Two connected systems for a commercial foodservice equipment manufacturer. Answers grounded in service documentation, followed by claims recommendations with reasons, confidence and human review.
Client identity kept private. Implementation and results from the original engagement.
Better decisions. Inside familiar tools.
Field technicians needed answers from dense service documentation. Manual lookup was slow, uneven and difficult to keep consistent across the field.
Warranty claims review had a similar problem. Incoming claims took human time, outcomes could vary by reviewer, and repeated decisions created fatigue.
Both systems needed to fit into the existing support environment. The technician assistant shipped first, followed by a claims workbench alongside it.
Two systems. Delivered in sequence.
System 1
Technician Support Assistant
Field technicians could ask natural-language diagnostic questions and get answers grounded in the manufacturer's own service manuals and documentation.
The system ran on a serverless AWS architecture. A submitted question entered a queue, a worker retrieved relevant manual context from a vector knowledge base, and an LLM generated a grounded answer.
The assistant was integrated into the existing CRM/support platform, so technicians did not need to learn a disconnected tool.
The second system evaluated incoming warranty claims and recommended Approve, Deny, or Investigate.
Each recommendation included High, Medium, or Low confidence plus written reasoning. It was designed for review, not blind automation.
The system compared each claim with device error history and similar past claims using Aurora Serverless with pgvector as a precedent store.
Approve, Deny, InvestigateConfidence scoringAurora Serverlesspgvector precedent store
From request to review.
Question or claim
Submitted in the existing support platform.
Queue
Work is routed without blocking the user experience.
Context
Relevant manuals, device history and precedent are retrieved.
Model
Amazon Bedrock generates an answer or recommendation.
Workbench
A person reviews, confirms or overrides with a reason.
AI recommends. People decide.
Reviewers see the recommendation, confidence level and written reasoning before they act.
Visible confidence
Recommendations are marked High, Medium or Low so reviewers know where to apply more judgment.
Controlled bulk review
Reviewers can confirm recommendations in batches. Overrides require a documented reason.
A record that improves future reviews
Overrides are logged and fed back into the precedent layer, reflecting real human judgment.
A reversible rollout
Feature flags, smoke tests and rollback notes protected the live technician assistant during the second launch.
What the system delivers.
Evaluation time and AI costs depend on the claim and the context retrieved. These figures describe this implementation.
5-12 sec
Typical claim evaluation time once the claim enters the AI review flow.
$0.01-$0.03
Approximate AI cost per claim evaluation, depending on claim size and context.
Full audit trail
Every recommendation, confidence level, reason, and override is recorded.
Zero disruption
The claims system shipped as an additive layer beside the live technician assistant.
What to plan for.
Ship the smaller system first
The technician assistant was useful on its own. Treating it as a standalone delivery made the second project easier to support internally.
Give precedent data an owner
The pgvector precedent store depends on the claims fed into it. A client-side owner was responsible for identifying bad precedent early.
Allow time to calibrate confidence
High and Medium thresholds were adjusted twice in the first month as reviewers worked through real cases. That adjustment period belongs in the delivery plan.