
First Commit AWS: Building Sentinel AML with Agentic AI
How we built Sentinel AML at First Commit using agentic AI, explainable risk, human review, and AWS cloud architecture for AML investigations.
Description
At First Commit — Bharat Builds Tour × AWS Builder Center, we built Sentinel AML, an agentic investigation system for suspicious financial activity. The project taught us that good AI is not just about prediction — it is about evidence, explainability, and knowing where humans must stay in control.
Some hackathons are about shipping fast.
First Commit made us think harder about what we were shipping.
On September 19, we attended the in-person build day of First Commit — Bharat Builds Tour × AWS Builder Center at Polaris School of Technology, Bengaluru.
The atmosphere was one of the best parts of the experience.
There were builders everywhere, AWS mentors helping teams unblock technical problems, people openly sharing ideas, and teams giving each other feedback instead of treating the room like a closed competition. The event itself was designed around workshops, feedback, and actual build time, which made the environment feel much more collaborative than simply "submit and wait."
Having Kunal Kushwaha and Hitesh Choudhary there also made the day memorable. The agenda included opening remarks from Kunal and a discussion with Hitesh and Kunal around skills, education, and what comes next for students in technology.
For our team — Devesh Raj and Mayank Gupta — the day became less about adding one more feature and more about asking better questions.
The idea
We started with a simple problem:
Can AI help investigators identify suspicious financial activity without turning the decision into a black box?
That became Sentinel AML.
Our first instinct was to build a familiar ML pipeline:
Transaction Data ↓Model ↓Suspicious / Not Suspicious
But very quickly we realized that this was not enough.
If a system flags an account, an investigator will naturally ask:
- Why was it flagged?
- Which transactions caused the concern?
- Was it a known AML pattern?
- Was it simply an anomaly?
- What evidence supports the score?
- What should happen next?
That changed the entire architecture.
Building an investigation system, not just a prediction model
Sentinel allows an analyst to ask questions such as:
Find structuring activity for account 803D95360
or:
Which customers made 10 or more transactions below $10,000?
or:
Detect circular movement of funds
The important part is that the system does not run the same pipeline every time.
It first understands the analyst's intent and then builds a query-specific execution plan.
For a direct threshold question:
Data Loader ↓Aggregation ↓Result
There is no reason to run machine learning simply because a model exists.
For a broader investigation:
Data Loader ↓Feature Engineering ↓AML Rules ↓Statistical Detection ↓Isolation Forest ↓Risk Scoring
And for network behavior:
Transactions ↓Graph Analysis ↓Cycles / Fan-in / Fan-out ↓Risk
That adaptive planning became one of the core ideas behind Sentinel.
Why we use multiple detection methods
There is no single algorithm that can explain every suspicious financial pattern.
So Sentinel combines:
Rule-based detection for known patterns such as structuring and rapid movement.
Statistical analysis using techniques such as Z-score and IQR to detect behavior far outside an account's normal baseline.
Isolation Forest for multivariate anomalies that are difficult to express with a fixed rule.
Graph analysis for transaction relationships such as circular movement, fan-in, fan-out, and concentration.
The outputs are then combined into an explainable risk score.
The goal is not to say:
"The AI thinks this account is suspicious."
The goal is to say:
"This account was escalated because these specific signals were observed, these detectors contributed to the score, and these transactions form the supporting evidence."
That difference matters.
Where AI fits — and where it should stop
One of our biggest learnings from the build was that generative AI should not be responsible for deciding guilt or innocence.
In Sentinel, analytical components calculate the evidence first.
The AI layer helps with:
- understanding the analyst's question
- planning the investigation
- summarizing evidence
- producing a readable explanation
The final decision still belongs to a human investigator.
Our end-to-end philosophy became:
Natural Language ↓Planning ↓Data Analysis ↓Suspicious Pattern Detection ↓Risk ↓Explanation ↓Human Review
Thinking in AWS services
The AWS workshops also pushed us to think beyond local code and ask:
How would this become a real cloud workflow?
Instead of adding AWS services just for the architecture diagram, we mapped each service to a clear responsibility.
Amazon S3 for transaction datasets, evidence files, and generated reports.
Amazon DynamoDB for investigation metadata, risk scores, status, and workflow state.
AWS Lambda for event-driven actions such as starting ingestion after a dataset arrives.
Amazon API Gateway as the controlled API entry point between the frontend and backend.
Amazon SNS for notifying analysts when a high-risk investigation is created.
Amazon CloudWatch for logs, failures, processing events, and operational visibility.
Conceptually:
Analyst ↓React Workspace ↓API Gateway ↓AML Backend ↓Dynamic Planner ↓Rules + Statistics + ML + Graph ↓Risk Score ↓DynamoDB ↓High Risk? ↓SNS Notification
S3 → datasets and reportsLambda → event-driven processingCloudWatch → monitoring
The most useful lesson was that cloud architecture is not about using the maximum number of services.
It is about giving every service a clear job.
What First Commit changed for us
We entered the hackathon thinking mostly about detection accuracy.
We left thinking much more about:
- explainability
- system boundaries
- responsible AI
- failure handling
- human review
- cloud architecture
- building something another person could actually trust
The environment helped with that.
Being surrounded by teams actively building, getting feedback from mentors during the day, and seeing people openly discuss architecture and trade-offs made the event feel much more like a builder community than a traditional competition. The presence of the AWS team and the hands-on AI sessions made it easier to move from "we have an idea" to "how would this actually run as a system?"
And listening to people like Kunal Kushwaha and Hitesh Choudhary in the same room reinforced something simple: learning technology is useful, but building something real with it is where the gaps in your understanding become visible.
The biggest takeaway
We started with this question:
Can AI detect suspicious transactions?
By the end, the better question was:
Can AI help an investigator understand suspicious activity without hiding how the conclusion was reached?
That is what we are trying to build with Sentinel AML.
Not an autonomous compliance officer.
Not a black-box risk score.
But a system where every important conclusion can be traced back to evidence — and where a human remains responsible for the final decision.
Team: Devesh Raj and Mayank Gupta
Project: Sentinel AML
GitHub:
Sentinel AML is a demonstration decision-support system. A risk score is an investigative lead, not proof of criminal activity.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article