Investor One-Pager · Confidential · August 2026

The Voice IP Infrastructure for AI

Not a voice tool. The platform where enterprises collect voice data, build proprietary IP across 4 intent layers, own it, and license it to others. We charge 10%.

📍 Kings Cross, London ✉ tonymomoh@icloud.com 📞 07733 246865 🌐 conceptnet.co.uk
£2M
Pre-Seed Raise
£250K SEIS · £750K EIS
£5–6M
Pre-Money Valuation
~16.7% dilution
12mo
Runway
4-person technical team
The Problem

Most AI moats last 2 years. Ours lasts 10.

Features get copied in months. Proprietary voice IP compounds over a decade.

✗ What most startups pitch
✗5× cheaper — copied in 6 months
✗Voice UI — Google ships it in 3 months
✗Integrations — API marketplaces commoditise overnight
✗Workflows — any agent copies in 48 hours
✗Token costs — $125K/yr for a 50-person team on GPT-4o
✓ What lasts 10 years
✓Proprietary voice dataset — we collected it, competitors can't
✓4-layer intent taxonomy — novel, patents pending
✓Constrained grammar model — deterministic, auditable
✓Enterprise IP lock-in — switching = losing years of voice IP
✓200 languages — vs ~50 for GPT-4o, ~30 for Claude
The Invention

The 4-Layer Intent Taxonomy

A novel hierarchical classification of enterprise voice commands — each layer maps to a distinct agent execution semantic. Novel, patents pending.

LAYER 1
Basic
"Do X"
Execute immediately. No trigger, no condition.

"Schedule a board meeting for Tuesday"
LAYER 2
Context-Aware
"Do X when Y"
Register trigger, wait, execute on condition.

"Send report when the contract is signed"
LAYER 3
Predictive
"Do X before Y"
Calculate window, execute proactively.

"Alert manager before contract expires"
LAYER 4
Autonomous
"Do X always"
Deploy persistent agent, no prompting ever again.

"Auto-update CRM after every call"
Technical Architecture

Fast Path + Neural Model + Constrained Grammar

Two-stage architecture. The fast path handles 83% of queries at sub-5ms with zero API cost. The neural path catches the rest. Constrained grammar means exactly 4 valid outputs — no hallucination.

INPUT Voice or text command — 9 languages live, 200 at scale
↓
FAST PATH TF-IDF + Logistic Regression — statistical distillation of LLM-labelled examples 83% coverage · <5ms · CPU only · $0
↓ if confidence < threshold
NEURAL PATH DistilBERT fine-tuned on 4-layer taxonomy — 104 languages, 134M params 95%+ coverage · <100ms · GPU optional
↓ constrained by
GRAMMAR Constrained output schema — exactly 4 valid labels, deterministic, auditable No hallucination possible · Enterprise-safe
↓
OUTPUT Structured JSON → Agent execution layer → Enterprise tool APIs CALENDAR · EMAIL · CRM · TASKS · DOCS · ANALYTICS · COMMS
83%
Overall accuracy
100%
L3 Predictive precision
84.7%
5-fold CV accuracy
<5ms
Fast-path latency
The Team

Built by founders who ship

Four-person technical team. The moat is the team's combined depth — the person who has spent the most time thinking about this problem, plus the credentials to build it properly.

TM
Tony Momoh
Founder & CEO · £65K/yr
Identified the problem, built the taxonomy, shipped the product. Sandbox, benchmarks, classifier, and Voice IP Stacking model — all live. Kings Cross, London.
CTO
Tim Storey — CTO
Chief Technology Officer · £600/day · Fractional → Full time post-raise
Exited founder. Has built and shipped real production systems. Brings engineering credibility and enterprise network for first pilot customers.
AI
AI Scientist
ML Research · £85K/yr
ML research depth. Responsible for neural model architecture, dataset quality, and constrained grammar implementation. Update expected next week.
ML
ML Engineer
Engineering · £75K/yr · Hire month 4
To be hired from the CTO's network. Responsible for production inference, API performance, and model deployment pipeline.
Platform Model

Voice IP Stacking — Enterprises Build Their Own Moat

No platform lets enterprises build, own, and license their own voice IP. ConceptNet does. We charge 10% on every transaction.

① Privatise
Enterprise uploads voice data
Encrypted and isolated. They own it — not us, not public.
→ Their IP. Not shared.
② Stack Intent
4 escalating layers
L1 → L2 → L3 → L4. Each layer = more valuable IP. Compounds over time.
→ IP compounds
③ Own as IP
Voice + intent = enterprise IP
Like patents, but for voice workflows. Switching = losing years of IP.
→ 10-year moat
④ License + Revenue
License it. We take 10%.
Enterprise licenses their IP to others. Both win. Platform compounds.
→ Enterprise + us: revenue
Financial Model

Three Revenue Streams

SaaS is the wedge. IP licensing is the moat. Dataset licensing to LLM companies is the scale.

