For 1517 Fund · May 2026

We gave voice to thought.

Non-Invasive Brain-to-Speech. No surgery. No movement.

The first non-surgical, no-movement system that reaches fully locked-in patients.

The next interface between human and machine is thought. Natural. Not operated. Not spoken to. Not typed at. A computer the brain accepts as part of itself — the way it accepts a limb — where it speaks and the computer speaks back. That is what we are building.

This is Phase One. We start where the need is most urgent — 70 million people who cannot speak. Not because they have nothing to say. Because nothing can hear them yet.

"God has commanded us to advance." — The sentence that started this.

Contact founder

SCROLL

The Proof

Watch what happens.

Real EEG signal. Real imagined sentences. The system is running on actual brain data — not synthetically generated, not reconstructed from text. Imagined speech neural activity patterns, decoded by the model.

Dataset: ChiSCO — the only large-scale published sentence-level imagined speech dataset · 23,000 recordings · Scientific Data, Nature 2024

▶

REAL-TIME DEMO — Running on Real EEG Data · Actual decoded output · Not simulated

↗ Open on YouTube

▶

CONCEPTUAL DEMO — EEG imagined speech visualization · Illustrative, not a live session

↗ Open on YouTube

The system decoded what they were thinking. 84% of the time, in the top 5 guesses. 24.5% exact match. From imagined sentences. No speaking. No movement. No benchmark existed for this task — in research or industry — before we built one.

84% on 197 sentences is not a research result awaiting productization. AAC research shows 100–200 core phrases cover 80% of daily communication needs. Our 197-sentence corpus was not chosen arbitrarily — it is aligned with thirty years of AAC research. The number is not a research constraint. It is the right number for the first product. The right vocabulary, reliably decoded — gives a fully locked-in patient the ability to communicate pain, need, love, goodbye. That product does not exist today without surgery. It exists now. And as the system learns each individual's EEG patterns over time, Top-5 becomes Top-1 for that person. The first guess is right. Every time.

Act II · The Gap

For decades, the promise was: technology will reach them.

Billions of dollars. Brilliant engineers. Brain-computer interfaces that changed lives. And yet — the fully locked-in were never reached.

Invasive BCIs

Open-skull surgery

Neuralink. BrainGate. Synchron.

High accuracy. Remarkable results for motor tasks.

Required: a neurosurgeon, a hospital, a procedure that carries real risk.

Reaches those willing — and able — to undergo surgery.

Non-invasive alternatives

Residual movement required

Eye-tracking. Muscle switches. Breath sensors. Blink patterns.

No surgery. Genuinely accessible.

Required: some fragment of voluntary movement. A twitch. A blink. A breath.

These are selection-based systems — the person chooses from displayed options using residual movement. Powerful tools that genuinely help the partially disabled. But they require movement, and they select rather than generate speech from thought. Zero movement means zero access. Not the fully locked-in.

The fully locked-in

Zero movement. Full consciousness. Nothing reaches them.

The gap has been open for decades.

Non-Invasive Modalities — Why EEG Is the Only Viable Path

Modality	Latency	Consumer Hardware	Natural Communication	Product Today
fNIRS	5–6s biological delay	Language-capable fNIRS requires 388-channel whole-head system ($150,000–$350,000). Consumer EEG for language decoding: ~$1,900.	No — 5–6s hemodynamic delay (biological, not engineering)	MindPortal/MindSpeech (professional high-density fNIRS)
MEG	Milliseconds	No — room-sized	No — not wearable	Brain2Qwerty (Meta, research)
fMRI	Seconds	No — hospital only	No	None
EEG ✓	Milliseconds	Yes — Emotiv, OpenBCI	Yes — millisecond-latency signal	Excelleve

EEG is the only non-invasive modality with millisecond latency, language-capable consumer hardware, and no biological ceiling. Consumer fNIRS exists but is limited to prefrontal focus monitoring (1–4 channels) — language decoding requires high-density professional systems. EEG's historical limitations — noise and poor spatial resolution — have been solved at the methodology level. The hardware is unchanged. The results are new.

