Anand Sampat

Architected and built an agentic AI platform end-to-end — training pipelines, ingestion, orchestration, and the application layer. Shipped Overline Agent to production in six months, delivering continuous tax, property, and insurance analysis for real estate investors. Designed a multimodal model strategy spanning open-source (GLM-4.6V, DeepSeek OCR, Gemma 3) and frontier models (Gemini 3, Opus 4.5), with post-training pipelines for orchestrators and RAG-based document analysis grounded in property DNA. Built infrastructure that securely ingests and analyzes a $1B+ portfolio across 10,000+ properties.

Built applied LLM and LVM work on novel AI hardware. Designed pipelines to parse documents, generate synthetic data, and pretrain / finetune TableQA and ChartQA models. Co-authored Composition of Experts, a modular compound AI system for routing across LLMs on enterprise tasks (MLSys). Established the team that delivered 2.4k TPS for Llama 3.2 multimodal on proprietary hardware, and ran ML training experiments across medical and document intelligence applications.

Engineering for early-stage ML startups. At WavvAI, built a copyright-free synthetic MIDI data pipeline grounded in musical first principles, then pretrained an autoregressive model — Musica — that generates production-ready EDM, validated with Grammy-winning musicians. At AnyCart, built the grocery recommendation system. At predictABill, built a scraper that parses and summarizes disparate health insurance policies.

Generalized pathology computer-vision models to new labs and scanners across five AI products, generating ~$5M in new revenue and cutting ML development time by 25% — four ArXiv preprints in three quarters (SC-MIL, S-DOTA, ContriMix, self-training for liver histopathology). Developed Multiple Instance Learning and graph neural network models to predict molecular biomarkers (RAS+RAF, ROS1, cMET, MSI) for lung and colorectal cancer. Built foundation tissue segmentation and cell classification models on par with disease-specific models, cutting deployment from weeks to days.

Initiated the COVID-19 stochastic contact-network modeling effort that became the COVID Calculator — a design-award-winning tool that enterprises used to inform return-to-office policy (PLOS One; patent). Streamlined data, model, and code versioning to cut model delivery time in half. Led the technical work behind the $100M SOMPO investment for disaster resilience in Japan.

Developed a novel CNN + NER + ASR combined algorithm for real-time multimedia retrieval — 1.5× accuracy and 10× latency reduction. Built an AWS-based MLOps platform and open-source experiment / environment tracker for CV, NLP, and traditional ML, deployed at scale to 1M+ end users.

At Berkeley (2011–2012), worked on computational materials science and quantum physics — first-principles simulation of electronic structure and device behavior. At Stanford (2012–2015), shifted to AI research, contributing to projects in computer vision, NLP, and machine-learning applications to biology.