Dhruv Joshi
AI & ML Engineer: LLM systems, RAG and agentic AI
I build production-grade LLM systems for regulated domains, including clinical AI, IP intelligence and legal NLP. I currently lead the technical work on Liver Digital Twin and HCC detection at Karyon.Bio.
About Me
Professional Summary
I am an AI & ML Engineer with 4+ years of experience building production-grade LLM systems, RAG pipelines and agentic AI. I've worked in biomedical AI, IP intelligence and legal NLP. I design and ship complete solutions, from retrieval and agent orchestration to evaluation, observability and deployment, using LangChain/LangGraph, FastAPI and modern MLOps tooling (Docker, Kubernetes, Airflow).
I care about rigorous, quantitative evaluation: calibration analysis, leakage-free test splits and failure-mode sampling guide every iteration. I also build privacy-first infrastructure so sensitive data like patient records never leaves controlled environments.
Core Expertise & Skills
| AI & ML | LLMs (GPT-4/5, LLaMA, Gemini, MedGemma), RAG, Semantic Search, Transformers, Prompt Engineering, LoRA / PEFT, Multimodal AI, Agentic AI, Digital Twins, Model Evaluation |
| Frameworks | LangChain, LangGraph, HuggingFace, PyTorch, TensorFlow, TabNet, FastAPI, Flask, Streamlit, Pydantic |
| MLOps & Cloud | Docker, Kubernetes, Airflow, MLflow, W&B, CI/CD, Qdrant, Chroma, Redis, AWS (S3, Lambda, EMR, Athena), GCP Vertex AI, Azure |
| Data & Languages | Python, SQL, Spark, Hadoop, Snowflake, PostgreSQL, MongoDB, Elasticsearch, JavaScript, R, MATLAB |
Work Experience
AI/ML Engineer, Karyon.Bio
Aug 2025 – Present · New York, NY- Shipped a clinical decision-support pipeline for liver fibrosis staging (F0–F4) on a biopsy-referred patient cohort, and removed evaluation leakage using confusion-matrix diffs, calibration-bin analysis and per-patient failure-mode sampling.
- Designed a LangGraph ReAct pipeline with deterministic safety gates grounded in AASLD/EASL guidelines, a PageIndex retrieval layer and a provider-agnostic LLM layer (OpenAI, Anthropic, Bedrock, Gemini, Moonshot).
- Built multimodal HCC detection from start to finish. It fuses genomic data, EHR records and liver MRI/CT imaging with a MedGemma + TabNet architecture, validated on held-out data across all three modalities.
- Engineered a fully offline, LangSmith-style observability stack (hierarchical JSONL spans, nested tool-call tracing, Streamlit dashboards) that keeps patient-derived data off third-party services.
AI/ML Engineer (Founding), Squirrel IP
Aug 2024 – Aug 2025 · New York, NY- Architected Artha, an AI-powered patent valuation engine. It turns granted patents into bank-grade credit scores using Cost + Market + Income triangulation and Monte Carlo simulation (N=10,000).
- Deployed a multi-agent LLM document intelligence system (LangChain, RAG, Qdrant) that integrates 6+ live data APIs and returns auditable JSON to financial institutions.
- Scaled the full pipeline (ingestion → RAG → valuation → report) on FastAPI, Kubernetes and Airflow, with access control and audit traceability.
Data Engineer, Go Digital Technology Consulting
Jul 2022 – Aug 2023 · Mumbai, India- Built and optimized PySpark ETL pipelines that process terabytes of financial data.
- Led the Hadoop → AWS migration (EMR, S3, Athena, Lambda), improving scalability and retrieval time.
Education
M.S. Computer Science
2025B.E. Electronics & Telecommunication
Publications & Research
Large-Language Memorization During the Classification of United States Supreme Court Cases
This paper benchmarks BERT, Legal-BERT, LLaMA 3 and DeepSeek on 279 legal issue categories, comparing fine-tuning with RAG-based approaches and studying memorization versus generalization. The best result was F1 0.624.
DOI: 10.48550/arXiv.2512.13654Multi-Modal Fusion Framework for HCC Prediction
This framework integrates transcriptomics, medical imaging and clinical attributes to improve hepatocellular carcinoma diagnosis. I built it with Kiran Adhikari, and my work covered the multimodal modeling (MedGemma + TabNet) and the pipeline.
Project page →Select Projects
AR Sign Language Accessibility App
An immersive AR app built in Unity. It recognizes sign language in real time from live camera data using CNN+LSTM models (70%+ accuracy) and has a text-to-speech system, so deaf and hard-of-hearing users can communicate in both directions. I built the full data pipeline, from dataset collection and annotation to model inference optimized for mobile AR.
Rosa: AI-Powered Women's Health Chatbot
A multi-agent LLM health assistant with session memory and RAG over a curated knowledge base. It has three specialized agents (mental health, physical health, triage) and scaled to 25K users.
CRM Multi-Agent AI System with Knowledge Graph
Audit-ready CRM automation with semantic search, Google Calendar integration and auto-tagging, deployed as FastAPI + Flask microservices with knowledge-graph memory.
Multilingual Document Analysis Assistant
An assistant that extracts technical terminology across languages using BM25 reranking, RAG and a fine-tuned term identifier. Built with Streamlit, Flask and Gemini Pro, it accepts PDF, DOCX and TXT files.
Contact Me
I'm open to conversations about AI/ML engineering, clinical AI and LLM systems.