Neuro-Symbolic Insurance Evaluator

A highly optimized, multi-modal pipeline for verifying damage claims via visual evidence. Built for the HackerRank Orchestrate 24-hour hackathon (June 2026).

Out of 15,295 registered participants, 1,773 successfully shipped a runnable AI agent and passed the final AI Judge interview. I finished #71 globally (Top 4%), placing 2nd in Germany and 6th in Europe! 🇩🇪🇪🇺

This system evaluates claims across cars, laptops, and packages by processing chat transcripts, user history, and image sets. It employs a Neuro-Symbolic architecture: using a Vision-Language Model (Claude 4.6 Sonnet) for visual reasoning, followed by a deterministic Python layer (engine.py) to enforce strict business logic and safety guarantees against "Critic Bias."

• Architecture: Unified multi-modal VLM call → Deterministic Python Guardrails.
• Performance: High concurrency via asyncio.Semaphore, zero-cost dev iterations via SQLite cache.py.

Read problem_statement.md for the full task spec, or check out code/README.md for my implementation details.

---

Contents

1. Repository layout
2. What you need to build
3. Where your code goes
4. Quickstart
5. Evaluation
6. Chat transcript logging
7. Submission
8. Judge interview

---

Repository layout

.
├── AGENTS.md                         # Rules for AI coding tools + transcript logging
├── problem_statement.md              # Full task description and I/O schema
├── README.md                         # You are here
├── code/                             # Build your solution here
│   ├── main.py                       # Suggested terminal entry point
│   └── evaluation/
│       └── main.py                   # Suggested evaluation entry point
└── dataset/
    ├── sample_claims.csv             # Inputs + expected outputs for development
    ├── claims.csv                    # Inputs only; run your system on these rows
    ├── user_history.csv              # Historical claim counts and risk context
    ├── evidence_requirements.csv     # Minimum image evidence requirements
    └── images/
        ├── sample/                   # Images referenced by sample_claims.csv
        └── test/                     # Images referenced by claims.csv

---

What you need to build

A system that, for each row in dataset/claims.csv, produces one row in output.csv.

Input fields:

Column	Meaning
user_id	User submitting the claim; use this to look up dataset/user_history.csv
image_paths	One or more submitted image paths, separated by semicolons
user_claim	Chat transcript describing the issue
claim_object	car, laptop, or package

Required output fields:

Column	Meaning
evidence_standard_met	Whether the image set is sufficient to evaluate the claim
evidence_standard_met_reason	Short reason for the evidence decision
risk_flags	Semicolon-separated risk flags, or none
issue_type	Visible issue type
object_part	Relevant object part
claim_status	supported, contradicted, or not_enough_information
claim_status_justification	Concise explanation grounded in the image evidence
supporting_image_ids	Image IDs supporting the decision, or none
valid_image	Whether the image set is usable for automated review
severity	none, low, medium, high, or unknown

Hard requirements:

• Must read the provided CSV files and local images.
• Must produce output.csv with the exact schema in problem_statement.md.
• Must include an evaluation workflow
• Must avoid hardcoded test labels or file-specific answers.

Beyond that you are free to bring your own approach: VLMs, LLMs, structured prompting, rule layers, batching, caching, evaluation pipelines, model comparison, or anything else.

---

Where your code goes

All of your work belongs in code/. The repo ships with empty starter files that you can grow into your full solution.

Suggested conventions:

• Put your main runnable solution in code/main.py, or document your own entry point clearly.
• Put evaluation code under code/evaluation/ or an evaluation/ folder included in your final code.zip.
• Write final predictions to output.csv.

---

Quickstart

Clone this repository:

git clone git@github.com:interviewstreet/hackerrank-orchestrate-june26.git
cd hackerrank-orchestrate-june26

You are free to use any language or runtime. Python, JavaScript, and TypeScript are all reasonable choices.

---

Evaluation

The evaluation report should include:

• metrics on dataset/sample_claims.csv
• at least two strategies, prompts, or model configurations compared
• the final strategy used for output.csv
• operational analysis covering model calls, token usage, image usage, approximate cost, runtime, and TPM/RPM considerations

---

Chat transcript logging

This repo ships with an AGENTS.md that modern AI coding tools may read. It instructs the tool to append conversation turns to a shared log file:

Platform	Path
macOS / Linux	$HOME/hackerrank_orchestrate/log.txt
Windows	%USERPROFILE%\hackerrank_orchestrate\log.txt

You will upload this log as your chat transcript at submission time. The chat transcript means your conversation with the AI coding tool you used to build the system. It is not the runtime logs, reasoning trace, or conversation history produced by the claim-verification agent you are building.

If you use multiple AI tools, include the relevant conversation logs from all of them in the same transcript file. Separate each tool's section with a clear divider and label it with the tool name.

Never paste secrets into the chat. If secrets are needed, use environment variables.

---

Submission

Submit the following files as instructed by HackerRank:

1. Code zip: zip your runnable solution, README, prompts/configs, and evaluation folder. Exclude virtualenvs, node_modules, build artifacts, and unnecessary generated files.
2. Predictions CSV: your final output.csv for all rows in dataset/claims.csv.
3. Chat transcript: the log.txt from the path in Chat transcript logging.

Before submitting, confirm:

• output.csv has one row per row in dataset/claims.csv.
• output.csv has the exact required columns in the exact required order.
• Your evaluation files are included in code.zip.

---

Judge interview

After submission, the AI Judge may ask about your approach, implementation decisions, model usage, evaluation strategy, and how you used AI while building the solution.

Be prepared to explain your solution in detail.