I am Darpan Ganatra, an applied AI engineer and ML leader. I build production LLM agents, evaluation systems, and research infrastructure for work that requires traceable evidence and predictable failure behavior.
What I build
My work spans applied AI and research engineering: agent architecture, tool and context design, constrained generation, trajectory-level evaluation, representation analysis, and artifact provenance.
At Publicis Health, I built the generative AI practice from scratch and remain hands-on while leading a team of five. A production clinical-reasoning system supports more than 20 pharmaceutical brands and reduced audience-development time from roughly 20 hours to roughly 2 hours.
Production evidence has changed several of my initial design choices. One monolithic agent became a multi-agent system, then moved back toward a single agent with deterministic workflows when that proved easier to operate reliably. An analytics agent emits a constrained DSL that cannot represent disallowed operations. The compiler enforces that boundary.
Public research engineering
An auditable investigation of cross-model representation geometry. A repaired one-pair pilot and preregistered bidirectional sensitivity study found stable linear translatability between Qwen2.5-0.5B and TinyLlama-1.1B, with a clear directional asymmetry across 20 held-out splits.
A fail-closed foundation for testing whether stable demographic representations causally transport into human-aligned subgroup response differences. The repository contains statistical and provenance infrastructure. No scientific finding is reported yet.
How I evaluate the work
For both model experiments and production agents, I want to know what was fixed before a run, what evidence would have changed the decision, and whether the artifact can be reproduced from its provenance. The result should also state precisely where the evidence stops.