Why Now — Not 10 Years Ago

Transformers have existed for years. Consumer EEG hardware has existed for years. The window is not that these tools arrived — it is that we are the first to combine them correctly. ChiSCO published in Nature 2024 gave the first sentence-level imagined speech EEG dataset at scale. We applied Topological Data Analysis — a method that extracts structure from EEG noise that standard signal processing cannot reach. We built a neural-language architecture that works with the signal quality TDA produces — and it is language-agnostic by design. Remove any one element and the result does not exist. The window is what we built with what was available.

February 2025 · Meta AI · Brain2Qwerty

Meta AI published Brain2Qwerty — an EEG/MEG system that decodes imagined typing. The user mentally simulates typing on a keyboard while the system reads the associated brain signals. Built by a full research team with institutional resources. Their paper states explicitly: "It is not applicable to locked-in individuals, who are completely unable to perform a typing task."

Their stated next step: "adapting the typing task into an imagination task." That is what Excelleve has already done.

Until now.

Act III · The Builder

The Bund, Shanghai · Where the fuel was lit

"God has commanded us to advance. There is a solution for everything in this world except death."

My father, when I asked him as a child if we could cure those who cannot speak

That sentence is why this company exists.

I was a child when I asked it. He said — we don't have the answer yet, but in the future we will. If you pursue it, you can find it too. I never forgot that. At 13, I was researching dark matter and the theory of everything. The curiosity was not a phase. It was a direction.

At 17, in A-levels, I learned linear regression — the first building block of machine learning. A line that fits data. I sat with that idea and felt something open: with this one concept, I have a solution to every problem. Every phenomenon in nature produces data. We just need the right function. That week, I found a TED talk by Conor Russomanno, co-founder of OpenBCI — EEG signals on a screen, controlling things. I had never heard the words brain-computer interface. Within an hour, two ideas that had been waiting to meet finally did: my father's answer, and linear regression.

During my time at NYU Shanghai, six months changed everything. The obsession that had been building since I was 13 was no longer containable in a classroom. I took Andrew Ng's Deep Learning courses. He didn't teach AI as calculus or statistics — he taught it as an art. Something that, when you truly befriend it, opens completely to you. BCI became that art for me. The mathematics stopped being obstacles and became intuitions.

I took academic leave. Came back to Lahore. No lab. No team. No funding. Personal hardware. One public dataset — ChiSCO, the only one in existence — and compute constraints so tight I could not use all of it. Six months later, these are the results. That is the context for every number on this page. A-levels lit the spark. Shanghai deepened it. Lahore is where it became real.

"If this is what six months looks like with these constraints — what happens next?"

— Muhammad Huzyafa Khokhar, Founder & CEO, Excelleve · Started at 17 · Deepened in Shanghai · Built in Lahore

Act IV · The Discovery

EEG Was Underestimated. Here Is What We Found.

EEG has real limitations — noise and poor spatial resolution are genuine, well-documented constraints. The hardware has not changed. What changed is how we process the signal. Modern ML pointed toward structure in the noise. Topological Data Analysis went further — finding what persists through noise without being told where to look. The result: noise minimized to a degree that makes sentence-level decoding possible. Spatial resolution solved specifically for the language task — the Broca-Wernicke circuit localized, the right hemisphere tonal processing regions identified (specific to Mandarin as a tonal language; the Broca-Wernicke circuit is universal and will be present in English data), without any linguistic prior. We did not prove EEG's limits wrong. We solved them for our task. That is what matters.

