Although AI systems are changing from simple tools into comprehensive collaborators, most infrastructure isn’t built to support models that plan, reason, or act with autonomy. In high-stakes fields like insurance, where inputs are fragmented and every decision carries inherent risk, that becomes difficult to ignore.
It’s the unfortunate intersection of legacy systems with slow, unpredictable, human-centered decisions. The required data exists, but it’s rarely structured or accessible in a way that AI systems can use reliably.
AI engineer Timofei Vasilevskii is working to change that at Strala, an agentic AI platform that retrieves, interprets, and acts on live claims data. Timofei leads the design aspect, focusing on making state-of-the-art methods, such as retrieval-augmented reasoning, dynamic memory systems, and schema-guided prompting, reliable enough for production in a complex, regulated setting.
With years of experience in machine learning research, competitive programming experience, and a deep understanding of applied systems thinking, Timofei is building what could become a reference design for deploying frontier AI methods into real-world workflows.
Here’s a closer look.
Insurance as a Stress Test for Autonomous AI
Few areas test the limits of autonomous AI like insurance. Every claim is different and often built from fragmented inputs, yet decisions must be consistent, auditable, and aligned with both policy and regulation.
Insurance is a rare convergence of widespread uncertainty, tight regulation, and real human consequences. Data often arrives incomplete or ambiguous, with lives and livelihoods frequently at stake.
A single claims process can span multiple channels, documents, and human interactions. AI systems operating in this field need to reason, adapt, and flag uncertainty that requires human judgment, which is what makes the insurance industry a fruitful proving ground for agentic AI.
“Insurance has a blend of high variability and high-stakes decision-making,” Timofei explains. “Claims vary wildly between cases, yet must follow strict protocols. Building AI here forces you to balance automation with empathy and precision.”
A New Frontier in AI Systems: Building for Agent Autonomy
Strala’s platform enables software agents to operate independently across the complex stages of insurance claims processing. While most AI systems only respond to single-step inputs, these agents can plan multi-step workflows, dynamically retrieve relevant information, and trigger real-world actions within tightly constrained operational processes.
To achieve this, Strala has built a full-stack system for agent autonomy. Retrieval-augmented planning enables agents to integrate reasoning with real-time data retrieval, allowing them to accurately and efficiently handle incomplete or fragmented data. The platform integrates agents with external workflows and APIs, allowing the AI to act rather than just generate text. The system also includes autonomous error recovery, which automatically adjusts workflows when agents encounter unexpected inputs or failures.
“We’re pushing the boundaries of multimodal AI,” says Timofei. “It’s a blend of LLM-centric design with robust systems thinking.”
A great example is how agents can review partial claims documents and decide when to escalate ambiguous cases to a human. This kind of dynamic judgment helps to balance automation with necessary oversight. Rather than focusing solely on cutting-edge AI research, the emphasis is on grounding processes for reliable production use, which is particularly important in a complex and regulated field like insurance.
Inside the Architecture: Building a Semantic Layer for Agentic Workflows
Timofei designed Strala to turn the ambiguity and variability of insurance data into machine and agent-readable signals. As he puts it, “it’s built to abstract away backend complexity and provide agents with a clean, reliable interface to the real world.”
This starts with a real-time ingestion layer that captures a range of inputs, including documents, call transcripts, images, and audio, before standardizing them for processing. From there, a semantic vector search layer can find relevant information by meaning rather than keywords, indexing the data and using schema-aligned metadata to make it more accessible to agents using structured queries.
Using that data, agents can operate through a memory-augmented task framework, which allows them to retain context across steps, reference prior actions, and adjust to new information as claims develop. In insurance, where cases can evolve over weeks and even months, this digital memory becomes critical.
Every layer of the system reduces friction, exposes meaning, and supports more consistent behavior within and across cases. An accident claim with an incomplete police report, a voicemail transcript, and several photo uploads can be clarified with inputs from past claims that had similar metadata, as well as a judgment on whether required documentation is needed to make a decision.
By the time an agent decides a claim, it is working with a rich and structured view of the case.
From Sandbox to Production: Deploying AI With Control
Bringing experimental AI into production environments is one of the hardest problems in applied machine learning.
In industries like insurance, systems must meet high operational standards before they ever touch real data, which is why new features always follow a structured path at Strala. Each one starts in an internal sandbox, progresses to a functional prototype, and then undergoes staged testing before being deployed into the core platform. “Speed matters, but not at the cost of stability or trust,” says Timofei.
The company’s process for bringing research-grade methods into a production-grade environment reflects that mindset. LangGraph is used to define agent workflows with clear logic and recoverable states. Supabase manages storage and access control, while LlamaIndex handles data pipelines across structured and unstructured sources. Finally, the system utilizes various AI models from providers such as OpenAI, Anthropic, Cohere, and Mistral, selecting the best fit for each specific task.
None of these components is treated as a black box. Instead, they’re selected and configured based on their ability to perform under load and integrate. Any experimental features have to fit within the broader stack rather than being retrofitted after the fact.
At Strala, reliability is a design requirement, not added as an afterthought.
Meet the Builder: Timofei Vasilevskii’s Journey to the Edge of AI Deployment
Timofei Vasilevskii began experimenting with machine learning during high school, driven by a deep interest in technical problem-solving. He studied computer science at a top Russian University and trained at the Yandex School of Data Analysis. He held roles in AI research, quantitative development, and software engineering before joining Strala at the age of 22.
As the company’s founding AI engineer, he now leads the system architecture for its core platform. By bringing state-of-the-art methods into real-world deployment, Timofei is shaping the future of AI where it matters most: under real constraints, in real environments.
“I’m excited by the challenge this domain presents,” he concludes. “After gaining industry experience and learning from senior colleagues, I felt ready to build AI software from the ground up. Being the founding engineer means dedicating myself fully to what I believe is the right approach.”
Kyle Lewis is a seasoned technology journalist with over a decade of experience covering the latest innovations and trends in the tech industry. With a deep passion for all things digital, he has built a reputation for delivering insightful analysis and thought-provoking commentary on everything from cutting-edge consumer electronics to groundbreaking enterprise solutions.
