AWS Builder Center
From AI Builders to the Stage: Building Agentic Decision Intelligence on AWS

From AI Builders to the Stage: Building Agentic Decision Intelligence on AWS

How I went from the AI Builders program to co-presenting a live Agentic Decision Intelligence system at FCAJ Community Day — combining Markov Chains, MDP, Monte Carlo simulation, Bedrock AgentCore, and Amazon QuickSight to solve a real consumer-lending portfolio problem on AWS.

AI Engineer at Renova Cloud - AWS Premiere Partner - 8x AWS Certified - First Cloud AI Journey
I never planned to present at a community event. I just wanted to build something real.
How it started
A few months ago, I joined the AI Builders program — specifically the "Build with AWS" track. At the time, my goal was straightforward: get hands-on experience with Amazon Bedrock AgentCore and understand how to design production-ready agent architectures. I'd been working with AWS services for a while, but the agent paradigm felt different. Not just calling an API — orchestrating reasoning, tool use, and human approval into one coherent flow.
The program delivered exactly that. We went deep into spec-to-code workflows, agent patterns, and how to think about multi-step decision systems that actually run in production. But the biggest unlock wasn't purely technical.
The aha moment: learning a domain I knew nothing about
Through the program, I connected with Xuan Le — a finance professional working on consumer-lending portfolio optimization. She had a real problem: forecasting revenue for a 60,000+ billion VND loan portfolio, understanding why forecasts deviate, and deciding what action to take — fast, not at month-end.
Here's what surprised me: I had zero background in banking or debt collection. But as Xuan walked me through the domain — how loans migrate across delinquency buckets, how write-offs hit the balance sheet, how a 5% forecast error translates to a 50 billion VND gap — something clicked. The problem wasn't just math. It was a decision system. States, transitions, policies, uncertainty. And that maps perfectly onto an agent architecture.
That was my real aha moment. Not a technical breakthrough, but realizing that understanding the business process — deeply — is what turns a generic AI demo into something a C-suite would actually trust.
What we built
Together, we designed what we call "Agentic Decision Intelligence" (ADI) — a closed-loop system that combines quantitative models with AI agents and human judgment:
  1. Model — Markov Chains capture how loans transition between states (current, 1-30 DPD, 31-60 DPD, write-off, etc.)
  1. Decide — Markov Decision Process (MDP) optimizes which collection action to take at each state
  1. Simulate — Monte Carlo runs thousands of portfolio paths to stress-test the decision
  1. Interpret — An LLM agent (powered by Bedrock) translates the output into plain-language executive insight
  1. Act — A human reviews, approves, and the system executes
The architecture runs on AWS: Bedrock AgentCore handles the reasoning and orchestration layer, and Amazon QuickSight serves as the decision interface.
My part: the QuickSight demo
On stage at FCAJ Community Day, Xuan presented the full business framework — the domain logic, the math, the "why." My job was to make it tangible.
I built an Amazon QuickSight dashboard from raw loan data — showing portfolio composition, state transitions, waterfall movement, and Monte Carlo simulation results. Then I demonstrated QuickSight Q: asking natural-language questions against the dataset and getting instant answers. No SQL, no pivot tables — just a question and a chart.
The moment that landed hardest with the audience: when I typed a plain question about which loan segment drives the most write-off risk, and QuickSight returned a clear, filterable answer in seconds. That's when the room understood — this isn't a reporting tool. It's a decision interface.
What I took away
Three things stay with me:
First, the most valuable AI systems start with domain understanding, not model selection. If I hadn't spent weeks learning how debt collection actually works in Vietnamese consumer lending, the architecture would have been technically correct but practically useless.
Second, the "agent" in Agentic AI doesn't replace human judgment — it structures it. The entire pipeline is designed so that the model recommends, but a human owns the decision. That's what makes it deployable in regulated finance.
Third, AWS gives you the building blocks to go from idea to live demo remarkably fast. Bedrock AgentCore for orchestration, QuickSight for visualization and NLP querying, all connected. The infrastructure didn't slow us down — it got out of the way.
Thank you
Grateful to the FCAJ community for the stage, to the AI Builders program for the foundation, and most of all to Xuan Le for trusting a non-banker to bring her vision to life on a dashboard.
Knowledge only becomes ours when we build it. That Friday, we built it together.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article