MNE EEG Language Circuit Topomap — NeuroScan SynAmps-2 QuikCap 128

32 of 122 electrodes · Language circuit identified by TDA without linguistic prior

NeuroScan SynAmps-2 · 128-ch QuikCap · Research lab environment

01 — Mathematical analysis

17,570 EEG trials

Topological Data Analysis (persistent homology) + Phase-Locking Value across every trial. Looking for structure in the signal — not assuming where it would be. ChiSCO dataset · Nature 2024 · 23,000 recordings.

02 — The findings

32 channels. Two discoveries.

Finding 1 — The language circuit: The Broca-Wernicke language circuit identified without any linguistic prior. Left hemisphere. Confirmed mathematically and statistically — accuracy improved 1.92× against the full 122-channel baseline, and the result is consistent with established neuroanatomy.

Finding 2 — Tonal processing: Right hemisphere channels dominant in tonal language processing — active specifically because Mandarin is a tonal language (Mandarin-specific — the core language circuit finding generalizes across languages). An algorithm that had never read a neuroscience paper confirmed what neurolinguistics established over decades. That is when I knew the signal was real.

122

full array

→

language circuit

03 — The result

1.92× cleaner signal

Fewer electrodes. Dramatically better signal quality. Not channel reduction — principled topology. This is what makes a consumer-viable wearable form factor possible.

↑ 1.92× signal quality vs full 122-channel array

Act V · The Numbers

The system works. Here is exactly how well.

No prior EEG benchmark existed at sentence scale. We created one. Every result below was produced in six months, on personal compute, using one public open-source dataset — ChiSCO — under compute constraints that prevented using all available data. Every number is real, reproducible, and holds across 4 subjects — not a single-subject result.

84%

Top-5 Accuracy — Imagined Speech Retrieval

197-sentence vocabulary · ChiSCO dataset · Scientific Data, Nature 2024

AAC research: 100–200 core phrases cover 80% of daily communication · This is the right vocabulary size

Note: ChiSCO is a Mandarin Chinese dataset — the only large-scale published sentence-level imagined speech EEG dataset in the world. Results proven on Chinese. English data collection is the immediate next step and the primary use of the data budget in this round.

24.5%

Top-1 Accuracy
Exact match, first guess

0.7378

Cosine Similarity
Free-form generation pipeline
Early-stage prototype

1.92×

Signal Quality
32 vs 122 channels

+53%

Over DeWave NeurIPS 2023
On a task going deeper — purely imagined speech, no physical action whatsoever

+53% over DeWave (NeurIPS 2023) — on a task going deeper — purely imagined speech, no physical action whatsoever. DeWave used visual reading with eye-tracking assistance. We decoded purely imagined speech with no visual stimulus and no movement.

This is the first EEG system to decode purely imagined speech naturally — at sentence scale, no screen, no visual stimulus, no movement. The field has been underestimating EEG for language. ChiSCO confirmed semantic signal exists in EEG. We scaled that to sentence level — a fundamentally deeper task — and built a system that decodes it at clinically useful accuracy.

Note on open vocabulary: MindPortal (MindSpeech) reports 30% accuracy on a word-cloud prompted task using fNIRS. Our closed-sentence approach at 84% restores 80% of daily communication for locked-in patients today. True open vocabulary at useful accuracy is our 12-month target — the clinical deployments fund the data.

The fundamental question — can EEG decode imagined speech at a useful level — is answered. 84% Top-5 is the starting accuracy for every new user. As the system calibrates to an individual — their specific EEG signatures, their patterns — Top-5 becomes Top-1 for that person. That is the product arc: start useful, become precise. More data across more subjects scales this to everyone. That is what this round funds.

Act VI · The Landscape

EEG's Limitations Were Solvable. We Solved Them.

Noise and poor spatial resolution are real EEG constraints — the hardware is unchanged. But constraints in physics are not the same as constraints in methodology. A decade of BCI research told us which approaches failed and why. We came in last and built first.

Six advantages. Each one defensible on its own. Together, they form a position that compounds with every deployment.

