Background
Srinivasan is an engineer at Google DeepMind who led engineering on the Gemini Image / Nano Banana line. Her public appearance is the Sequoia podcast with Brichtova on how the model achieved character consistency, where she is candid that the qualities users actually care about — likeness, aesthetic quality — resist quantification, and that the team's method was obsessive focus on specific problems such as text rendering rather than a general metric.
That is a rare admission from an engineering lead and a useful one: it means the internal evaluation stack for the most successful image model of 2025-26 rests on human judgement by design, not by omission.
What they run now
Engineering for the Gemini Image model family: training and evaluation infrastructure, the specific capability pushes (character consistency, text rendering) and the human-evaluation processes the team trusts over benchmarks.
What they buy
Inference, anchored in her own statement of the problem. What an engineering lead in her position needs is graded evaluation sets that behave like benchmarks but encode human judgement: stable, versioned, with per-item rationale and known inter-rater agreement, so a model change can be scored without convening a room of people every time. That is a different product from raw preference pairs and a more defensible one. Also plausible: adversarial or diagnostic sets for specific failure modes (text rendering at small sizes, identity drift across edits).
Career
- presentSoftware engineer → engineering lead, Gemini Image (Nano Banana), Google / Google DeepMind