INTELLIGENT DYNAMICS
INVESTOR PRESENTATION
SPEECH / LANGUAGE / EXPRESSION
01 / 16
OUR AMBITION
Robust Solutions Despite Constraints.
Developing high-performance, robust models in data-constrained environments.
VOICE / EXPRESSION
But what if we don’t have enough data?
Audio example: audiodynamics.dev
INTELLIGENT DYNAMICS
PROBLEM
SPEECH / LANGUAGE / EXPRESSION
02 / 16
THE PROBLEM
A specialized task with little to no data.
01
The data gap.
Organizations may need robust specific solutions without having the amount of data that standard methods may require.
02
The execution gap.
Methods that rely on massive scale may fall apart when there is little data or very limited computational resources.
Our goal: bring intelligence to a data-constrained environment.
INTELLIGENT DYNAMICS
PRODUCT + PROGRESS
SPEECH / LANGUAGE / EXPRESSION
03 / 16
DEVELOPMENT APPROACH
A solution that solves your problem, not someone else’s.
INPUT
Task-specific data.
Process and augment data.
ADAPTATION
Training pipeline.
Design a custom training pipeline for your solution.
APPLICATIONS
Custom
software
New model.
→
→
IMPLEMENTED
Fine-tuning pipeline + Jeju text-to-speech system
DEMONSTRATED
Recorded song and dialogue dubbing studies
INTELLIGENT DYNAMICS
TARGET CUSTOMERS
SPEECH / LANGUAGE / EXPRESSION
04 / 16
TARGET CUSTOMER PROFILES
Why not use another solution?
OTHER SOLUTIONS
IN-HOUSE TRAINING
Cost and efficiency.
It can be inefficient to use someone else’s solutions to your problem. For example, our extremism classifier targets lower-cost, task-specific inference.
Why not train ourselves?
Training in a data-constrained environment is a very particular domain. Methods such as adversarial weight perturbation need task-specific adaptation and validation.
Commercial priority: validate the strongest initial customer segment.
INTELLIGENT DYNAMICS
TARGET CUSTOMERS
SPEECH / LANGUAGE / EXPRESSION
05 / 16
TARGET CUSTOMER PROFILES
Our target customer is
B2B / ORGANIZATIONS
B2C / INDIVIDUALS
Someone whose needs are not met.
Perhaps their current approach is too expensive, not performant enough, or not efficient enough.
Someone with little data.
They might work in a data-constrained environment, such as health care.
Commercial priority: validate the strongest initial customer segment.
INTELLIGENT DYNAMICS
THE AMBITION
SPEECH / LANGUAGE / EXPRESSION
06 / 16
Let’s take a look at some of our solutions.
You can view a folder of our audio examples here.
audiodynamics.dev
EXPLORE THE DEMOS
SCAN TO OPEN
INTELLIGENT DYNAMICS
DEMO / 01
WATCH ONLINE / audiodynamics.dev
07 / 16
RECORDED PERFORMANCE STUDY
One song. Two languages.
CHINESE → ENGLISH
00:28
Original
Chinese
00:58
Generated
English
The challenge: preserving meaning and pacing across languages.
INTELLIGENT DYNAMICS
DEMO / 02
WATCH ONLINE / audiodynamics.dev
08 / 16
RECORDED PERFORMANCE STUDY
One scene. Three performances.
JAPANESE → ENGLISH
01
Original Japanese
Source performance
02
Previous English
Earlier version
03
New English
Experimental audition
Here’s another real world use case.
INTELLIGENT DYNAMICS
MARKET SIZE + OPPORTUNITY
SPEECH / LANGUAGE / EXPRESSION
09 / 16
EXTERNAL MARKET CONTEXT
A large category.
A narrow entry point.
GLOBAL AI VOICE GENERATORS · SOFTWARE + SERVICES
2026 / ESTIMATE
$7.7B
2030 / FORECAST
$21.8B
TARGET OPPORTUNITY
Custom speech
where standard
fit is not enough.
Specific languages.
Specialized workflows.
Limited task-specific audio.
Category estimates, not our serviceable market or revenue forecast.
Source: Grand View Research, AI Voice Generators Market (2024–2030) · accessed 15 Sep 2026
INTELLIGENT DYNAMICS
COMPETITIVE LANDSCAPE
SPEECH / LANGUAGE / EXPRESSION
10 / 16
SELECTED ALTERNATIVES
Audio is a crowded field, but we believe our solutions are robust.
