Gradient Disco helps organisations build proprietary, sovereign and sustainable AI capabilities - optimised to your processes, running on any infrastructure, and compounding in value with every model generation.
We see these challenges across industries - in companies, public sector organisations, and NGOs that have invested in AI but are not getting what they expected.
Prompts and agent configurations tuned manually to a specific model break when the next version ships. Because the tuning was done by hand, re-tuning is a manual project - with the same cost and uncertainty as the first time, every time.
Token costs scale with every use case you add. API pricing is volatile. Model availability is not guaranteed. What starts as a pilot budget becomes a structural dependency you cannot easily walk back.
When your processes, documents, and customer data flow through hyperscaler models, you are not just using a service - you are training your future competition. Customers, transactions, and operating knowledge become someone else's asset.
Most AI projects spend the first year building infrastructure. Use cases come last - often with off-the-shelf prompts. By the time the platform is ready, the budget is spent and the results are hard to justify.
Large Language Models (LLMs) are by design extremely limited in detecting their own errors and lack the capability to tell a safe workaround from an unsafe one. Yet implementers promise reliable multi-step process automation ("Agentic AI"), ignoring the statistical properties of compounding error and known safety risks of LLMs. The result creates frustration, commercial risk, and increases pressure on human operators that need to mitigate with excessive "human-in-the-loop".
Not a platform project. Not a prompt consultancy. A structured process and reusable components that build durable AI capital.
Gradient Disco turns your organisation's data and processes into proprietary AI - optimised automatically, deployed on your terms, and designed to handle uncertainty at every step. Model the inputs, outputs and quality metrics instead of getting lost in prompts and skills. Let the optimization do the rest.
Gradient Disco's Uncertainty Quantification module gives you the layer that calculates model confidence, and thus allows agents to score their options based on a cost vs. payoff model - to pick the action with the greatest expected value. We introduce self-healing capabilities to processes and a mathematical method to safely graduate AI automations to autonomy.
We work with your organisation to collect what matters - then our optimisation engine runs automatically. The output lives in your artifact registry, fully versioned and auditable. New model available? Your library endures. Simply rerun the optimization and benefit from the improvements.
AI systems are inherently uncertain. Our runtime framework makes that explicit with two methods - picked by deployment context - to calculate an uncertainty value. We use it to decide whether to commit, retry, self-heal, or escalate to a human, based on what being wrong actually costs.
In a dynamic process, where agents collaborate across multiple systems, observing each outcome makes decisions accountable. Every step is a cost-vs-payoff decision. After a burn-in period, the Interaction Model enables agents to pick the strongest option. The preferred method for paid-API deployments.
For edge cases, running multiple versions of a model with different input parameters returns a distribution of results, instead of a single output. Measuring the spread of the Ensemble captures the uncertainty inherent in every model output. Not a feeling - a number. Ideal when you run your own compute.
Putting Interaction Models and Ensembles to action, let your models commit automatically when confidence is high, retry carefully when it wavers, and escalate to a human when the stakes are too high to risk. Our Decision Framework ensures every decision threshold is tuned to the economic cost of being wrong.
AI processes graduate to greater autonomy based on an economic performance function. If signals start diverging in production, the system reverts - automatically, before the damage compounds.
Run several copies, compare.
If they disagree → unsure.
Run once, predict expected.
Big gap to prediction → surprise.
AI is becoming a competitive battleground.
Most organisations are on the wrong side of it.
Frontier model providers are moving fast - and not just on capability. They are building consulting arms, entering joint ventures with private equity firms, and beginning to differentiate access to their most capable models based on commercial relationships. Equal access to cutting-edge AI cannot be assumed. Organisations are racing to adopt, but ignoring that their Intellectual Property leaves with the data.
At the same time, open-source and smaller models are closing the performance gap with frontier labs faster than most organisations realise. The conditions now exist to run highly capable, task-specific AI on your own infrastructure - at a fraction of the cost, with full control over your data.
Your processes, customer data, and domain knowledge are your competitive advantage. Every interaction with a third-party model is a transfer of that knowledge to infrastructure controlled by others. A sovereign AI stack keeps it inside your organisation.
Organisations that build around a single vendor are exposed to pricing changes, access restrictions, and commercial dynamics outside their control. A methodology that works across model families gives you real options - now and as the landscape evolves.
Organisations that start building proprietary AI capital today - their own training data, their own optimised models - are creating a durable structural advantage. The cost of waiting is not standing still. It is increasing lock-in and a widening gap to catch up.
We start where the value is clearest and the entry cost is lowest - and build your AI capital from there.
A sales team creates customer-facing PowerPoint presentations every day - pulling data from previous decks, adapting slide content, applying the corporate template. Done manually with a frontier model, the output regularly breaks the layout, misses brand guidelines, or requires three re-runs to get right. Each rework costs time and money. Each model upgrade resets the prompts.
Where Gradient Disco helpsWe provide a layer of abstraction that focuses on the inputs, the expected outputs, and the quality criteria - and let the system generate the most efficient prompts and skills, creating a valuable data asset that persists beyond model versions and works across providers.
At the end of the engagement, you will have inputs, expected output and quality metrics in a data library; the tuned prompts and skills in your artifact registry; and the Gradient Disco infrastructure to re-run the prompt tuning when anything changes.