∑

Beyond the Noise Limit

EEG is notorious for two hard limits the field accepted as physics: extreme noise and poor spatial resolution. We did not work around them. We went through them. Topological Data Analysis extracts structure that persists through noise — not by reducing noise, but by finding what survives it. The result: 1.92× signal quality improvement on 32 of 122 channels. The language circuit — Broca, Wernicke, right hemisphere tonal processing (specific to tonal languages — Mandarin in ChiSCO) — localized without a linguistic prior. Spatial resolution that EEG is not supposed to have. This is not a signal processing improvement. It is a different class of approach. Our PLV + TDA pipeline is specific to this architecture — the moat is not the trial count but the understanding of which circuit to interrogate, why topology finds it where signal processing cannot, and how the result connects to the decoder downstream.

⊞

Benchmark Ownership

No comparable benchmark existed for sentence-level imagined speech decoding from EEG — in research or industry. We built it. The person who defines how the problem is measured owns the category. Every future competitor will be evaluated against our task definition, our vocabulary size, our methodology, our numbers. When English results publish, the benchmark expands. We will be the only team to have built and validated sentence-level imagined speech decoding across multiple languages. Defining the benchmark creates an early category advantage. The methodology, the task definition, and the baseline numbers are ours.

◉

Data Flywheel

Every user generates labeled EEG training data. We reach users first because we need no surgery and no implants — just a cap. Each new user strengthens the model globally. Each session calibrates for that individual — Top-5 becomes Top-1 over time. The first data collected is English — the consumer and clinical language — built on an architecture already proven in Mandarin. The dataset compounds. The product improves with every use.

◈

EEG Is Where This Starts

We did not choose EEG despite its constraints. We chose EEG because the path is clear. Consumer hardware already exists — Emotiv, OpenBCI, Neurosity. And we are building our own: a purpose-built EEG wearable designed specifically for this task — only the channels that carry the language circuit, engineered for stable electrode contact, minimal motion artifact, and zero-prep deployment. As natural to wear as AirPods. EEG also has the only millisecond-latency signal among non-invasive modalities — the physics that makes natural real-time communication possible. EEG was always the right foundation — and we are building the hardware that makes it feel like nothing at all.

◎

Last Mover, First to Market

We started in 2025 with the entire map already drawn — which means we go to clinical deployment faster than anyone. No surgery means any clinic, anywhere. No expensive research-grade hardware. Consumer EEG caps deploy anywhere — any clinic, any home, any country. No 5–6 second hemodynamic delay means real-time communication — that ceiling is biological physics that fNIRS cannot engineer away. We will be the first non-invasive imagined speech system to reach patients at scale. Being first to clinical deployment creates structural advantages in data, clinical relationships, and reimbursement pathways that compound from day one.

⟁

Six Months, One Dataset

Every result was produced in six months, on personal compute, using one public open-source dataset — ChiSCO — under compute constraints that prevented using all of it. This is what constrained looks like. The scaling relationship is already visible in the data.

Act VII · Beyond EEG

The next interface is not measured.
It is felt.

What comes after EEG has a precise technical name: a closed-loop neural interface. A system that reads from the brain and writes back to it — simultaneously, continuously, in a feedback loop that never stops. EEG decoding is Phase One of building toward this. This is not speculative — closed-loop neural interfaces exist in clinical literature. What does not exist yet is one that works non-invasively, at the cognitive level, for natural language. That is the specific gap this company is building toward.

What Closed-Loop Means

Every current BCI is open-loop — it reads from the brain and stops there. A closed-loop interface reads and writes simultaneously. But the critical distinction is which side does the learning. In a true closed-loop system, the interface continuously adapts to your neural patterns — learning your specific signal, your timing, your language circuit in real time. The brain does not need to learn new behaviors or control patterns. The interface does the learning — continuously adapting to your neural signal in real time, updating its model of your specific patterns until communication becomes effortless. Not because you adapted to the machine. Because the machine became fluent in you.

