Case Studies

Customer snapshot

Eli Lilly is a global pharmaceutical company and one of the world's largest producers of insulin and other critical medications. As part of an AWS ProServe engagement, Cloudtech was brought in to automate and scale Lilly's Innovation Readiness Assessment process — a complex, manual workflow requiring significant effort from domain experts across multiple data sources.

Goal

Design and implement an enterprise Agentic AI platform on AWS that automates Eli Lilly's Innovation Readiness Assessment process — replacing manual SME effort with a standardised, scalable, evidence-backed AI workflow.

Key results

Challenge

Manual assessments could not scale

Eli Lilly's Innovation Readiness Assessment required domain experts to manually collect, analyse, and evaluate information across internal, public, and licensed data sources. Each assessment was time-consuming, difficult to standardise, and challenging to replicate across different products and therapeutic areas.

Evidence was fragmented and inconsistent

Information lived across enterprise knowledge bases, public APIs, licensed data sources, and proprietary tools. Without a unified retrieval layer, assessments lacked traceability and consistency — making it difficult for SMEs to validate findings or understand how conclusions were reached.

SME oversight could not be eliminated

Despite the need for automation, human expert review and override capability had to be preserved. The solution needed to augment SMEs, not replace them — with confidence scoring, source evidence, and human-in-the-loop controls built into every assessment.

Solution

Cloudtech designed and implemented an enterprise Agentic AI platform inside Eli Lilly's AWS environment. Eight specialised AI agents operate in parallel to evaluate IRA attributes, retrieve supporting evidence, generate structured scores and rationales, and consolidate findings into a comprehensive readiness assessment.

Multi-agent orchestration

A multi-agent architecture using LangGraph and LangChain enables eight specialised domain agents to execute in parallel and consolidate their findings into a standardised IRA assessment — reducing assessment time from days to hours.

Evidence-grounded retrieval

RAG implemented across enterprise knowledge, public APIs, licensed data sources, and custom data tools provides traceable, contextually relevant insights — with confidence scoring, contradiction detection, and coverage-gap identification built into every output.

Human-in-the-loop controls

Domain SMEs can review, validate, and override AI-generated results at any point. Persistent assessment history and conversational refinement allow SMEs to interact with completed assessments and understand the reasoning behind every finding.

Scope + Timeline

‍

Outcomes

• Eight specialised AI agents operating in parallel — replacing a manual, SME-intensive process with a standardised, repeatable workflow
• Evidence-grounded retrieval across enterprise knowledge, public APIs, and licensed data sources — with full source traceability on every finding
• Confidence scoring, contradiction detection, and coverage-gap identification built into every assessment output
• Human-in-the-loop controls maintained — SMEs retain full ability to review, validate, and override AI-generated results
• Persistent assessment history enables conversational refinement — SMEs can interrogate findings and understand reasoning
• Reusable enterprise Agentic AI pattern established — extendable across products, therapeutic areas, and other knowledge-intensive workflows

Technology stack

Agentic AI: LangGraph, LangChain, Claude Models
RAG and Embeddings: Retrieval-Augmented Generation, Embedding Models, Evidence-Grounded Generation
Backend: Python, FastAPI, REST APIs
Data and Storage: Amazon DynamoDB, Amazon S3
Deployment: Amazon EKS, Docker, Amazon ECR
CI/CD: GitHub Actions, ArgoCD, AWS CloudFormation
Monitoring: Amazon CloudWatch
Interface: React.js Web Portal, Microsoft Entra ID Authentication

Want similar results for your business?

Schedule a call with the Cloudtech team to scope an Agentic AI engagement for your use case. We typically go from baseline to production in four to six weeks.

‍

Customer snapshot

Ascend BPO is a US-based healthcare business process outsourcing company managing appointment scheduling and
patient intake for healthcare providers across the United States. With high inbound call volumes and strict HIPAA compliance
requirements, Ascend BPO needed an AI voice agent capable of handling thousands of monthly calls accurately, securely,
and with seamless escalation to human agents when needed.

Goal:

Deploy a HIPAA-compliant AI voice agent on AWS to automate inbound appointment scheduling at scale — handling 2,500–
5,000 calls per month with intelligent human escalation.

Key results

‍

Challenge

High call volume, limited staff capacity


Ascend BPO's healthcare clients receive thousands of appointment scheduling calls every month. Human agents were
spending the majority of their time on routine intake workflows that could be automated — creating bottlenecks and
inconsistent caller experiences

HIPAA compliance at every layer


Every interaction involved Protected Health Information — patient names, dates of birth, insurance IDs, and appointment
details. Any solution needed end-to-end HIPAA compliance covering secure storage, encrypted transmission, access
controls, and full audit trails.

