M1CR0W4V3/campus-assistant-router

Campus Assistant Multi-Label Edge Router (ONNX)

This is an optimized, ONNX-quantized DistilBERT model trained to route student queries to 18 specific university departments. It is designed for FAANG-grade multi-agent orchestrator architectures.

Because it is quantized to ONNX, it achieves a ~4x latency reduction on CPU compared to standard PyTorch models, making it ideal for edge deployment.

Supported Domains (18)

The model supports multi-label classification (can detect multiple intents at once) across the following domains: Academics, Admissions, Athletics & Recreation, Campus Security & Safety, Career Services, Dining & Nutrition, Disciplinary, Facilities, Fees & Finance, General, Health & Wellness, Housing, IT Helpdesk & Tech Support, International, Library Services, Registration, Student Life, Transportation & Parking.

How to use this model

Anyone can pull this model and use it in their own Python applications using the optimum library.

1. Install Dependencies

pip install transformers optimum[onnxruntime] torch

2. Run Inference

Here is the exact code snippet to run multi-label classification:

import torch
import json
from transformers import AutoTokenizer
from optimum.onnxruntime import ORTModelForSequenceClassification

# 1. Load Model & Tokenizer
model_id = "your-username/campus-assistant-router" # Change this to the repo name
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = ORTModelForSequenceClassification.from_pretrained(model_id)

# 2. Define Query
query = "I need to pay my tuition and book a study room in the library"
inputs = tokenizer(query, return_tensors="pt", truncation=True, max_length=128)

# 3. Run Inference
with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.sigmoid(outputs.logits)[0] # Sigmoid for multi-label

# 4. Extract Domains (Threshold = 0.2)
threshold = 0.2
above_threshold = (probs > threshold).nonzero(as_tuple=True)[0]

domains = []
# You can load the id2domain mapping from label_mapping.json in the repo!
id2domain = model.config.id2label 

if len(above_threshold) > 0:
    for idx in above_threshold:
        domains.append(id2domain[idx.item()])
else:
    # Fallback to highest confidence
    confidence, predicted_class = torch.max(probs, dim=-1)
    domains.append(id2domain[predicted_class.item()])

print(f"Query: {query}")
print(f"Routed Domains: {domains}")

Intended Use

This model is intended to act as the "Semantic Router" for Multi-Agent AI systems (like LangGraph or AutoGen). Instead of a single LLM trying to answer everything, this ONNX model instantly categorizes the user's intent and triggers the correct sub-agent downstream.