the live agent and verifies each stored field against ground truth. It caught a data-loss defect (two facts in one message, one saved) that no unit test flagged.
• Built the longitudinal memory the agents reason over: a HIPAA-safe personalized health knowledge graph, synced from connected EHR data, that carries structured context across sessions and reports.
• Cut agent retrieval latency 49% (1486ms to 753ms) with an access-pattern audit, and held it sub-linear: 175x more stored data for 1.53x latency.
• Kept the longitudinal health memory service reliable on Docker, GitHub Actions CI/CD, and Kubernetes: stopped the OOM evictions crashing it by cutting pod memory 33% (1.5 GB to 1.0 GB), and cut deploy time 10x (31 min to 3 min).

• LangGraph
• RAG
• Agent evaluation
• Knowledge graph
• EHR sync
• HIPAA-aware
• DynamoDB
• Kubernetes

Laid CareGene’s model and evaluation foundations: what to fine-tune, what to serve, and how to measure trust.

• Fine-tuned DistilBERT with focal loss over 9 intent classes to power backend guardrails and intent-aware prompting.
• Quantized that classifier to ONNX INT8 for serving: 75% smaller (256 MB to 64 MB) and ~2x faster on CPU at ~2x throughput, with accuracy held at 98.62% vs 98.94% in FP32.
• Ran a QLoRA supervised fine-tune of Llama 3.2 on EC2, served with vLLM FP8, as a pilot to test whether a custom rare-disease model was worth it.
• Built CareGene’s first LLM and RAG evaluation framework in DeepEval and G-Eval, with a custom metric suite for factual correctness and clinical-safety deferral, aligned to product requirements.
• That evaluation redirected the product: across roughly 7,000 rare diseases with constantly changing findings, retrieval beat a fine-tuned model on accuracy and freshness, so we kept a frontier model and invested in RAG and memory.

• DistilBERT
• QLoRA
• vLLM FP8
• ONNX INT8
• DeepEval
• G-Eval

Built Sambodhan, a multilingual civic-tech platform that auto-routes citizen grievances by department and urgency.

• Fine-tuned XLM-RoBERTa to ~95% F1 across Nepali and English, with focal loss so rare urgent complaints were not drowned out by routine ones.
• Deployed the department and urgency classifiers as containerized FastAPI microservices on Hugging Face Spaces.
• Built continuous-learning MLOps: a GitHub Actions loop retrains on misclassification drift and promotes a model only when it beats the incumbent on F1-macro.

• XLM-RoBERTa
• Hugging Face
• GitHub Actions
• FastAPI