It Exists. Not Here Yet.

Closed-loop neural interfaces are well-established in research. Deep brain stimulation with adaptive feedback is a deployed clinical example — the stimulator reads brain state and adjusts its output in real time. Sensory-motor prosthetic research at BrainGate and UCSF demonstrates that bidirectional signal flow fundamentally changes the brain-device relationship — evidence that the concept is real, not theoretical. What does not exist is a non-invasive closed loop working at the cognitive level — for language, for thought, without surgery or implants. The concept is proven at the motor and clinical level. The non-invasive cognitive application is the open problem. That is the specific gap this company is building toward.

Why EEG Is the Foundation

The writing side of a closed-loop interface requires a precise map: which circuits carry the target signal, where they sit spatially, and at what timing they fire. General neuroscience tells you language involves Broca's area. That is not sufficient for engineering. You need the exact topological signature — which positions capture the circuit, what the spatial pattern looks like, at what latency the signal appears. That is what our TDA pipeline produces. This map is the prerequisite for knowing where and when to write. And it is the prerequisite for designing hardware that requires far less than a full EEG array — once you know precisely which circuits matter and where they are, you do not need to cover the whole head. The eventual hardware may need only a single contact point, or no direct contact at all. The EEG work does not define the form factor of the final interface. It defines the knowledge that makes a minimal one possible.

Two Parallel Tracks. One Company.

This research is not phase two. It is happening now — in parallel with every clinical session, every dataset, every result on this page. Phase One builds the product and the map simultaneously. The closed-loop research uses the map.

Non-invasive stimulation approaches exist today — transcranial magnetic stimulation, transcranial alternating current stimulation, focused ultrasound. None have been combined with a real-time language decoder in a closed-loop architecture. That is the specific gap. That is where this goes.

The interface adapts to you. Not the other way around. It learns your neural language — continuously, in real time — until you stop thinking about the interface and simply think. That is neural symbiosis. That is what we are building.

You stop thinking about the interface.
You just think.
A channel that runs in both directions.

The knowledge of the world. Your assistant. Every interface — present the way thought is present. The brain receives it the way it receives sensory input: directly, without interpretation overhead.

Act VIII · The Vision

The mission is the interface. Phase One is the patients who need it most.

The company is not an EEG decoder. EEG is Phase One — where the need is most urgent and the proof is most achievable. Every phase that follows compounds on what Phase One builds: data, expertise, clinical relationships, and the neural map that makes the next interface possible.

The Mission

Where This Is Going

Decoding the brain has a fundamental limit: secondary measurement captures a representation of intent, not intent itself. What we are building toward is a different paradigm — a closed-loop neural interface where the interface continuously adapts to the brain's signal rather than requiring the brain to adapt to it. The brain does not learn the interface. The interface learns the brain. EEG proves that the brain's signals can be worked with. The closed-loop research builds on that foundation. The two tracks run in parallel. One produces the product. The other produces the platform.

Now

Thought to Speech

84% Top-5. Working AAC device for ALS, locked-in syndrome, and severe motor disability. Silent thought — decoded, confirmed, spoken.

Over time, the system calibrates to the individual — their specific EEG signatures, their vocabulary, their patterns. Top-5 becomes Top-1. For that person, the first guess is right. Every time. That is their final product.

12 Months

Neural-Language Interface

Open-vocabulary free-form generation. Any sentence, any context. Prototype at 0.74 cosine similarity (early-stage). Full capability scales with data.

Consumer

The Everyday Thought Interface

The clinical decoder proves the concept. We are building toward consumer deployment in parallel with Phase One. Consumer hardware launches alongside open vocabulary. A purpose-built EEG wearable designed specifically for this task — engineered for stable electrode contact, minimal motion artifact, and zero-prep deployment. Minimal form factor. Optimized entirely for the task. Gaming. AR and VR. Silent productivity. Hands-free control. Every person who thinks becomes a user.