Unpredictable real-world callers


Real callers don't follow scripts. Confused patients, incorrect insurance details, urgent booking requests, and off-topic
questions are all part of everyday call volume. The system needed to handle each scenario gracefully — and escalate to a
human agent without losing context or forcing the caller to repeat themselves.

‍

Solution

Cloudtech designed and deployed a HIPAA-compliant AI voice agent inside Ascend BPO's AWS environment, built on
Amazon Bedrock, Amazon Transcribe, and Amazon Polly. The agent handles the complete appointment scheduling workflow
end to end — from greeting to confirmed booking — without requiring a human for routine calls.

‍

‍

Scope + Timeline

‍

Outcomes

• ,5 00–5,000 inbound calls per month handled autonomously — no human required for routine scheduling
• Identity and insurance verification completed in under 60 seconds per call
• Warm transfer to human agent in under 2 seconds, with full context handed over seamlessly
• Fully HIPAA-compliant — encrypted PHI handling, access controls, and audit trail on AWS
• Stress tested across complex real-world scenarios including confused callers, aggressive behaviour, tangent
  conversations, and urgent bookings
• Complete patient lifecycle supported — create record, verify and update insurance, check availability, book and confirm
  appointment

Salesforce-to-S3 Data Pipeline

Source One Spares needed a secure, cost-effective way to back up Salesforce data to AWS, without managing complex
infrastructure. They required encryption, auditing, compliance controls, and a solution their on-prem team could use with
their existing stack.

Goal:

Build a production-ready AWS backup pipeline, no compute overhead, no ongoing management, no complexity.

Key results

‍

Challenge

No infrastructure overhead

Source One Spares didn't want to manage servers, Lambda functions, or complex pipelines. They needed a backup solution that just works, set it and forget it.

‍

Enterprise-grade security


Salesforce data is sensitive. They required encryption at rest and in transit, strict access controls, and full audit trails for compliance.

‍

Cost efficiency at scale


Storing years of Salesforce backups can get expensive fast. They needed intelligent tiering to minimize costs without sacrificing access when needed.

‍

What We Built

A production-ready AWS backup pipeline; zero compute, zero complexity.

‍

Cost Optimization

‍

Deliverables

• Architecture diagrams ( VPC + S3 backup flow)
• IAM & Access Runbook with sample code
• Validation Report with security verification
• Live handoff session with connectivity test

‍

7 Security Controls Enforced

• SSE-KMS encryption
• HTTPS-only
• Least-privilege IAM
• CloudTrail logging S3 access logs
• Versioning enabled
• Blocked public access

‍

Want Similar Results for your business?

Schedule a call with the Cloudtech team to scope a voice-agent pilot for your use
case. We typically go from baseline to production in four weeks.

‍

‍

A family-owned travel company in Myrtle Beach built an outbound voice agent that qualifies leads naturally, then hands off to
human bookers , cutting cost-per-call 67% while keeping their U.S. based phone team at the center of every booking.

‍

Key results

‍

Customer Snapshot

Monster Reservations Group has booked vacations for thousands of families across 50+ destinations since 2006.

Their competitive edge has always been their U.S.-based phone team — fast, friendly, and the reason customers come back. The challenge was scaling this team without losing that quality.

‍

The Challenge

Slow responses broke the flow

The existing AI agent took 1.5 seconds to respond; long enough for callers to notice, interrupt, or lose

confidence. Conversations felt robotic, not human.

Manual outreach was inefficient

Agents spent time on initial calls gathering basic preferences before they could focus on actually booking

vacations. This limited how many customers they could serve.

The Bar for Success

Agents spent time on initial calls gathering basic preferences before they could focus on actually booking vacations. This limited how many customers they could serve.

‍

The Solution

CloudTech built an outbound AI voice agent designed for one job: qualify leads efficiently. Built on AWS using Amazon

Bedrock for reasoning, Amazon Connect for telephony, and ElevenLabs for natural voice synthesis.

Outbound calling

The AI agent initiates calls to customers and engages in natural conversation about their vacation preferences.

Preference gathering

Captures destination interests, travel dates, group size, and budget through guided dialogue.

Seamless handoff

Once preferences are gathered, the call transfers to a human agent via Amazon Connect with full context. No need for the customer to repeat anything.
‍

Scope & timeline

‍

Outcomes

• Projected 67% reduction in cost-per-call by automating initial outreach
• 500ms response latency, conversations feel natural, not robotic
• 95%+ accuracy in gathering customer preferences
• Human agents now start conversations with full context, ready to book
  ‍

What we tuned

Model selection

Amazon Bedrock for reasoning, ElevenLabs for natural voice

Conversation flow

Refined dialogue to gather preferences naturally

Handoff timing

Smooth transfer via Amazon Connect with full context
‍

Where it can be replicated

Outbound sales & lead qualification at scale
Teams that need to make hundreds or thousands of initial contact calls where the goal is to qualify, not close.