ALTERNATIVE
Documented capability.
Goals to validate.
ElevenLabs
Multilingual dubbing; speaker
identity and emotional tone.
Live and on-device performance.
Microsoft
Azure Speech
Custom speech recognition
and custom synthetic voices.
Highly expressive audio.
OpenVoice
Short-reference voice cloning
and cross-language style control.
Control and performance.
Our test to win: less task-specific data at matched quality.
Sources below; vendor-reported capabilities, not head-to-head results.
ElevenLabs
Microsoft: recognition
Microsoft: voice
OpenVoice
Reviewed 15 Sep 2026
INTELLIGENT DYNAMICS
UNIQUE ADVANTAGE / MOAT THESIS
SPEECH / LANGUAGE / EXPRESSION
11 / 16
MOAT THESIS
From custom models
to cumulative know-how.
BUILT
Adaptation
capability
Fine-tuning pipeline
and Jeju text-to-speech
development.
TO VALIDATE
Training-data
efficiency
Fewer task-specific
recordings at the same
quality level.
POTENTIAL TO COMPOUND
Reusable
methods
Carry useful adaptation
methods into the next
language or workflow.
INTELLIGENT DYNAMICS
BUSINESS-MODEL DESIGN
SPEECH / LANGUAGE / EXPRESSION
12 / 16
COMMERCIAL MODEL
Custom scope.
Recurring access.
B2B / CUSTOM PROJECTS
B2C / SUBSCRIPTION + USAGE
Project revenue
Pricing reflects the language,
workflow and scope of the
custom speech solution.
Recurring access
Subscriptions and pay-as-you-go
fees give consumers access
to speech tools.
Next commercial proof: a priority buyer, willingness to pay and delivery economics.
INTELLIGENT DYNAMICS
PIPELINE REUSE
SPEECH / LANGUAGE / EXPRESSION
13 / 16
We are also quite experienced with natural language processing.
Let’s attempt to create a language model from scratch.
Since I did, this mostly just for fun and trained on my laptop 100 million parameter, 60 hour estimate is wrong. I actually paused the training run several times so I could play video games or do other stuff.
You can run a slightly nerfed version of this model on your browser via WebGPU on our website.
INTELLIGENT DYNAMICS
PIPELINE REUSE
SPEECH / LANGUAGE / EXPRESSION
13.5 / 16
What about computer vision?
Let’s do something no one’s done before.
There are two models. First is an image decomposition model which decomposes existing images into valid Photoshop sequences. The second takes textual prompt and creates novel sequences to create a novel image.
INTELLIGENT DYNAMICS
TEAM BACKGROUND
SPEECH / LANGUAGE / EXPRESSION
14 / 16
THE FOUNDERS
Research depth. Hands-on execution.
Oskar Jonientz
CO-FOUNDER
Andrew
CO-FOUNDER / UCLA COMPUTER SCIENCE
Model engineering
Built a hybrid Mamba–Transformer
masked-diffusion model in PyTorch.
Customer-facing software
Built and sold software; supported users
and incorporated product feedback.
Cross-team execution
Boeing internship, summer 2024: business
case development and technical proposals.
Published AI researcher
Worked with Georgetown professors on
agentic ML for high-dimensional biomedical data.
Language models + retrieval
Experience building and evaluating
language-model and retrieval systems.
ML pipelines + evaluation
Python-based machine-learning pipelines
and rigorous model-performance testing.
INTELLIGENT DYNAMICS
DEVELOPMENT STAGES
SPEECH / LANGUAGE / EXPRESSION
15 / 16
PROPOSED PROOF POINTS
Make it work.
Then make it repeatable.
01
Measure
Compare data required at
matched quality against
named baselines.
02
Validate
Test a specific buyer’s
workflow, budget and
acceptance criteria.
03
Repeat
Reuse what works; measure
delivery cost and
repeat demand.
Quality evidence first. Commercial evidence next.
INTELLIGENT DYNAMICS
THE AMBITION
SPEECH / LANGUAGE / EXPRESSION
16 / 16
Audio Examples
You can view a folder of our audio examples here.
OSKAR JONIENTZ / CO-FOUNDER
oskarjonientz@gmail.com
audiodynamics.dev
EXPLORE THE DEMOS
SCAN TO OPEN