Policy PDF + rater corpus
Accept a local policy-document path and the supplied
raters/ directory. Your implementation may support
uploads or URLs as an extra.
Engineering × Insurance
Teach an agent to read the fine print.
Build an explainable insurance brokerage agent that turns messy policy PDFs and inconsistent insurer grids into one defensible, source-cited answer.
The premise
The rate is never in just one place.
Indian insurer commission grids arrive as spreadsheets, PDFs, circulars and images. Each one invents its own vocabulary for geography, vehicle segments, policy types and exceptions.
Your job is to build the missing reasoning layer: extract the policy facts, traverse the right grid, refuse unsupported cases, and show exactly how the answer was reached.
RTO / zone
make / model
fuel / CC
vehicle age
policy type
NCB
premium slab
Problem statement
Build the brokerage agent.
Given a motor policy PDF and a directory of insurer rate cards, return the expected brokerage with an evidence trail another engineer—or broker—can audit.
Rate + exact provenance
Return extracted facts, OD and TP rates, granular source citations, confidence, and a plain-language explanation of every lookup step.
Extract what changes the answer
Insurer, make/model, fuel, RTO, CC, manufacture year, policy type, NCB and premium breakup. Preserve uncertainty instead of guessing through it.
Keep lookup logic deterministic
LLMs are welcome for document extraction. Grid resolution must be testable, reproducible and explainable—not a black-box answer.
Cite the precise source
Name the file and the exact sheet, row, cell or PDF page used. “Found in the rate card” is not enough.
Know when not to answer
If a segment is declined, ambiguous or absent, surface the gap and say what evidence would resolve it.
Four policies. Four lookup grammars.
Your test bench
-
01
PDF · included
HDFC ERGO · Comprehensive
Zone × premium slab × fuel × NCB
-
02
PDF · included
Go Digit · Stand-alone TP
RTO cluster × fuel / CC segment
-
03
PDF · included
Reliance · Comprehensive
RTO zone × model segment × footnote
-
04
PDF · included
Tata AIG · Stand-alone TP
Coverage gap detection—not just lookup
Every supplied policy is a metadata-clean synthetic fixture. Names, contacts and customer-specific identifiers have been replaced; city/RTO context, vehicle rating facts and premium figures remain intact. The bundle includes a public verification summary and checksums.
The corpus
Messy by design. Realistic by default.
Use the original source formats. Normalising them is part of the challenge; hard-coding the four answers is not.
4 × PDFDe-identified test policiesFor example: HDFC ERGO comprehensive and Go Digit SATP
3 × XLSXOriginal-format brokerage gridsFor example: Go Digit multi-sheet and Tata AIG 70+
columns
1 × PDFOriginal-format rate gridHDFC ERGO · two pages with zone and slab tables
DocsBrief, verification and manifestTask contract, privacy summary and SHA-256 checksums
What good looks like
Correct is the start. Defensible is the bar.
Accuracy
Extract the right facts and reach the right rate—or the right refusal.
Traceability
Make every transformation and source lookup inspectable.
Resilience
Handle missing fields, exclusions, footnotes and unsupported segments.
Engineering
Readable code, meaningful tests, sensible boundaries and a useful README.
“The most impressive submission may be the one that confidently says I cannot determine this rate from the supplied evidence—and proves why.”
No deadline. No permission slip.
Build it because the problem bothers you.
Publish your implementation in a repository with setup instructions, tests, sample outputs and a short note on the trade-offs you made. When it is ready, email the repository link and your four decision traces to careers@vaatun.com.
□ source code
□ run instructions
□ tests
□ four decision traces
□ assumptions