Applied AI · knowledge · decisions · automation
We design and build systems that turn data and information into usable knowledge, content, predictions, decisions and actions.
We combine machine learning, generative models, knowledge engineering, agents, software and automation to solve problems that cut across information, people, applications and, when necessary, the physical world.
02 — What we can help to solve
Finding, relating and making usable what an organisation knows.
Documents, databases, conversations, code and expert knowledge contain different parts of what an organisation knows. We build search, RAG, research and knowledge systems able to find, relate and contextualise that information so that people, models and agents can use it.
Generating quality content from data, information and knowledge.
We build systems able to organise information, select what to communicate, plan content and generate high-quality pieces adapted to different audiences, languages, durations and channels. We combine NLG, generative models, terminology, localisation and multimodal production to create text, reporting, explanations, voice, subtitles and video from a single informational base.
Turning data into predictions, signals and decisions that can be measured and improved.
We apply machine learning, forecasting, classification, ranking, anomaly detection, optimisation and decision models to extract useful signals from complex data. We design the models around the decision they have to improve, its constraints, the uncertainty and the result that can be observed afterwards.
Coordinating information, decisions and actions between people, software, agents, sensors and robots.
We build systems that do not end once they produce an answer: they query information, take or prepare decisions, act on applications, coordinate agents, handle exceptions and check results. When the operation carries on into the physical world, we integrate perception, vision, sensing, navigation, manipulation and robotic control within the same operational journey.
03 — Systems in practice
Our own projects are a way of researching by building. We work on real problems of knowledge, language, machine learning, software, agents, robotics and physical systems, and we use those projects to test architectures, methods and technologies before turning them into reusable capabilities.
ERP-Bots
ERP-Bots gathers the context of the case, applies the policy that corresponds to it, acts through typed tools on the ERP and the business applications, and asks for authorisation when the action requires it. The journey preserves its state through waits, retries and exceptions, and checks the effect in the source before considering it finished.
The complete trace of a case can be examined: what context was retrieved, what policy decided, who authorised and what signal made it possible to close it.
DocuLogic · KIR
DocuLogic reconstructs the logic and the operational knowledge that today live spread between code, configurations, documents, tables, tests and traces, and preserves them as knowledge that is legible, versionable and joined to its evidence.
Each rule arrives with its provenance and its state: observed behaviour, inferred logic, validated knowledge or agreed decision. The map of rules of a domain can be examined, along with the evidence that sustains each one and the contradictions the system shows instead of resolving on its own.
Finaz
Finaz researches how to turn complex data and information into high-quality content systematically. It separates data, calculations, content selection, narrative planning and linguistic realisation in order to produce different explanations, audiences, languages and durations from a single representation.
On that architecture it combines NLG, generative models, localisation, voice, subtitles and video. The same information can become a short briefing, a long explanation, a narration or an audiovisual piece without treating each format as an independent problem.
Project Memory RAG
A RAG that is useful for complex projects needs to do more than search for semantically similar fragments. Project Memory RAG combines lexical and vector retrieval with specialised indexes for documentation, conversations and code, and adds reranking, context expansion and structural retrieval.
In code it can relate symbols, modules, calls, dependencies, tests and changes in Git; in documents it preserves structure and context; and in conversations it turns decisions, problems, constraints and results into reusable memory.
The aim is that an agent or a person can retrieve not just similar content, but the context they really need in order to carry on with the work.
SuperSocrates · SuperQuestions
SuperSocrates looks for what is missing and what does not fit: information gaps, data that does not match between sources, decisions that contradict one another and assumptions nobody got round to writing down. Instead of closing an answer, it proposes the next useful question and shows what evidence would resolve it.
SuperQuestions is a different product: it represents and manages the questions of a project together with their evidence, their decisions, their dependencies and the work they open. A specific collision can be examined with the two sources that gave rise to it and the question it opens.
SuperPythagoras · SecondOpinion
SuperPythagoras is the harness with which we build software when people and bots share the work: it assigns roles, sets delivery contracts, orders the handovers and demands review before a change advances. SecondOpinion adds an independent perspective on every change with consequences, formulated by whoever did not write it.
The journey of a change can be examined: who proposed it, under what contract, what the second opinion observed and what evidence remains of why it was built that way.