A client has millions of documents - contracts, technical files, product descriptions, regulatory filings - that need to be enriched with structured metadata and have specific facts extracted from them. Running this through a frontier model API is prohibitively expensive, raises data privacy concerns, and requires a new manual setup every time the model is updated.
Where Gradient Disco helpsWe provide a layer of abstraction that focuses on the inputs, the expected outputs, and the quality criteria - and let the system train an open source model to the use case. This creates a durable AI asset that provides a competitive advantage and can complete the task at predictable cost.
At the end of the engagement, you will have your own proprietary AI model, the assets that got you there, and the Gradient Disco infrastructure to re-build the model.
Finance teams processing large volumes of supplier invoices face a consistent pattern: extraction errors, mismatched purchase orders, and incorrect GL coding that only surfaces at month-end. A frontier model can handle individual steps - but a multi-step agentic workflow means errors compound. One uncertain decision contaminates the next, and by the time a human catches it, the damage is done.
Where Gradient Disco helpsWe document the full process, then apply the right optimisation to each step - prompt engineering and skill optimisation for structured extraction, reinforcement learning for routing decisions that improve with volume. An ensemble of models runs on high-stakes fields, and an Interaction Model tracks outcomes across the entire pipeline. The decision framework commits automatically when confidence is high, escalates to a human when it is not, and tunes thresholds to the actual cost of an error.
At the end of the engagement, you will have your own proprietary trained AI model or tuned prompts, the assets that got you there, and the Gradient Disco infrastructure to re-build the model/prompts. You will have a burnt-in and ready Interaction Model for uncertainty quantification and an operational instance of the Gradient Disco Decision Framework.
AI assistants like OpenClaw and Hermes Agent can handle email drafting, scheduling, research, and task management. But deploying them from scratch means weeks of configuration, prompt engineering, and discovering failure modes the hard way - often without guardrails to prevent the assistant from taking actions it should not.
Where Gradient Disco helpsWe provide a pre-compiled version of your chosen agent, ready to deploy with sensible defaults and guardrails built in from day one. Our optimisation engine then personalises it to you: using the emails you have already written, it learns your tone, response patterns, and priorities - so drafts arrive that actually sound like you. The result is an assistant that is useful immediately and improves with every interaction.
Work with people who have been there and back again. We are a small team by design. You work directly with the persons who built the methodology and wrote the code. We believe in lean, transparent and effective project work and prefer tangible results over shiny presentations.
Julian has spent the past decade working on machine learning systems - from research to production. He has trained models, built agentic pipelines, led engineering teams, and previously ran a company in the AI space. He designed the optimisation methodology at the core of Gradient Disco and builds the technical infrastructure for every engagement. As a seasoned tech advisor, Julian is the right person to help design your future proof AI stack.
Gregor's background spans AI, consulting, and large programme delivery across Austria and internationally. In previous roles as Managing Director of Cloudflight Austria and Executive Vice President at EBCONT, he has led teams on digital transformation projects, worked on AI strategy and implementation across a range of sectors, and held full P&L responsibility for services organisations of more than 300 people. He works directly with customers from the first conversation through to delivered results. As a strategy and transformation advisor, he helps organisations build sustainable tech and leadership structure that allows them to grow in a volatile and unpredictable world.
Gregor likes to leave the car at home and ride his bike; when off work, he practices dealing with uncertainty on all kinds of human-powered adventures that typically involve mountains or water.
We work with a small number of organisations at a time to ensure the depth the approach requires. Reach out to start the conversation.
Tell us about your organisation and the AI challenge you are facing. We will get back to you within 48 hours.
Based in Vienna, Austria. Working globally.
Work with us to create proprietary, sustainable AI that compounds in value over time and builds your own AI IP. Our method includes automated prompt optimisation for off the shelf and frontier models, and fine tuning and reinforcement learning to create your own proprietary domain model.
The hard truth: Large Language Models (LLMs) are both inherently uncertain and ignorant of their own uncertainty. To automate processes of any significance, it is foolish to ignore the statistical reality of compounding errors. Our mathematical model of Uncertainty Quantification provides a much needed layer to safely graduate your AI automation from supervised to fully autonomous, implements methods for self-healing of the AI process, and gives you a robust escalation mechanisms to avoid disaster.
If you want AI transformation to accelerate, your team will need hands-on experience and the freedom to experiment. We set up "The Lab" in your organisation, work with your team and provide the guidance so this process cuts the curves and dead ends. Not as a one-time experiment, but as a living test bed for new technology and new methods of human-technology collaboration. Our engagement covers a focused and time-boxed onboarding and setup phase including team training and management advisory (technology and change leadership to build enduring value), followed by regular check-ins to support progress.
In a world of fast-moving tech, geo-economic disruptions and rapid commodification, it's crucial to build on a strategy that creates redundancies and allows your organisation to benefit from change. In these challenging times for leaders, we provide a fresh perspective and guidance - discreet, vendor neutral and down-to-earth.
Deploy AI assistants like OpenClaw and Hermes Agent to automate your work - but without the hassle of self-hosting, and with solid guardrails. This service often goes hand in hand with build-to-own and uncertainty quantification methods to reduce errors.
Read the answers to common questions about how we work, or explore our thinking on AI strategy, model optimisation, and sovereign AI in the blog.
Who is Gradient Disco for, what does "sovereign AI" mean, what infrastructure do you need, and how do you get started — all answered in one place.
Thinking on AI strategy, model optimisation, fine-tuning vs. prompting, and building AI systems that compound in value over time.
