Skip to main content
SYS_INIT // DWARKESH.LAB
DWARKESHRAMANI
Computer Engineering · Full-Stack Engineer
INITIALIZING DWARKESH // BUILD SYSTEM...12%
INDEX / RETURN TO SELECTED WORK
05/Lead Backend & AI Engineer·2026·SHIPPED

HackerRank Orchestrate

Visual evidence verification for automated damage claims.

Python 3.11Multimodal AIComputer VisionPydantic Schema ValidationPyTestAutomated Evaluation
VIEW SOURCE ON GITHUB

01 // OVERVIEW

Visual evidence verification for automated damage claims.

THE PROBLEM

Evaluating visual insurance and return claims manually is slow, inconsistent, and vulnerable to fraud. Ingesting multiple photos alongside conversational claims requires rigorous edge-case handling.

THE IDEA & APPROACH

An automated decision engine that ingests claim conversations, user history, and multiple submitted images, rigorously classifying evidence as Supporting, Contradicting, or Inconclusive.

02 // SYSTEM ARCHITECTURE & DATA FLOW

Pydantic-validated evidence pipeline executing multi-image inspection, claim alignment reasoning, and standardized JSON output validation.

FIG 1.0 // SYSTEM TOPOLOGY & EXECUTION PIPELINEINGESTIONFASTAPI / AIPERSISTENCE
DETAILED EXECUTION SEQUENCE
01 →Inbound claim payload with conversational logs, user metadata, and visual attachments
02 →Schema validator validates payload integrity against strict competition constraints
03 →Multimodal vision model evaluates visual damage specifics against claimed object type
04 →Cross-verification logic checks consistency with historical claims and conversation transcripts
05 →Decision output formatted according to strict benchmark schema with audit trail

03 // TECHNICAL DECISIONS & TRADE-OFFS

DECISIONStrict Pydantic models with automated test suite
ALTERNATIVES CONSIDEREDDecision Schema
WHY

Guaranteed zero schema mismatches during the automated 24-hour evaluation harness runs.

04 // CHALLENGES & RESOLUTIONS

CHALLENGE // 01

Distinguishing genuine damage from glare, dust, or intentional distortion in low-light camera captures.

CHALLENGE // 02

Maintaining execution speed within strict hackathon test runner timeouts.

05 // VERIFIED OUTCOMES

Completed under intense 24-hour hackathon constraints with high test pass rate.

Transparent audit logging capturing exact reasoning tokens for every claim decision.

06 // LESSONS & TAKEAWAYS

Rigorous type validation and deterministic schemas are the bedrock of reliable AI applications.
STATUS: VERIFIED ON GITHUB
INSPECT REPO