NM NitinMane
AURANGABAD, IN Let's Talk
Home / Projects / AWS Agent Engineer
Credential Spotlight — AWS AI & ML Scholars

Future AWS Agent Engineer

A fully funded Udacity Nanodegree earned through the AWS AI & ML Scholars program and completed on October 8, 2026. Three agent systems on Amazon Bedrock AgentCore, each one harder than the last: a guarded chatbot, a support agent with six tools, and a team of five agents that answers questions from three knowledge bases.

Verify Certificate Explore the Three Builds Read the Journey
Verified by Udacity

Verified Certificate of Nanodegree Program Completion

Future AWS Agent Engineer

One of three tracks in Phase 2 of AWS AI & ML Scholars, a program run by AWS and Udacity. Up to 100,000 learners started the free Challenge phase, and the top 4,500 on the assessment were given a fully funded Nanodegree. The track covers Amazon Bedrock, Amazon Nova, retrieval-augmented generation (RAG), and single- and multi-agent systems.

Recipient
Nitin Mane
Completed
October 8, 2026
Issuer
Udacity
Pathway
AWS AI & ML Scholars
Certificate ID
ba1d41ec-900d-11f1-9972-03d0f79bde06
View certificate on Udacity About the scholarship
03 Agent systems built From one routed chatbot to five cooperating agents
120/120 Official test score Multi-agent RAG assignment, Udacity AWS sandbox
0.95 Bedrock correctness LLM-as-a-judge score, harmfulness 0.00
6/6 Live scenarios passed Orders, refunds, RAG, memory, maths, browser
34 Automated tests 22 core + 12 UI tests, all passing
0 AWS resources left running Teardown checked after every project
01 — The Pathway

From scholar to agent engineer.

The scholarship has two phases: a free, self-paced Challenge, and then a Nanodegree for the learners who score highest on the assessment. I built one project each month.

  1. Spring 2026

    Challenge phase

    Self-paced AWS AI Practitioner learning plan, plus hands-on work in Amazon PartyRock and Amazon Bedrock.

  2. July 2026

    Assessment & selection

    An assessment to rank learners. The top 4,500 receive a fully funded Udacity Nanodegree.

  3. August 2026

    Project 01 — Chatbot

    A three-route support chatbot on an AgentCore Harness. It saves bug reports to DynamoDB through a Gateway tool.

  4. September 2026

    Project 02 — Support Agent

    A Strands agent on AgentCore Runtime with six capabilities: Gateway tools, a Knowledge Base, Memory, Code Interpreter and Browser.

  5. October 8, 2026

    Project 03 & graduation

    NovaMart, a multi-agent RAG system that scored 120/120 on the official tests, and the Nanodegree completed the same day.

02 — Three Builds

One agent, then six tools, then five agents.

Each project made the system harder to keep under control. Each time, the useful work was deciding which parts the model handles and which parts plain code enforces.

Agentic AI August 2026

Customer Support Chatbot

A support chatbot with three separate routes, where plain code decides when the database may be written to.

The problem

Support bots fail in two common ways: they make up answers, and they save incomplete tickets. This one has three routes that never overlap. It collects bug reports over several turns, answers only from the shop's FAQ, and sends everything else to a person.

What I built

  • An AgentCore Harness on Amazon Nova Pro 1.0 that keeps conversation state across turns.
  • An AgentCore Gateway that gives the model one tool, defined by an OpenAPI contract: a Lambda function that writes bug reports to DynamoDB.
  • A routing guard in plain code that blocks early writes, removes <thinking> tags from replies and handles fallbacks.
  • A Flask web app with login and a panel that shows which route was chosen and the source of each answer.
  • Bedrock evaluation on a JSONL dataset, with a model scoring the answers and a per-case metrics table.
0.95Correctness
0.00Harmfulness
34/34Tests passing
10/10Eval cases valid
Bedrock AgentCore Amazon Nova Pro AWS Lambda DynamoDB CloudFormation Flask pytest
View Repository Read the Build Notes
Agentic AI September 2026

AI Support Agent

One conversation can track an order, issue a refund, look up a policy, remember the customer, work out a discount and read a website.

The problem

Real support requests need several backend services. An agent with many tools is only useful if every tool has a clear job, and if the final reply keeps the numbers the tools returned.

