Skip to content

About

Grounded multi-domain support-triage agent (HackerRank Orchestrate Hackathon). 8-layer DAG fusing ChromaDB vector + BM25 lexical retrieval, with an adversarial-critic pass that blocks ungrounded answers — 93% groundedness across a 772-document corpus.

Topics

Resources

Stars

2 stars

Watchers

0 watching

Forks

Repository files navigation

🚀 HackerRank Orchestrate: Grounded Multi-Domain Triage Agent

License: MIT Accuracy: 93% Architecture: 8--Layer%20DAG

A production-grade support triage agent built for the HackerRank Orchestrate Hackathon. This agent autonomously handles tickets for HackerRank, Claude, and Visa with zero-hallucination guarantees.


💎 Key Technical Innovations

  • 8-Layer DAG Architecture: Orchestrates retrieval, reasoning, and adversarial verification in a resilient graph.
  • 93% Groundedness: Every response is strictly mapped to the corporate corpus; no parametric "guessing."
  • Adversarial Critic (L7B): A secondary LLM pass designed specifically to "catch" and block potential hallucinations before output.
  • Hybrid Search: Fuses ChromaDB Vector search with BM25 Lexical search for sub-millisecond, highly relevant retrieval.

📁 Repository Structure

  • code/: The core engine (Logic, Retrieval, Layers).
  • docs/: Technical deep-dives (Architecture, Post-mortem).
  • support_tickets/: Evaluation inputs and the high-accuracy output.csv.
  • data/: The support corpus (772 files).
  • log.txt: Full AI-collaboration transcript for transparency.

🛠️ Installation & Setup

See the Code README for detailed environment setup and run instructions.

📈 Technical Deep-Dives

  1. Architecture Breakdown: How the 8-layer pipeline works.
  2. Post-Mortem & Optimizations: Lessons learned on throughput and rate-limiting.

⚖️ License

MIT License - See LICENSE for details.


Created by Hitanshu (@hitanshu04) for the HackerRank Orchestrate Challenge 2026.

About

Grounded multi-domain support-triage agent (HackerRank Orchestrate Hackathon). 8-layer DAG fusing ChromaDB vector + BM25 lexical retrieval, with an adversarial-critic pass that blocks ungrounded answers — 93% groundedness across a 772-document corpus.

Topics

Resources

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages