You can deploy this model and start earning money today!

Ttimms/
sovereign-edge-intent-router

Name: sovereign-edge-intent-router
Availability: InStock

intent-classification

edge

jetson

distilbert/distilbert-base-uncased

quantized

mit

Views

No views yet

sovereign-edge Intent Router

DistilBERT fine-tuned for 5-class single-label intent classification, routing user queries to one of five domain experts in a LangGraph multi-agent system deployed on a Jetson Orin Nano.

Status: snapshot, not under active development right now. This is the router model behind the sovereign-edge project, which is mid re-architecture. It may be retrained or superseded — check the GitHub repo for the current state before assuming this is the latest version.

Architecture

mermaid

graph TD
    Q["user query"]
    T1["Tier 1 - embedding similarity"]
    T2["Tier 2 - THIS model: DistilBERT 5-class intent<br/>INT8 ONNX 64 MB - &lt;10 ms on Jetson Orin Nano CPU"]
    T3["Tier 3 - keyword fallback (always available)"]
    E["one of 5 LangGraph expert subgraphs<br/>spiritual / career / intelligence / creative / goals"]
    Q --> T1 --> T2 --> T3
    T1 --> E
    T2 --> E
    T3 --> E

What this is part of

This model is Tier 2 of a three-tier intent router: embedding similarity (Tier 1) → this DistilBERT classifier (Tier 2) → keyword matching (Tier 3, always-available fallback). The system degrades gracefully without this model — see packages/router/src/router/classifier.py in the GitHub repo for the full routing logic.

Labels

ID	Label
0	spiritual
1	career
2	intelligence
3	creative
4	goals

Each label routes to a dedicated LangGraph expert subgraph. The shipped classifier has these 5 trained labels; the project's router adds a 6th GENERAL class, handled by the keyword-fallback tier and a low-confidence threshold rather than by this model — so project docs that describe a "6-class" router are consistent with this 5-label checkpoint.

Formats included

Path	Format	Size	Use case
`hf_model/`	HF `transformers` (`safetensors`)	256 MB	Fine-tuning, evaluation, re-export
`router_fp32.onnx` (+ `.onnx.data`)	ONNX, fp32	536 MB	Reference export, re-quantization source
`router.onnx`	ONNX, INT8 quantized	64 MB	Production — deployed on Jetson Orin Nano CPU, <10ms inference

Usage

HF transformers

python

from transformers import AutoModelForSequenceClassification, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("Ttimms/sovereign-edge-intent-router", subfolder="hf_model")
model = AutoModelForSequenceClassification.from_pretrained("Ttimms/sovereign-edge-intent-router", subfolder="hf_model")

ONNX (production path)

python

import onnxruntime as ort
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("Ttimms/sovereign-edge-intent-router", subfolder="hf_model")
session = ort.InferenceSession("router.onnx")  # download router.onnx from this repo
inputs = tokenizer("How do I plan my week better?", return_tensors="np")
outputs = session.run(None, dict(inputs))

Training

Trained with scripts/train-router.py in the GitHub repo — standard HF Trainer fine-tune of distilbert-base-uncased for 5-way single-label sequence classification.

License

MIT — matches the upstream project license.