What I built

  • An async Strands app on AgentCore Runtime using Amazon Nova 2 Lite.
  • An MCP Gateway with two targets: an API Gateway endpoint for orders and a Lambda function for refunds.
  • Bedrock Knowledge Base search over a product catalog in S3, indexed in OpenSearch Serverless.
  • AgentCore Memory hooks that bring back customer preferences in later sessions.
  • Code Interpreter for loyalty discounts (Python Decimal, exact rounding) and Browser for reading live web pages.
  • A check on every reply that copies the calculator's exact figures, so the model cannot restate them wrongly.
6/6Live scenarios
$99.00Exact discount total
$139.99Refund approved
2Sessions, one memory
Strands Agents AgentCore Runtime MCP Gateway Bedrock Knowledge Bases OpenSearch Serverless Code Interpreter Python 3.13
View Repository Read the Build Notes
Agentic AI Capstone October 2026

NovaMart Multi-Agent E-commerce RAG

A supervisor agent sends each question to four specialist agents, which share state and search three knowledge bases at once.

The problem

An e-commerce support question often needs order data, a refund decision and policy text at the same time. A single agent ends up mixing these jobs together. NovaMart gives each job to its own agent and keeps all of their work in one shared, versioned record.

What I built

  • An OrchestratorAgent (Claude Haiku 4.5) with five routing tools, and four Claude Sonnet 4.5 agents for inventory, policy, refunds and the final reply.
  • Parallel RAG: the PolicyAgent searches the Returns, Shipping and Warranty knowledge bases (stored in S3 Vectors) at the same time using a ThreadPoolExecutor.
  • A shared WorkflowState table in DynamoDB with optimistic locking: each write must match the expected version or it retries.
  • AgentCore Memory for seven-day session summaries and a Bedrock Guardrail on the runtime.
  • CloudWatch and X-Ray tracing, CloudFormation and CDK for infrastructure, and GitHub Actions for validation and deployment.
120/120Official tests
5Cooperating agents
3Knowledge bases in parallel
0Failed trace nodes
Strands Agents AgentCore Runtime & Memory S3 Vectors DynamoDB X-Ray AWS CDK GitHub Actions
View Repository Read the Build Notes
03 — Side by Side

How the design changed.

The same questions came up in every project: where state lives, where answers come from, and how to prove the system works.

01 · Chatbot
02 · Support Agent
03 · NovaMart
Pattern
One agent, three fixed routes
One agent, six tools
Supervisor and four specialist agents
Model
Amazon Nova Pro 1.0
Amazon Nova 2 Lite
Claude Haiku 4.5 + Sonnet 4.5
Knowledge
FAQ included in the prompt
Knowledge Base on OpenSearch Serverless
3 Knowledge Bases on S3 Vectors, searched in parallel
State
Harness session
AgentCore Memory across sessions
Versioned DynamoDB state + Memory
Safety
Routing guard in code
Reply check against tool output
Bedrock Guardrail + routing rules
Proof
Bedrock LLM-as-a-judge, 34 tests
6 recorded live scenarios
120/120 official tests, X-Ray trace map
04 — Skills

What I can now build.

Multi-agent orchestration

Supervisor and worker design, routing rules, specialist agents and a fixed final step.

Strands AgentsAgent graphs

Retrieval (RAG)

Bedrock Knowledge Bases on S3 Vectors and OpenSearch Serverless, parallel search, and answers that cite their sources.

Knowledge BasesS3 Vectors

Agent tools

AgentCore Gateway (MCP and OpenAPI), Lambda and API Gateway targets, Code Interpreter and Browser.

MCPOpenAPI

State and memory

Session state, memory across sessions, and versioned DynamoDB writes so agents don't overwrite each other.

AgentCore MemoryDynamoDB

Safety and evaluation

Guardrails, guards written in code, prompt-injection defences, model-graded evaluation and tests run against the live system.

GuardrailsBedrock Evaluation

Operations

CloudWatch logs, X-Ray traces, CloudFormation and CDK, GitHub Actions, and a checked teardown so nothing keeps costing money.

X-RayAWS CDK
Let's Build

Building agents on AWS?

I'm interested in agent systems that touch real operations: support, plant floors and edge devices. I can help design the architecture, the guardrails and the evaluation.

Start a Conversation See All Projects