The intent classification layer for enterprise AI agents, government workflow automation, and autonomous defence systems.
Raising £2M Seed · EIS Eligible · Kings Cross, London · conceptnet.co.uk
An independent ML researcher from the Hugging Face community ran adversarial holdout tests designed to catch fake accuracy numbers. He tried to break it.
| Test | Result |
|---|---|
| Standard test set | 100% (epochs 4 and 5) |
| Adversarial holdout | 99.3% confirmed |
| Grouped lexical-family holdout | 99.78% |
| Issues found | 4 |
| Issues fixed | All 4 — within 24 hours |
"The obvious train/test leakage explanation did not survive that check." — Independent researcher, Hugging Face community
ConceptNet classifies enterprise voice and text commands into 4 intent layers — producing structured JSON for agent execution automatically.
Token-free. 9 languages. No LLM required. Locally deployable. Air-gap capable.
Every enterprise voice command — in any language, any industry, any country — fits into exactly one of these:
| Layer | Pattern | Execution | Example |
|---|---|---|---|
| L1 Basic | "Do X" | Immediate | "Schedule a board meeting" |
| L2 Context-Aware | "Do X when Y" | Conditional trigger | "Send report when contract is signed" |
| L3 Predictive | "Do X before Y" | Proactive | "Alert manager before deadline expires" |
| L4 Autonomous | "Do X always" | Persistent agent | "Auto-update CRM after every call" |
Nobody had defined this taxonomy. Nobody had built a dedicated classifier for it. We did. It's patented.
Input (voice or text — 9 languages)
↓
FAST PATH — TF-IDF + Logistic Regression
83% coverage · <5ms · CPU only · zero cost
↓ if confidence below threshold
NEURAL PATH — DistilBERT fine-tuned
100% accuracy · <100ms · GPU optional
↓ constrained by
GRAMMAR LAYER — exactly 4 valid outputs
No hallucination · Deterministic · Auditable
↓
Structured JSON → Agent execution → Enterprise tools
| Property | What it means |
|---|---|
| Constrained grammar | Exactly 4 outputs — hallucination mathematically impossible |
| Token-free | No US cloud dependency — 3× cheaper than GPT-4o |
| Air-gap capable | No internet required — defence and government grade |
| Locally deployable | Data never leaves the organisation |
| Patents pending | Novel taxonomy — no prior art |
| Threshold | Fast coverage | Fast accuracy | Final accuracy |
|---|---|---|---|
| 0.50 | 69.2% | 95.0% | 95.9% |
| 0.55 | 60.3% | 98.9% | 99.3% |
| 0.65 | 43.2% | 100% | 100% |
Mixed semantics: L4 > L3 > L2 > L1
{
"text": "Send the report when the contract is signed",
"intent_layer": 2,
"intent_label": "Context-Aware",
"execution_mode": "conditional",
"tool": "EMAIL",
"action": "send_message",
"condition": "contract is signed",
"confidence": 0.94,
"language": "en",
"latency_ms": 3
}| Metric | Value |
|---|---|
| Total examples | 757 |
| Languages | 9 — EN, FR, ES, DE, IT, PT, ZH, AR, RU |
| Intent layers | All 4 — balanced distribution |
| L3 surface forms | "before", "ahead of", "in advance of", "prior to", "in time for", "by the time" |
| Format | JSON + CSV |
ConceptNet is available as a plugin for all major Chinese AI agent frameworks:
| Platform | Plugin | Community |
|---|---|---|
| DeepSeek Harness | conceptnet-dsh-plugin | Discussion #5559 — 211K ⭐ |
| Qwen / Alibaba | conceptnet-qwen-plugin | QwenLM ecosystem |
| Kimi / Moonshot | conceptnet-kimi-plugin | Moonshot AI ecosystem |
| GLM / Zhipu | conceptnet-glm-plugin | THUDM ecosystem |
| Doubao / ByteDance | conceptnet-doubao-plugin | ByteDance ecosystem |
| Asset | Link |
|---|---|
| Live Sandbox | conceptnet.co.uk/sandbox/ |
| Hugging Face Model | huggingface.co/conceptnetUk/intent-classifier |
| HF Space | huggingface.co/spaces/conceptnetUk/voice-ip-sandbox |
| API Documentation | docs/API.md |
| Investor One-Pager | conceptnet.co.uk/docs/investor.html |
| Metric | Number |
|---|---|
| GitHub clones | 188 from 75 unique developers |
| HF model downloads | 17+ |
| Independent peer review | ✅ Adversarial holdout confirmed |
| API token requests | Active — first within minutes of announcing |
| Plugin ecosystem | 5 Chinese AI platforms |
| Government pipeline | Nigeria FIRS, Mauritius, Rivers State, Qatar, Ivory Coast |
| Defence | UK Defence Innovation submitted — DIOL233749 |
| Built before raise | £448,000 / $600,000 |
ConceptNet/
├── README.md
├── LICENSE # MIT
├── CNAME # conceptnet.co.uk
├── requirements.txt
├── index.html # Website homepage
├── privacy.html
├── terms.html
├── core/
│ ├── train_classifier.py # Fast-path training
│ ├── train_neural.py # Neural model training
│ └── intent_grammar.txt # Constrained grammar
├── data/
│ ├── conceptnet_dataset_v2.json
│ └── conceptnet_dataset_v2.csv
├── docs/
│ ├── investor.html
│ ├── API.md
│ └── ARCHITECTURE.md
├── sandbox/
│ └── index.html # Live sandbox
├── benchmarks/
├── examples/
│ └── quickstart.py
└── api/
- Patents pending — 4-layer taxonomy, Voice IP Stacking, constrained grammar architecture
- Classifier logic — production inference code not in this repo
- Dataset — released for research only, commercial use requires licence
- © 2026 ConceptNet Ltd — all rights reserved
The model that makes ConceptNet a 10-year moat:
① PRIVATISE → Enterprise data encrypted. They own it entirely.
② STACK → Classified across 4 intent layers. Compounds over time.
③ OWN → Classified dataset = proprietary IP. Switching = losing it.
④ LICENSE → Enterprises license IP to others. ConceptNet takes 10%.
| Item | Detail |
|---|---|
| Amount | £2M Seed |
| Pre-money | £15M |
| EIS eligible | 30% tax relief for UK investors |
| Built before raise | $600,000 / £448,000 |
| CTO | Tim Storey — confirmed fractional |
| Defence | DASA DIOL233749 submitted |
| Contact | tonymomoh@icloud.com · 07733 246865 |
🇬🇧 English · 🇫🇷 French · 🇪🇸 Spanish · 🇩🇪 German · 🇮🇹 Italian · 🇧🇷 Portuguese · 🇨🇳 Chinese · 🇸🇦 Arabic · 🇷🇺 Russian
ConceptNet Ltd · Kings Cross, London · © 2026 · Patents pending tonymomoh@icloud.com · 07733 246865 · conceptnet.co.uk