HackerRank Orchestrate

HackerRank Orchestrate 24-hour hackathon (May 1–2, 2026).

Goal : Build a terminal-based AI agent that triages real support tickets across three product ecosystems; HackerRank, Claude, and Visa — using only the support corpus shipped in this repo.

Problem Statement : problem_statement.md

🤖 HackerRank Orchestrate AI Agent

An intelligent terminal-based AI support agent that triages and responds to real-world support tickets across:

• 🧑‍💻 HackerRank
• 🤖 Claude
• 💳 Visa

Built in a 24-hour hackathon, this system combines RAG (Retrieval-Augmented Generation), LLM reasoning, and safety guardrails to produce accurate, grounded, and reliable responses.

---

🎯 Problem

Support tickets are:

• noisy
• ambiguous
• sometimes risky (fraud, billing, account access)

The challenge:

Convert messy user queries into structured decisions + safe responses using ONLY the provided support corpus.

---

🧠 Solution Overview

We designed a hybrid AI system that combines:

• 🔍 Retrieval (RAG) → fetch relevant support docs
• 🤖 LLM reasoning → understand & generate responses
• ⚖️ Rule-based safety layer → enforce escalation for high-risk cases

---

⚙️ System Architecture

Checkout : README.md

User Ticket
   ↓
Preprocessing (subject + issue)
   ↓
Hybrid Retrieval (Dense + Sparse)
   ↓
LLM Reasoning (classification + response)
   ↓
Safety Layer (risk override)
   ↓
Structured Output (CSV)

---

⚡ Demo

Input Ticket:

"My Visa card was charged twice for the same transaction"

Output:

• Status: escalated
• Product Area: billing
• Request Type: product_issue

Response: Duplicate charges may indicate a billing issue or transaction error. Please contact your issuing bank to initiate a chargeback...

Justification: High financial risk detected → escalation required

---

🔍 Key Features

✅ Hybrid Retrieval (RAG)

• Dense (embeddings) + Sparse (TF-IDF)
• Improves relevance and reduces hallucination

✅ Structured AI Reasoning

• Outputs strict JSON:

  • status
  • product_area
  • response
  • justification
  • request_type

✅ Safety-First Design

• Fraud / billing / security → auto escalation
• Prevents unsafe or incorrect responses

✅ Cost Optimization

• Precomputed embeddings (cached locally)
• Reduced token usage drastically

✅ Deterministic Overrides

• Combines LLM intelligence with rule-based control

---

📂 Repository Structure

.
├── AGENTS.md                       # Rules for AI coding tools + transcript logging
├── problem_statement.md            # Full task description and I/O schema
├── README.md                       
├── code/                           
│   └── main.py                     #   Entry point
│   └── build_index.py
│   └── classifier.py
│   └── llm.py
│   └── processor.py
│   └── retriever.py
│   └── risk.py
│   └── temp.py
│   └── utils.py
├── data/                           # Local-only support corpus (no network needed)
│   ├── hackerrank/                 #   HackerRank help center
│   ├── claude/                     #   Claude Help Center export
│   └── visa/                       #   Visa consumer + small-business support
│   └── embeddings.npy
└── support_tickets/
    ├── sample_support_tickets.csv  # Inputs + expected outputs (for development)
    ├── support_tickets.csv         # Inputs only (run your agent on these)
    └── output.csv                  # RESULTs

---

🧾 Output Format

Each ticket produces:

Column	Description
status	replied / escalated
product_area	domain classification
response	grounded user answer
justification	reasoning behind decision
request_type	issue classification

---

🚀 Quickstart

git clone https://github.com/Kshitijasharma/hackerrank-orchestrate-ai-agent.git
cd hackerrank-orchestrate-ai-agent/code
python main.py

---

🔒 Constraints

• Uses ONLY local support corpus (no external knowledge)
• Avoids hallucinations
• Escalates high-risk cases safely

---

📊 Results

Final outputs available at:

support_tickets/output.csv

---

🧠 Key Learnings

• LLMs need retrieval + constraints to be reliable
• Pure AI → unsafe
• Hybrid systems → production-ready

---

🏁 Hackathon

Built for HackerRank Orchestrate (May 2026) Results announced: May 15, 2026

---

✨ Author

Kshitija Sharma

---

⭐ If you found this interesting

Give it a star ⭐ and explore the architecture!