AI at The Bench: Getting More From Each Experiment

Author: Jeremy Elser, Head of Science, Palantir Collection: UNLOCK-2026-Sessions

Summary

Jeremy Elser, founding CEO of Shipthesis (a biologics biotech) and forward deployed engineer (FDE) at Palantir, demonstrates how his lab implemented AI-powered data capture, experiment management, and protocol generation using Palantir Foundry. The central problem he addresses is institutional knowledge loss: when a scientist leaves an organization, the value of their experimental work degrades rapidly because critical metadata — reagent concentrations, batch identifiers, procedural nuances — is never captured in sufficient detail. Scientists lack the time and inclination to fully document their work, and even the most diligent researcher cannot anticipate what future questions will need answering.

The solution Elser built on Palantir Foundry for Shipthesis (referred to internally as "Ichor") combines an electronic lab notebook (ELN) integration with eLab, AI-powered extraction and classification of experimental metadata, an ontology connecting all laboratory objects (experiments, protocols, scientists, customers, journal articles), and an AI agent called the "Automatic Jeremy" that proactively queries scientists for missing metadata. In return for scientists' cooperation, the system automatically generates charts and statistical analyses — including Kaplan-Meier survival curves — across all historical experiments. The system classifies experiment types (e.g., cell proliferation assay) and identifies proteins/drugs used without requiring explicit labeling by scientists.

Elser extends the system to document summarization and IND (Investigational New Drug) application evaluation: the AI can read charts embedded in final reports, summarize studies, and assess how close a drug program is to meeting FDA guidance for a given indication. He also demonstrates the "Logic" module in Palantir Foundry as a prompt-chaining workflow builder that, given access to journal articles and FDA guidance, generates complete experimental protocols including animal model selection and positive control choices (e.g., minoxidil for an alopecia study). His broader thesis is that this infrastructure could help address the reproducibility crisis in science, since the root cause of irreproducible experiments is often tacit procedural knowledge that was never recorded — not scientific fraud.