Session

Building and Scaling a Multi-Agent App Platform

Overview

ExperienceIn Person
TrackApplication Development
IndustryEnterprise Technology, Financial Services
TechnologiesUnity Catalog, Databricks Apps, Databricks Agents
Skill LevelIntermediate
DOWNLOAD SESSION SLIDES

How we designed, scaled and shipped a multi‑agent AI app platform for Skuld — built 100% natively on Databricks — and the unified App, AI and data patterns that made it possible. Maritime insurance claims depend on shifting legal facts, mixed‑format case files and strict audit requirements, and standard chunking plus top‑k filtering lost provenance and broke under load. At Skuld, we built a multi‑agent app on Databricks: Lakeflow Jobs runs parsing and enrichment pipelines, Model Serving powers synchronous agent calls, Databricks Apps handles secure user workflows and Unity Catalog governs access end-to-end. The talk dives into our evidence‑first retrieval pattern, where documents become source‑linked fact cards, conflicting facts stay visible for review and agents return exact citations instead of opaque answers. Attendees leave with a hosting decision framework and benchmark data on parsing throughput, queues, token cost, rate limits and latency.

Session Speakers

Domonkos Pal

/Director of Digital Product Development
Hiflylabs

Miran Hadziomerovic

/AI Lead
Skuld

Full Summary

Scaling marine insurance claims with multi-agent AI at Skuld

Marine insurance claims span multiple jurisdictions, vast document sets, and timelines measured in years. Individual handlers juggle dozens of such cases in parallel, each with contracts, surveyor reports, and correspondence that can run from hundreds to tens of thousands of files. The conversation outlines how Skuld is building a multi-agent stack on the Databricks Platform to lift that cognitive load without compromising traceability, control, or trust.

FAQ


RAG is effective for targeted Q and A over bounded datasets, but claims work requires synthesizing tens of thousands of documents across long timelines. Prompt inconsistencies and LLM non-determinism lead to unstable answers at scale, so a deterministic multi-agent approach proved more reliable.

It means keeping routing, storage, and orchestration deterministic, and using AI only where it provides clear value, such as extracting insights. Each agent has a narrow remit and strict guardrails, which limits the surface where model unpredictability can cause issues.

Skuld builds gold-standard comparisons from closed cases handled by experts, asks fixed questions, and compares expert answers to system outputs. Evaluation is embedded at each pipeline step so changes to prompts or agents can be traced to their impact, positive or negative.

User overrides are recorded as separate, first-class records rather than overwriting agent outputs. Downstream agents consume a view where user overrides win, while upstream components preserve both original and corrected versions to prevent future runs from undoing human corrections.

Three options are used: model serving for scalable endpoints, Lakeflow Jobs for pipeline-integrated workloads where latency is less critical, and Databricks Apps for low-latency production hosting with built-in access control. In practice, teams often mix these based on latency and integration needs.