‍
Information gathering before human conversation
Businesses where every human conversation is more valuable when it starts with full context
healthcare intake, financial pre-qualification, service scheduling.

‍
Automated outreach with human closing
Any team that wants to automate the first 30 seconds of every call without losing the human touch on
the booking, sale, or decision moment.

Want Similar Results for your business?

Schedule a call with the Cloudtech team to scope a voice-agent pilot for your use case. We typically go from baseline to production in four weeks.

‍

Executive Summary

A leading provider of SaaS solutions for OTT video delivery and media app development partnered with us to modernize their backend infrastructure and enhance their analytics platform on AWS. The project aimed to transition core APIs, user analytics, and media streaming orchestration to a cloud-native, serverless architecture. This modernization significantly improved platform scalability, reduced latency for global users, and enabled real-time analytics across content, user interactions, and overall performance.

Challenges

The customer faced several challenges that hindered scalability, performance, and operational efficiency:

• API Modernization: Legacy, monolithic API services caused inefficiencies and limited scalability.
  
• Global Latency: Content delivery, particularly video, had inconsistent performance for users across different regions.
  
• Real-Time Analytics: A lack of real-time data insights made it difficult to track user engagement and optimize content.
  
• Operational Complexity: High operational overhead due to manual processes and limited automation.
  
• Disaster Recovery: Ensuring high availability and data redundancy was a top concern.

Scope of the Project

We were tasked with modernizing the SaaS platform by migrating core APIs and backend services to AWS microservices architecture. This involved microservice decomposition using Amazon API Gateway and AWS Lambda to replace monolithic API services. We also optimized video content delivery by utilizing CloudFront, S3, and MediaConvert for low-latency streaming and global delivery. A serverless analytics pipeline was built to process and analyze user events through Kinesis, Lambda, Redshift, and Glue. We ensured high availability by implementing multi-region failover with Route 53 and Aurora Global. Finally, CI/CD workflows were automated using CodePipeline, with monitoring and observability ensured through CloudWatch and X-Ray.

Partner Solution

We designed a fully serverless, scalable architecture using a combination of AWS Lambda for event-driven API services, CloudFront for global content delivery, and Redshift for analytics. The solution leverages Kinesis for real-time data ingestion, and Glue for data transformation and storage in S3.

The architecture includes:

• API Gateway for routing requests to Lambda and ECS/Fargate microservices.
  
• CloudFront for content delivery with S3 and MediaConvert integration.
  
• Redshift for querying and analyzing user interaction data.
  
• Aurora Global Database for cross-region failover and high availability.
  
• AWS Backup for disaster recovery and cross-region replication of data.
  
  

Our Solution

API Modernization

• Microservice Decomposition: We migrated monolithic APIs to a microservices-based architecture on AWS using API Gateway to manage routing and AWS Lambda to handle serverless execution.
  
• ECS/Fargate: Containerized components are managed through Amazon ECS (Fargate) for flexible, cost-efficient compute.
  
• API Gateway: Securely exposed the APIs to the frontend, validating requests and integrating with backend services using IAM roles for access control.
  

Streaming Optimization

• CloudFront CDN: Used CloudFront to cache content at the edge, reducing latency and speeding up content delivery globally.
  
• S3 & MediaConvert: Leveraged Amazon S3 for storage and MediaConvert for adaptive bitrate transcoding, enabling smooth video streaming on various devices.
  
• Global Distribution: Ensured optimal performance and reduced buffering by using CloudFront to serve video content efficiently to users across the globe.
  

Real-Time Analytics Pipeline

• Data Ingestion: Kinesis Data Streams was used to ingest user interaction events (e.g., play, pause, share) in real time.
  
• Data Enrichment: AWS Lambda processed and enriched the data streams before being stored in S3.
  
• ETL with Glue: AWS Glue performed ETL (Extract, Transform, Load) processes, converting data into an analytical format for consumption by Redshift.
  
• Analytics: Amazon Redshift was used for fast querying and reporting, enabling real-time insights into user behavior.
  

High Availability and Fault Tolerance

• Multi-AZ Deployment: Deployed critical services like Redshift and Aurora Global in multiple Availability Zones to ensure high availability.
  
• Route 53 Failover: Set up Route 53 with latency-based routing and health checks to ensure automatic failover between regions if one region faces issues.
  
• Auto-Scaling: Configured Auto Scaling Groups and ALB to automatically scale compute resources based on demand.
  

Benefits

• Enhanced Streaming Performance: By implementing CloudFront and MediaConvert, video buffering was reduced by >90% and latency minimized across regions.
  
• Real-Time Analytics: The Redshift + Glue pipeline enabled real-time analytics, empowering the customer to optimize user engagement based on live data insights.
  
• Operational Efficiency: Automating CI/CD with CodePipeline and using serverless components significantly reduced manual intervention, lowering operational overhead by 40%.
  
• High Availability: Route 53 routing and Aurora Global deployment ensured 99.95% uptime across regions, offering the customer peace of mind.
  
• Scalable Storage: Using S3 Intelligent-Tiering and Redshift Concurrency Scaling provided optimal storage management, ensuring cost efficiency as data grew.
  

Outcome (Business Impact)

• Enhanced User Satisfaction: Reduced video buffering to <3%, improving content delivery speed and user experience globally.
  
• Improved Match Accuracy: Real-time Redshift + Glue pipelines improved engagement tracking, reducing query times from 30s to <5s.
  
• Operational Efficiency: Automation through CI/CD and serverless components cut manual intervention by 40%, increasing overall team productivity.
  
• High Availability: Route 53 and Aurora Global ensured 99.95% uptime, even during regional outages.
• ‍Cost Efficiency: Optimized storage with S3 Intelligent-Tiering and Redshift Concurrency Scaling drove down operational costs.
  

Executive Summary

BeNotable is a platform dedicated to connecting music students with colleges. To stay ahead in a competitive landscape, BeNotable aimed to leverage Generative AI to enrich students’ audition experience and differentiate their service. We assessed their existing AWS-based data infrastructure (Amazon S3 and DynamoDB), technical maturity, and business objectives. The assessment highlighted an opportunity to introduce the “Aria Audition Lab Coach”, giving students instant, AI-generated feedback on tone, rhythm, and expressive quality. This case study outlines how we implemented a secure, scalable, and cost effective Generative AI workflow on AWS.

‍

Challenges

• Provide high quality AI feedback on large volumes of audio while maintaining low latency.
• Protect student data and intellectual property with robust security controls.
• Ensure end to end observability and graceful failure handling across asynchronous workloads.
• Integrate seamlessly with BeNotable’s existing AWS foundations without disrupting live users.

Scope of the project

• Discovery & Readiness : Assessed data quality, security posture, and AI objectives.
• Architecture & PoC :Designed an event driven, serverless architecture and validated model choice in Amazon Bedrock.
• Implementation : Built secure upload, processing pipeline, AI inference, and feedback delivery using API Gateway, Lambda, S3, DynamoDB, SQS/SNS, and EventBridge.
• UAT & Launch : Performance, security, and user acceptance testing with staged rollout.
• Enablement:Delivered IaC templates, runbooks, and a roadmap for multilingual expansion.

Partner Solution

• Cloud native platform - That matches music students with colleges via audition submissions.
• Web and chatbot interfaces - For students to upload recordings and receive feedback.
• Existing AWS foundations - Amazon S3 for raw audio and Amazon DynamoDB for metadata storage.
• Key business goals -  Deepen student engagement, enrich learning experience, and stand out from competitor platforms.
• ‍Secure Upload – Students authenticate with Amazon Cognito; requests are filtered through AWS WAF and served via Amazon API Gateway to a “PUT /upload audio” Lambda function.
• Storage Layer – Raw recordings land in an Amazon S3 bucket; Lambda captures metadata (student, instrument, timestamp) and writes to Amazon DynamoDB.
• Processing Pipeline – An SQS queue triggers a processor Lambda that transcribes audio and invokes Amazon Bedrock (Anthropic Claude or AI21) to generate feedback. Events are coordinated with Amazon EventBridge.
• Messaging Layer – Results are published through Amazon SNS. A Dead Letter Queue retains failed messages for replay and root cause analysis.
• Observability & Monitoring – Amazon CloudWatch Logs, metrics, and AWS X Ray traces provided full visibility, while AWS Config & IAM manage compliance and least privilege access.
• Scalability & Resilience – The design is serverless and fully managed, automatically scaling with usage and isolating faults through queue based decoupling.

‍

Solution Architecture Diagram

‍

‍

Metrics Used to Measure Success & Lessons Learned

• Engagement: +30 % increase in average session duration; 2× rise in audition uploads.
• Latency: p95 feedback delivery < 4 s.
• Reliability: < 0.2 % message failure, all captured in DLQ
• ‍Cost Efficiency: ~40 % reduction in operational overhead via serverless pay per use.

Lesson Learned

• Prompt engineering with few shot and chain of thought examples is key to nuanced music feedback.
• RAG with Titan Embeddings grounds generative output in music theory references for factual accuracy.
• Comprehensive observability accelerates latency tuning and error resolution.
• Early educator feedback loops refine model prompts and sustain content authenticity.

‍

Outcome (Business Impact)

• Students receive immediate, high quality feedback, increasing practice frequency and quality.
• Colleges gain richer audition insights, improving talent fit decisions and placement rates.
• BeNotable differentiates as an AI driven innovator, attracting new users and institutional partners.
• Serverless architecture scales elastically with peak audition seasons while aligning costs to usage.

‍

‍

‍