The Engine Of AI Trust: Building Dependable Frontier Models

Sushant Mehta from Google DeepMind highlights the importance of engineering discipline in developing dependable AI models. By automating evaluations and embedding user feedback, he aims to transform theoretical AI performance into real-world success.

In today’s AI landscape, where new models debut faster than features can ship, the real bottleneck is not capability but dependability. The gap between a compelling research demo and a system that millions can rely on is where trust is either built or broken. This is not a problem of raw model intelligence, but one of engineering rigor.

At the heart of this challenge sits the important discipline of post-training large language models. This is where raw model potential is refined into predictable behavior. It is a discipline of constraints, metrics and relentless validation.

Few understand this frontier better than Sushant Mehta, a Senior Research Engineer at Google DeepMind. As the quality lead for Gemini’s coding and tool-use features, Mehta spearheaded the launch of the Gemini Data Analysis Agent and built the critical infrastructure that turns research breakthroughs into production-ready features. His work focuses on a deceptively simple goal: making AI more powerful and, better still, more accountable.

“The leap from a promising model to a trustworthy agent is solved by more than mere scale,” Mehta shares. “It’s solved by engineering discipline. You are more than just tuning parameters. You are building a system for measurable, reproducible trust.”

The Real-World Stakes of AI Evaluation

AI performance is often theoretical until it collides with a user’s task. A model might ace an academic benchmark but fail to write usable code. It might summarize a document perfectly, yet misinterpret a user’s nuanced request. This is the “last-mile” problem for AI, where failure can cause user friction and erode user trust.

This was the core challenge for the Gemini Data Analysis Agent and Gemini-GitHub integrations, which he led to launch. The goal was more than just correct code generation; it was enabling a user to analyze repositories, manipulate spreadsheets and generate charts through natural language, helping users work faster and more intuitively. Success had to be reframed from accuracy to real user outcomes.

“Benchmarks give you a false sense of security,” Mehta observes. “The real test happens when a user's workflow depends on your model. That's where we moved from abstract accuracy to tangible user success.”

Mehta’s approach was to close the feedback loop between training and real-world impact. By embedding outcome-based signals directly into the reinforcement learning process, his team shifted the model’s optimization target from static metrics to live user success. The result was not just more features, but a new engineering doctrine: measurable progress must replace subjective confidence.

Scaling Trust Through Automation: The AutoRater System

For most AI teams, evaluation is the silent bottleneck. Human review cycles stretch for weeks, stalling iteration and leaving even the most advanced models waiting for human validation. Mehta’s answer was to automate what could no longer scale manually.

This philosophy was a central theme of his discussion as a featured speaker and 1v1 interview at the AI Infra Summit, where he detailed the transition from manual checks to automated, AI-driven validation systems.

“Without rapid feedback, model development is navigation without a compass. You might be moving, but you have no idea if you're heading toward a cliff or solid ground. AutoRaters gave us that directional certainty, transforming a blind exploration into a guided journey.”

He helped design and deploy AutoRaters, an intelligent evaluation system that uses AI to benchmark AI outputs with human-grade fidelity. AutoRater compares model responses to curated standards, quantifies alignment and continuously recalibrates its own scoring through supervised feedback. The result: evaluation loops that once took weeks now conclude within hours.

This automation did more than just accelerate reviews, it rewired the experimentation pipeline of post-training teams at DeepMind. Teams could now run hundreds of training iterations in the window previously allotted for one or two. The savings in time and cost were matched by a qualitative gain: reproducible, data-driven confidence in every launch.

From Code to Culture: Trust as an Organizational Capability

Engineering trust at scale requires more than sophisticated tools; it demands a cultural shift where evaluation evolves from a final checkpoint into a continuous, organizational capability.

Mehta’s framework for scalable evaluation brought together research, product and compliance under a shared language of metrics. At DeepMind, this meant creating end-to-end dashboards that linked training experiments to launch metrics and quality reports. These systems made it possible to trace performance regressions back to data versions, detect inconsistencies automatically and enforce privacy constraints across all evaluation stages.

This systems-thinking mindset is rooted in his co-authored scholarly paper titled, “Unifying Mixture of Experts and Multi-Head Latent Attention for Efficient Language Models,” where he explores novel architectures that fundamentally improve the trade-off between model performance and computational efficiency, a core challenge in deploying large-scale AI reliably. At Google Maps, Mehta led privacy-safe personalization, introducing On-Device Location History and differential privacy techniques. This allowed models to learn from global patterns without ever exposing individual data, a foundational principle that now also governs DeepMind’s own model accountability strategies.

“We were more than just unlocking model capabilities,” Mehta notes. “We were building a system where a lapse meant a breach of trust or a failed user task. That changes how you design everything.”

Rethinking What 'Dependable AI’ Really Means

AI’s determining factor in the future, will not be the biggest model, but rather the most dependable one. Trust cannot be an afterthought; it must be the core material from which these systems are engineered.

“The next breakthrough won't come from a bigger model,” Mehta, who also serves as an academic paper reviewer for the NeurIPS 2025 conference, concludes. “It will be a more trustworthy one. We're moving from an era of demonstration to an era of dependability.”

Engineers like Sushant Mehta are defining this path, from abstract concepts to concrete launches, through operational discipline.

Because when a system touches a user's work, trust is no longer optional. It is engineered.