Case study: enterprise GenAI, Genpact
A GenAI copilot on Amazon Bedrock, from requirements to production
I led a GenAI ETL copilot on Amazon Bedrock from technical requirements to production and monitoring. It drafts and checks data-pipeline code, and 70+ projects moved onto it safely. Delivery became 3x faster, with 70% less manual engineering effort.
- Context
- Enterprise data platform for a large US telecom client, delivered through Genpact
- My role
- Lead Consultant. Led the technical side end to end: requirements, implementation, delivery and monitoring
- When
- Jan 2022 – Aug 2024
- Worked with
- Engineering, testing and operations leads across the US and India
- Stack
- Amazon Bedrock, AWS data services, Spark and SQL pipelines
The request
The client ran a large data platform. Every new or changed pipeline followed the same manual loop: an engineer turned the spec into code, ran it, then checked inputs and outputs by hand. Delivery was slow and code differed from project to project.
The request was to bring in a GenAI tool on Amazon Bedrock that drafts this code and checks it. The challenge was moving 70+ live projects onto it without breaking anything: many teams, production data and no room for surprises.
What the tool does
Context in
Specs, table structures and existing code for the data sources.
Draft
A foundation model on Amazon Bedrock drafts the pipeline code.
Check automatically
Scripts compare inputs to outputs and check run time against the old job.
Human approval
Engineer review and user acceptance testing before anything reaches production.
Key decisions
A validation gate to earn trust
Engineers didn't trust AI-generated code, so I defined success metrics and a human-in-the-loop gate: automated Hive and Athena data and runtime checks, then UAT.
One reusable structure
Standardising delivery on one structure let 70+ projects move onto the tool in planned phases.
How we rolled it out
Technical requirements
Turned the goal into technical requirements: inputs, checks, success criteria and how we would implement it.
Roadmap
Broke the move of 70+ projects into phases, with owners and dates.
Team alignment
Met leads from every affected team to agree responsibilities, timing and hand-offs.
Development
Built and tried the setup in development first.
Test and UAT
Tested with the teams and got sign-off on real use cases.
Production
Released with backups and the rollback plan ready.
Making it safe
Backups first
Full backups of every affected project before implementation, so nothing could be lost.
Rollback plan
A defined way to return to the previous pipelines quickly if a release failed.
Monitoring plan
What to watch after release, who watches it and what to do when something goes wrong.
Daily automated checks
Automation that checks the pipelines every day and flags issues so they get fixed fast.
Adoption and change management
Training the teams
Ran sessions so each team knew how to use the tool, read the checks and raise issues.
Proving it kept working
Daily automated monitoring showed results over time, which built confidence for wider use.
Results
These results come from the copilot I led, delivered together with the engineering, testing and operations teams. The client name, internal systems and commercial details are left out for confidentiality.
What I learned
With AI tools, going live is a product decision, not only a technical one. Teams adopted the tool because every output was checked, there was a clear way back if something broke, and daily monitoring showed it kept working. Next time I would also track how often the first draft passes its checks, to make the case for wider use even stronger.