https://github.com/swayam-mishra/hackerrank-orchestrate
HackerRank Orchestrate hackathon AI agents: a multi-modal damage-claim verifier (3rd of 1,773) and a support-triage RAG agent (60th of 1,349). The model describes; plain code decides
https://github.com/swayam-mishra/hackerrank-orchestrate
ai-agents hackathon llms multimodal prompt-engineering rag
Last synced: about 16 hours ago
JSON representation
HackerRank Orchestrate hackathon AI agents: a multi-modal damage-claim verifier (3rd of 1,773) and a support-triage RAG agent (60th of 1,349). The model describes; plain code decides
- Host: GitHub
- URL: https://github.com/swayam-mishra/hackerrank-orchestrate
- Owner: swayam-mishra
- License: mit
- Created: 2026-06-07T16:28:14.000Z (23 days ago)
- Default Branch: main
- Last Pushed: 2026-06-27T08:09:15.000Z (3 days ago)
- Last Synced: 2026-06-27T09:07:14.175Z (3 days ago)
- Topics: ai-agents, hackathon, llms, multimodal, prompt-engineering, rag
- Language: Python
- Homepage:
- Size: 10.3 MB
- Stars: 1
- Watchers: 0
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# HackerRank Orchestrate: my hackathon projects
This repo holds my entries to **HackerRank Orchestrate**, a monthly 24-hour event where you design
and ship an AI agent. Each month lives in its own folder with its own data, code, and a single README
that walks through the whole thing. If you're here to learn from the code, start with whichever
project sounds more interesting and read its README top to bottom.
| Month | Project | What it does | Model | Result |
|---|---|---|---|---|
| **June 2026** | [Multi-Modal Damage-Claim Verification](./hackerrank-orchestrate-june/) | Looks at the photos on a damage claim and decides whether they back up the claim, contradict it, or aren't enough to tell. Fills in 14 explained columns. | Claude Opus 4.8 (vision) | 🥉 **3rd of 1,773** |
| **May 2026** | [Support Triage Agent](./hackerrank-orchestrate-may/) | Reads support tickets for three help desks (HackerRank, Claude, Visa) and either answers them from local docs or sends them to a human. Fills in 5 columns. | Claude Haiku 4.5 | **60th of 1,349** |
Each project keeps its own scraped final leaderboard:
[`leaderboard-june.json`](./hackerrank-orchestrate-june/leaderboard-june.json) and
[`leaderboard-may.json`](./hackerrank-orchestrate-may/leaderboard-may.json).
---
## June 2026: Multi-Modal Damage-Claim Verification
The whole project rests on one idea: **the vision model only describes what it sees, and plain code
makes the actual decision.** Keeping those two jobs apart pays off in a few ways. You get the same
answer every time (the decision is just code, so you can save the model's observations and re-run the
logic offline with no API calls). It's auditable, since every column traces back to a specific rule.
And it resists trickery: an instruction hidden inside an image can confuse one observation but can
never reach the code that decides. A person's claim history can add caution, but the structure makes
it impossible for history to flip a verdict on its own.
**[Read the June README](./hackerrank-orchestrate-june/README.md)** for the architecture, the
commands, the decisions, and the results.
## May 2026: Support Triage Agent
A command-line agent that answers or escalates support tickets using only a local set of help docs,
with no internet calls while it runs. Each ticket goes through a pipeline: filter out junk, clean up
the search query, search the docs (fast keyword search, then a small local model re-orders the best
hits), score the confidence, check a few hard escalation rules, have Claude draft a reply, validate
and if needed repair it, strip anything the model made up, and log the reasoning. As in June, the
model only writes; the routing decisions are plain code you can read.
**[Read the May README](./hackerrank-orchestrate-may/README.md)** for setup, a file-by-file guide,
the decisions behind it, and the results.
---
## How this repo is organized
```
.
├── hackerrank-orchestrate-june/ June project (code/, dataset/, README.md, leaderboard-june.json)
├── hackerrank-orchestrate-may/ May project (code/, data/, support_tickets/, README.md, leaderboard-may.json)
├── .gitignore one shared ignore file for the whole repo
└── README.md you are here
```
## Good to know
- There's **one `.gitignore` at the top** covering both projects (secrets, virtual environments,
caches, and generated run files). Each project keeps its own `requirements.txt` and `.env.example`.
- **Secrets stay out of git.** The `.env` files are ignored and never committed. To run either agent,
copy that project's `.env.example` to `.env` and paste in your own `ANTHROPIC_API_KEY`.
- The challenge **datasets, problem statements, and rulebooks belong to HackerRank.** Each project's
README says what's included and what isn't.
- **License.** My code is released under the [MIT License](./LICENSE). That covers my code only, not
HackerRank's problem statements or any bundled help-center data, which stay their owners' property.