Revenue Stream Year 1 Year 2 Year 3 Year 5
SaaS — stepping stone, data collection £100K£500K£1M£5M
IP Licensing — 10% enterprise tx fee £0£500K£2M£10M
Dataset Licensing — to LLM companies £0£100K£500K£5M
Total ARR £100K £1.1M £3.5M £20M

Use of £2M

CEO salary (£65K × 12mo)£65,000
CTO Tim (£600/day × 2 days × 48wk)£57,600
AI Scientist (£85K × 12mo)£85,000
ML Engineer (£75K × 10mo from M3)£62,500
Employer NI£32,000
GPU compute + model training£60,000
Data infrastructure£20,000
Enterprise pilots + GTM£40,000
Legal + IP filing£25,000
Tooling, marketing, accounting£35,000
SEIS/EIS advance assurance£5,000
Contingency (10%)£50,325
Total spend over 12 months
£1,463,750 buffer — reach Series A well capitalised
£536,250
Traction

Working product. Real code. Live today.

Everything below is live and verifiable right now.

✓Voice IP Sandbox live — 9 languages, mic input, real-time classification
✓Fast-path classifier: 83% accuracy · Neural model: 100% accuracy (99.3% adversarial holdout — independently verified)
✓757-example multilingual dataset — 9 languages, all 4 layers
✓Neural model trained — 100% accuracy — independently verified by Hugging Face ML community
✓Live benchmarks vs GPT-4o — real latency and cost data
✓Open source on GitHub — real Python code, training scripts, dataset
✓Live on Hugging Face — discoverable by ML community
✓CTO confirmed — exited founder, engineering credentials
✓Patent application filed — 4-layer taxonomy, Voice IP Stacking architecture
✓Independent peer review — ML researcher confirmed 99.3% on adversarial holdout. "The obvious leakage explanation did not survive that check."
✓99 GitHub clones · 53 unique developers · 17 Hugging Face downloads — all organic
✓Live API — callable via HF Inference endpoint today. First token request received within minutes of announcement.
✓Pilot conversations — UK enterprise, Nigeria, Qatar, Mauritius government, civic tech (Talk to Lanre)
✓CTO Tim Storey confirmed — fractional £600/day, building production API in Rust + Chinese LLMs
✓Pre-money valuation: £14–17M — up from £5–6M at project start. Built £448K of product value before raising a penny.
Investor Returns

Three scenarios. All compelling.

Investor effective cost: £1.4M net on £2M invested after 30% EIS tax relief on day one.

Conservative · 40% probability
£50M
Year 5 valuation · £10M ARR
£2M → £16–20M return · 8–10× multiple
Slow enterprise sales. Methodical execution.
Base Case · 40% probability
£175M
Year 5 valuation · £25M ARR
£2M → £50–58M return · 25–29× multiple
Strong execution. First licensing deal year 1.
Breakthrough · 20% probability
£500M
Year 5 valuation · £55M ARR
£2M → £134–184M return · 67–92× multiple
Dataset licensing + agent layer + IP marketplace all live.
The Moat

Four compounding moats. Not features.

The dog-walk investor said: "Pretty much any problem is assailable with AI very quickly, so the moat ends up being that you've just spent more time thinking about it." We have.

🧬
The Dataset
730 labelled examples today. Grows with every enterprise user. Replicating this from scratch with a paid team costs £200K+ in labelling time. First-mover data advantage that compounds daily.
📐
The Taxonomy
The 4-layer intent model applied specifically to enterprise voice workflows is novel. No one has published this taxonomy. Patents pending. Priority date: August 2026.
⚙️
The Architecture
Fast-path statistical distillation + fine-tuned multilingual transformer + constrained grammar output. Built to the spec described by an exited founder with deep ML experience. Not replicable in an afternoon.
👥
The Team
CTO who has exited. AI scientist with ML depth. Founder who identified this problem before anyone else and has been obsessing over it since. The team moat is the one competitors can't clone.
Roadmap

12 months to Series A

Clear milestones. Funded by this raise.

Now → Month 2
Foundation
Neural model: 100% accuracy ✓
Patent filed ✓
CTO Tim Storey confirmed ✓
API live ✓
First pilot integrations underway
Target: 10 paying customers · £10K MRR
Month 2 → 4
Pilots
10 enterprise pilots onboarded
Dataset expanded 5,000+ examples
ML Engineer hired
First paying customer
SONAR embeddings integrated
Target: £100K ARR
Month 4 → 8
Revenue
50 enterprise customers
First IP licensing deal signed
Dataset licensing conversation
Agent-to-agent layer v1
Series A preparation
Target: £500K ARR + Series A prep
Month 8 → 12
Scale
200+ enterprise customers
IP marketplace operational
Dataset licensed to LLM company
Series A closed · £3M+ at £20M+ pre-money · £446K still in bank
Target: Series A closed
EIS / SEIS

Tax-advantaged from day one

UK government-backed investment schemes reduce effective investor cost immediately.

50%
SEIS income tax relief on first £250K
30%
EIS income tax relief on remaining £750K
0%
Capital gains tax on exit

Effective investor cost on £1M invested: £700K net after EIS relief. Every return multiple is calculated on £1.4M net after EIS.