Inside Google’s AI Accelerator ft. Andrew McKishnie

This is a summary of an episode of Pioneers, an educational podcast on AI led by our founder. Join 3,000+ business leaders and AI enthusiasts and be the first to know when new episodes go live. Subscribe to our newsletter here.

TL;DR:

• Multimodal was selected for Google’s AI Accelerator for building agentic AI in highly regulated industries like finance and insurance.
• Andrew emphasizes the importance of Gemini 2.5 Pro and synthetic data in scaling AI without customer data bottlenecks.
• Multimodal is shifting from custom builds to out-of-the-box offerings to better serve mid-market and enterprise customers.
• Deep customer relationships allow Multimodal to embed institutional knowledge directly into AI workflows.
• Google’s support—from mentorship to co-selling via Marketplace—is accelerating both product development and go-to-market motion.

‍Before we dive into the key takeaways from this episode, be sure to catch the full episode here:

Meet Andrew - VP of Engineering

Andrew McKishnie, VP of Engineering at Multimodal, brings a background in linguistics and a passion for language models to the world of agentic AI.

He joined Multimodal as an individual contributor and quickly rose to lead engineering, shaping everything from product design to customer success. Andrew plays a key role in integrating Google’s AI stack, like Gemini, into Multimodal’s agent platform, while also strengthening how institutional knowledge gets embedded into AI workflows.

An advocate for thoughtful engineering and high-impact deployment, Andrew is focused on scaling Multimodal’s platform to serve both enterprise and mid-market clients through more composable, out-of-the-box solutions.

From mentoring sessions to bootcamp insights, Andrew leverages the Google AI Accelerator experience to ensure agentic AI meets the operational realities of finance and insurance.

Inside Google’s AI Accelerator: Why Multimodal Was Selected

Andrew explains that Multimodal was selected for Google’s AI Accelerator because of its focus on building agentic AI for highly regulated industries like banking and insurance.

Google recognized that Multimodal had real enterprise traction and a practical go-to-market strategy, not just theoretical demos.

Andrew says the selection process involved multiple interviews where Multimodal’s ability to translate business needs into actual product was clear.

What stood out was Multimodal’s deep focus on verticalized problems and the credibility of its team. Being chosen by Google gave the team access to new tools, mentorship, and co-selling opportunities.

“The accelerator has been a really, really cool forcing function.” — Andrew McKishnie

For Andrew and the engineering team, it also meant the chance to pressure-test ideas with Google’s top AI engineers and integrate cutting-edge models like Gemini 2.5.

Gemini 2.5 Pro and the Power of Synthetic Data

One of the key benefits Andrew highlights from being in the accelerator is access to Gemini 2.5 Pro.

He explains that the model’s longer context window has been a game changer for finance and insurance workflows. It allows agents to reason across multiple documents without splitting up context.

Andrew also describes how Multimodal uses synthetic data to train and test agents before real client data is available. This lets them simulate client-specific workflows without waiting for lengthy approvals.

The engineering team uses this approach to create agents that are production-ready by the time real documents arrive.

“What we’re learning from Google’s engineering team is bleeding into how we build everything.” — Andrew McKishnie

Combining Gemini’s capabilities with synthetic training reduces deployment time and helps the team build safer, more capable agents from day one.

From Custom Builds to Out-of-the-Box Agentic AI

Andrew described a shift in Multimodal’s product strategy. Originally, every agent was a custom build for each enterprise client. But now, the team is investing in composable agent templates that allow faster deployment and reuse.

He notes that this change was accelerated by insights from the Google AI Accelerator, where other startups and mentors emphasized the importance of balancing customization with scalability.

“Multimodal’s new approach focuses on defining common workflows and packaging them into modules that can be quickly configured. “

This lets the company serve both large enterprises and mid-market clients. Andrew sees it as a way to bring the benefits of agentic AI to more organizations without sacrificing reliability, quality, or speed of implementation.

Scaling Without Real Data: How Synthetic Training Accelerates Go-Live

Andrew explains that one of the biggest bottlenecks in AI deployment is waiting for client data.

Legal, compliance, and procurement reviews can delay access for weeks or months.

To solve this, Multimodal uses synthetic data based on public records and structured test cases to begin agent training in parallel. Andrew shares that this technique allows the engineering team to simulate how the agent will interact with documents, test edge cases, and refine the behavior long before client onboarding finishes.

When access is finally granted, agents are already tested and can be deployed quickly. This significantly shortens the time from contract signing to go-live and ensures agents are better prepared to operate in complex real-world scenarios.

Deep Client Embedding: The Secret to Institutional Knowledge Transfer

A key theme Andrew returns to is the idea of deep client embedding.

He shares that the best AI agents are not just technically accurate but also reflect the tone, culture, and processes of the client.

The Multimodal team achieves this by pairing engineers directly with customer-facing leads during early deployments. Andrew explains that this hands-on approach allows engineers to observe the real pain points and workflows their agents will be augmenting.

Over time, this results in agents that are more trustworthy and aligned with how a client actually operates.

“The agents we’re building reflect the institutions they work inside.” — Ankur Patel

By training agents with specific examples from the client’s workflow, they can help retain and scale institutional knowledge in ways that static documentation cannot.

How Google’s Mentorship Model Shapes Engineering Culture

Andrew speaks highly of the mentorship he and the team received through Google’s AI Accelerator.

Weekly technical sessions gave them feedback on everything from agent design patterns to compliance tradeoffs. He highlights how hearing from Google engineers about how they make infrastructure decisions at scale helped the Multimodal team reconsider how they manage complexity.

One example he gives is rethinking how they handle edge cases within agents, borrowing ideas from large-scale customer service implementations. This exposure influenced not only their product but also how they collaborate internally.

Andrew says the accelerator helped reinforce a culture of iterative development, faster experimentation, and more structured engineering reviews across the team.

Engineering for Compliance in Finance and Insurance

Andrew shares that building AI for finance and insurance requires a mindset focused on trust and auditability.

He explains that enterprises want to understand how an agent arrived at a decision, what data it accessed, and what steps it took.

This requires engineering teams to go beyond performance and prioritize logging, transparency, and approval workflows. Andrew says Multimodal builds agents with these guardrails in place from day one.

“Our agents are built with guardrails from day one to support enterprise compliance.” — Andrew McKishnie

Rather than seeing compliance as a blocker, they treat it as a design constraint that leads to better, safer systems.

He notes that their ability to support complex audit requirements and chain-of-custody expectations is part of what makes their product suitable for regulated industries.

What’s Next: Gemini-Powered Deployments and One-Click Setups

Looking ahead, Andrew says the goal is to make it even easier to deploy agents. One major initiative is using Gemini models to power agents across more steps of the workflow, not just summarization or data extraction.

He also mentions a push toward one-click setup for common use cases, especially for mid-market customers.

“That’s the future—Gemini powering real enterprise workflows.” — Ankur Patel

The idea is to remove friction during onboarding by offering prebuilt workflows that can be activated with minimal configuration. Andrew believes this is key to expanding beyond highly customized enterprise deployments and reaching a broader customer base.

With Gemini and better tooling, the team is focused on faster launches, easier integrations, and lower overhead for customers adopting agentic AI.

Interested in learning more about Multimodal and our other partnerships? Check out our accelerator partnerships and how we’re driving innovation and leading the future.