BaakiBatao - tell me the rest
Your insurer paid less than you claimed. Was that lawful?
That question became BaakiBatao, built by Kartik Chavan, Eshan Nahar and
Avadhut Noola for First Commit, Bharat Builds × AWS.
The problem is explanation, not rejection
Policies have room rent limits, deductibles, co-payments, sub-limits and non-payable items.
A lower settlement is not proof of wrongdoing.
So we did not build a tool that labels deductions suspicious. We built one that reconstructs
the settlement independently and then compares.
Every deduction lands in one of three buckets:
Bucket | What it means |
|---|
Correctly applied | We can reproduce it from the policy |
Incorrectly applied | It conflicts with a clause we can cite |
Unresolved | We do not have enough to make a defensible call |
That last one matters. If the documents do not explain something, we say so.
Reconstruction, not detection
Deductions happen in sequence. A room rent cap affects related charges. A deductible changes
the payable base. Co-pay applies to that result.
In our worked example the total deduction is ₹2,06,680:
| Amount | Why |
|---|
Correctly applied | ₹84,480 | Reproduced from the policy |
Incorrectly applied | ₹1,19,200 | Conflicts with rules we can cite |
Unresolved | ₹3,000 | Unexplained |
Total | ₹2,06,680 |
|
But the answer is not "ask for ₹1,19,200 back." Restore those deductions and the policy's
10 percent co-pay still applies:
₹1,19,200 disputed, with a clause behind every rupee
− ₹11,920 the 10% co-pay that lawfully rides on a restored deduction
────────────
₹1,07,280 actually challengeable, with citations
That gap is the entire argument for running the full calculation.
The arithmetic has to close
If the difference is ₹2,06,680, our findings sum to ₹2,06,680. Exactly.
- Money is stored in integer paise, never floats.
- Every finding carries a clause behind it, or it is not a finding.
The engine runs seven ordered stages:
Order matters, so it lives in a versioned rulepack as data rather than scattered through
application code. Same input, same rulepack, same output.
Where AWS fits
Service | What it does here |
|---|
Amazon S3 | Documents land through presigned uploads |
Amazon Textract | Extracts structure from bills, deduction sheets and policies |
AWS Step Functions | Orchestrates the workflow, including stages that pause for human correction |
AWS Lambda | Runs each stage |
Amazon DynamoDB | Holds case state and events |
AWS KMS | Signs the final certificate |
Amazon Comprehend | The redaction gate, plus deterministic checks |
Amazon Bedrock | Turns already redacted findings into readable prose |
Before text moves downstream it passes the redaction gate, keeping raw documents separated
from later stages.
Bedrock sits in the pipeline but never decides the money. The number comes from rules
and arithmetic.
Failing honestly
OCR is noisy and bill descriptions vary. If a line cannot be normalised confidently, the
workflow stops and asks instead of attaching the wrong rule to the wrong charge.
We also tested that lawful deductions are left alone. A system that flags everything demos
well and helps nobody.
The standard we kept returning to was never "can we detect a bad deduction." It was:
Can someone else verify why we reached this result?
BaakiBatao. Tell me the rest.
Loading article