Horizon

Minds to Minds — Co-Thinking

A purpose-built wearable — compact form factor, natural to wear, minimal contact with the body. And at the far end — a closed-loop neural interface deployed at consumer scale. Not a better keyboard. Not a better touchscreen. Not better voice control. The interface that makes all of them unnecessary — every device, every application, every piece of digital knowledge accessible through thought. The end state is not a computer you use. It is one you stop thinking about — because you are simply thinking. The moment it works at scale, everyone can have it — natural, portable, invisible. The way computing shifted from mainframes to desktops to phones, the next shift is to thought. That shift makes every human being a user.

The interface we are building toward has no precedent.
Not a tool you use. An extension of who you are.

Act IX · The Ask

Raising

$1M

Pre-Seed · $5M Post-Money

Data Collection — English EEG Sessions$450K · 45%

Team Building$200K · 20%

Closed-Loop Interface Research$150K · 15%

Compute & Infrastructure$120K · 12%

IRB Approval & Ethics Review$50K · 5%

Hardware Prototype$30K · 3%

Every dollar maps directly to results. Data collection drives accuracy. Hardware prototype validates the consumer form factor. IRB approval enables human subjects research. The scaling relationship is visible in current data.

Path to Series A — 18 months

6 Months

Closed-sentence pilot — 200–300 sentences

Deploy with real patients. English data collection begins alongside deployment. Right vocabulary — medical needs, daily life, family. More data opens the path to cross-subject generalisation. First non-invasive imagined speech system to reach clinical deployment.

May 2027

Open-vocabulary generation + consumer hardware

Generation pipeline deployed to real patients — any sentence, no predefined list. Consumer hardware launches alongside: a purpose-built EEG wearable designed specifically for this task. Engineered for stable electrode contact, minimal motion artifact, and zero-prep deployment. Minimal form factor. Optimized entirely for the task.

Nov 2027

Series A — $8–12M

Six months of real patient deployment: open-vocabulary generation in active use, consumer hardware validated in real-world conditions, data compounding. By November: clinical pilot data + English results + live open-vocabulary deployment demonstrated to Series A investors. Closed-loop interface research producing first architectural results.

2028+

Closed-Loop Research — First Milestones

Non-invasive stimulation mapping against TDA-identified language circuit. First closed-loop architecture experiments. This research track runs in parallel from day one — Series A funds its acceleration.

Horizon

Closed-Loop Neural Interface — Consumer at Scale

When the closed-loop neural interface is achieved, consumer deployment is not a product launch. It is a consequence. The moment it works, everyone can have it — natural, portable, invisible. Every human being becomes a user.

"The next interface between human intent and the world will not be a screen, a voice, or a gesture. It will be thought. Excelleve is building that interface."

The core retrieval question is now demonstrably tractable. The system works, today, at a level that helps real people. What this round funds is not the discovery — it is the scale.

The initial product deploys quickly — 200–300 sentences, calibrating to the individual, Top-1 accuracy over time. Real patients. Real communication. That deployment collects the data that trains the generation pipeline — any thought, any sentence, naturally. The first product paves the way. Every person we help today gives us the data to help everyone tomorrow.

Current results are on ChiSCO — the only large-scale published imagined speech EEG dataset in the world, in Mandarin Chinese. We built on the only foundation available. The architecture is language-agnostic. English data collection is the first use of this round — fast to collect because the architecture is proven, the language circuit is mapped, and the pipeline is validated — English channel identification will happen through the same TDA process that worked for Mandarin.

The end state: a computer the brain accepts as a limb — not because it learned to, but because the loop runs in both directions. The knowledge of the world, your assistant, every interface — present the way thought is present. We start with the 70 million people who cannot speak. We finish with every human being who thinks.

Reach out directly.

huzyafa@excelleve.com excelleve.com

Muhammad Huzyafa Khokhar · Founder & CEO · Excelleve