I am an AI transformation and program leader with nine years at Amazon in marketplace trust and, before that, three years running a bank branch P&L. My work is turning VP-level ambiguity into shipped, governed systems.
Below are five records from the ledger, each with its baseline, its result and what I actually did. Every figure is the same one that appears on my resume and in my interview answers.
Case records
tick = baseline, bar = result
GenAI / Zero to One
Founding a GenAI and Agentic AI program from a blank charter
SHIPPED
I started with no team and no budget, only a charter I wrote myself. The program delivered RAG-grounded contact classification, automated risk scoring and workflow generation to 2,500 associates, and it was measured on adoption rather than deployment.
Adoption, weekly active use (two quarters)0%43.2%
Program NPS (15 months)-37.253, now 91 in yr 3
Model hallucinationbaselinedown 50%
Annualized savings-$9.9M
What I actually did
Wrote the VP-approved charter and carved engineering and data-science capacity from the existing org
Personally wrote and iterated the RAG-grounded prompts and conversation flows
Ran adoption as a product problem, with instrumentation, feedback loops and an intake pipeline that sorted ideas rather than culling them: 150 ideas raised, 14 productized in wave one, 12 carried into wave two, 124 industrialized through other channels and none cancelled
Boundary I did not write production code and did not own model infrastructure. I owned the charter, the workflow design, the evaluation and the number.
Executive Judgment / Cost
Reversing a $22.5M cost mandate into a $30M roadmap
SHIPPED
A VP directive to cut $22.5M in opex would have eliminated 235 roles and the growth programs with them. I challenged the brief instead of executing it.
Savings plan$22.5M cut$30M roadmap
Net headcount cut235 roleszero
Growth programs protectedat risk58
What I actually did
Built the financial case independently over a three-month cross-functional review, covering scenarios, sensitivities and trade-offs
Presented the alternative to the VP and won endorsement on data, not narrative
Delivered part of the roadmap through the GenAI program above, so the two records compound
Boundary I did not hold budget authority. I changed the decision by producing better evidence than the brief rested on.
Governance / Org Building
Founding the organization's only Program Management Office
SHIPPED
Forty-five programs, two hundred stakeholders and fifty teams were running with no shared intake, stage-gates or risk discipline. I built the operating system first and then the intelligence layer on top of it.
On-time delivery60%90%
Portfolio under governancepartial100%
Plan-vs-actual forecast variance36%7%
Resourcing ask, one cycle66 people12
What I actually did
Designed intake, stage-gate governance, RAID, and decision registers from a blank page
Built the RAG-status portfolio reviews and QBR cadence feeding annual planning and FP&A
Moved leadership from slide-driven status to system-driven visibility
Boundary I founded and ran the PMO without a dedicated PMO headcount; it was built inside an existing team.
Convergence / Delivery at Scale
Converging two global organizations onto one operating model
SHIPPED
This was a first-of-its-kind convergence with no playbook, a hard six-month deadline and peak sale-event blackout windows in the middle of it. Two organizations, whose work spanned 27 marketplaces and 10M+ annual contacts, moved onto a single operating model.
Authored the Director-approved cutover playbook (risk-ranked wave sequencing, dual-run staffing, blackout windows, daily defect triage) and framed the go/no-go on program health
Arbitrated cross-org trade-offs between two global organizations on data-backed evidence
Held SLA, quality, and speed through blackout windows when volume spikes and launches freeze
Boundary I did not own either organization. I owned the mechanism that moved them, and the go/no-go call went to a Director.
Hands-On AI / Claude + Gemini
Building the AI tools myself, with no engineering team
SHIPPED
I personally designed, prompted and shipped a family of Claude- and Gemini-powered systems: a support chatbot validated in a live bake-off, real-time risk-context agents and planning assistants.
Information turnaroundbaselinedown 90%
Investigator context assembly72 hours2 hours
Handle time (12 months on)17.5 min2.5 min
Capacity freed-120,000 FTE hours
What I actually did
Fused live platform risk signals with historical behavior data into agent-assembled case context
Validated the chatbot against the incumbent tool in a live seller-facing bake-off
Published the personal-project designs as open source, where I designed them and an AI assistant wrote the code: github.com/prateek-ratnakar
Boundary I work at the design, requirements and evaluation layer, and I do not write or review application code. I designed, prompted and tested these tools, and I tracked system integration points at diagram level, not interface-spec level.
Also on the ledger
GenAI program14 systems in production, with 12 more in a second wave
Intelligent Priority Routing10M+ contacts a year, executive escalations down 90%
International launchesA zero-defect Ireland launch, which stood up its contact center as a service, and a move from two country launches a year to five
Before Amazon
I was a Branch Manager at Bank of India with a full P&L and credit-sanctioning authority. Deposits grew 26.95%, credit grew 21.02%, non-performing assets fell 71.4%, and the branch ranked in the top 1% nationally for cross-sell. I learned risk and recovery from the inside, not from a deck.
I hold a B.Tech in Chemical Engineering from IIT (BHU), with First Class Honours, and an MBA in Strategy, Operations and Finance from IIM Bangalore. I am a two-time Amazon North Star Award winner (2021, 2025) and was Business Leader of the Year (2019), and I hold the IIT-BHU COLOR (2011) and BLUE Honorable Mention (2012).
The next record is open.
I am looking for Staff / Principal TPM, Director of Program Management, Chief of Staff, Founder's Office and AI transformation mandates. Hand me the ambiguous, unowned problem.