Medlyze AI Data Assistant

Introduction & Background

Medlyze entered the market focused on taming mandated healthcare price transparency files. Their tooling already served data-savvy users, but adoption lagged with analysts, policy teams and consultants who needed quick answers without writing SQL. tuniverstudio, fresh off AI agent deployments for boutique retailers, partnered with Medlyze to prototype a conversational interface that could prove the value of AI-assisted analytics to their broader customer base.

• Strategic push to make complex pricing benchmarks self-serve
• tuniverstudio's mission: accessible AI copilots for regulated industries

Problem & Constraints

Healthcare price transparency data is noisy, high-volume and compliance-sensitive. Medlyze needed a way for non-technical teammates to explore negotiated rates, spot outliers and export usable views without risking misinterpretation or exposing sensitive partner data. Any assistant had to cope with the idiosyncrasies of payer/provider terminology, gracefully recover from malformed queries and respect the guardrails of public versus private datasets.

• Conversational UX had to map to intricate schemas and healthcare vernacular
• Compliance rules limited which datasets could power demos and marketing collateral

Solution Overview

We architected an AI analyst copilot that goes end-to-end: it interprets natural language, composes SQL, validates the response shape and presents insights as narrative-ready visuals. Suggested prompts help first-time users explore safely, while contextual disclaimers explain how each answer was generated. The assistant can export both raw tables and presentation-ready charts, ensuring downstream teams can continue their workflows without friction.

• Plain-English queries route through a guardrailed LangChain orchestration
• Dynamic visualisation picker chooses the most legible chart for each insight

Implementation Journey

We grounded the LLM in Medlyze’s canonical BigQuery schemas and added schema-aware planning so the assistant only touches approved tables. A retry handler diagnoses BigQuery errors, corrects column names and resubmits without surfacing raw stack traces to end-users. Front-end guidance, including spotlighted suggested queries and state-aware buttons, keeps the experience discoverable.

• LangChain tooling with deterministic prompt templates and schema tags
• Anthropic Claude during experimentation, tuned towards OpenAI GPT for launch
• SvelteKit shell embedding the assistant alongside knowledge panels

Results & Early Signals

Early pilot users cut the time to first meaningful insight from hours to minutes. Teams outside core data engineering now explore reimbursement patterns independently, supporting Medlyze’s positioning as an indispensable pricing intelligence partner. Public demo datasets power safe conference demos and marketing videos without risking PHI leakage.

• Faster analyst onboarding and reduced dependency on SQL specialists
• Demo environments seeded with anonymised public data for go-to-market impact

Looking Ahead

With the Medlyze launch, tuniverstudio has a reusable blueprint for regulated analytics copilots. Next on the roadmap: deeper integration with Medlyze Mentor, enterprise-ready RBAC and customer-specific fine-tuning. The collaboration showcases how lean AI teams can unlock outsized value with thoughtful orchestration and empathetic UX.