PentaEmbeddings
PentaEmbeddings researches specialised vector representations for financial information and time series. The aim is that patterns of prices, returns, volatility, volume, microstructure and market context can be compared, retrieved and combined with other modalities inside machine learning and search systems.
As opposed to embeddings designed solely for language, we explore representations able to capture temporal structure and market behaviour: finding similar periods, recognising regimes, retrieving precedents and connecting financial series with textual and fundamental information.
It is a line of work at the intersection between embeddings, time series, multimodal representation and financial machine learning.
Excorpora
Excorpora researches how to model what must change, what may change and what must remain unaltered when we transform language between varieties, languages, registers and contexts.
It combines contrastive research, computational terminology, synthetic corpora, semantic representation, generative models and adversarial evaluation in order to treat translation and localisation as a problem of linguistic engineering, not solely of fluency.
MagdalenOS · MontojOS · Visort · IoRT · Utobot
Our work in Physical Intelligence brings together several complementary lines. MagdalenOS researches robotic manipulation: geometry, kinematics, motion planning, tools, loads and collision avoidance. MontojOS works on navigation and control of mobile robots, combining perception, state estimation, localisation and trajectory planning.
Visort explores computer vision as a perception layer for recognising, locating and characterising whatever a system has to decide or act upon. IoRT extends these capabilities with an Internet of Robotics Things architecture that connects robots, sensors, cameras, models, agents and digital systems within a single operation.
Our spin-off Utobot takes part of this experience into applied robotics, with systems for fields such as kitchens, logistics, distribution, cleaning and the automation of spaces.
Together, these lines research one and the same problem: how to make perception, decision, software and physical action part of a single intelligent system.
Kinematics · planning · tools · loads · collisions
Localisation · perception · trajectories · control · autonomy
Computer vision · perception · localisation · inspection
Robots · sensors · agents · models · software · events
Kitchens · logistics · distribution · cleaning · automation
Each project makes a journey visible: what comes in, what criterion is applied, what acts and how the result is checked.
04 — How we build these systems
Different problems require different forms of intelligence. We combine systems that organise knowledge, components that interpret and decide, and operational architectures able to keep the work running until it produces a result.
Establishes what the system knows, where it comes from and what it means.
We bring together documents, data, language, code, rules and expert knowledge; we resolve entities, versions and contradictions; and we build a common representation with meaning, provenance and validation status.
That base can feed research, search, reporting, translation, localisation, NLG composition, assistants, agents and project memory. It can also make examinable the logic that today lives spread between software, configurations, documentation and experience.
They interpret the context and coordinate the next step within defined limits.
We build agents with typed tools, bounded memory, durable workflows, policies and human intervention, able to interpret the context and choose the next step.
Orchestration coordinates that work between models, software and people, and preserves the objective, the state, the permissions and the completion criterion through to a result that can be evaluated.
Preserves the state and carries the work through to a checkable result.
We represent the state of the work, apply logic and policies, coordinate agents, software and people, keep processes alive through waits and exceptions and verify the effect in the appropriate source.
The operation can live in an ERP, cross several applications, carry on for months or extend into the physical world. The system preserves who holds the next responsibility, what authority it needs and what signal makes it possible to consider it completed.
05 — Capabilities
We work from the sources and the representation of knowledge through to the models, the software and the orchestration that sustain decisions and operations. Each capability occupies a defined function and connects with the others through contracts, state and evaluation criteria.
We explore new technologies and build on lasting foundations. The selection is argued by function, data, cost, latency, responsibility, maturity and result.
Explore capabilities06 — The real work
The work of an organisation rarely fits into a single tool or a single answer.
01
Documents, data, code, conversations and people's knowledge contain different parts of the context.
02
Expert people spend time searching, cross-checking, calculating and gathering information before they can decide.
03
A decision then has to become actions, approvals, changes in an ERP, messages, waits or physical work.
04
When exceptions appear it is difficult to know what happened, what rule was applied, what is missing and who holds the next responsibility.
We build systems to join those four parts: knowledge, decision, execution and checking.
07 — Culture 3.14
We are interested in problems that demand understanding before simplifying. We listen to different positions, clarify concepts, examine assumptions and build systems that make it possible to bring ideas into contact with reality.
Ambition leads us to explore technologies and connect disciplines. Humility allows us to distinguish between intuition, evidence, prototype and operational capability; and to review a decision when the work demonstrates something different.
08 — Physical Intelligence · Laboratory
AI stops being software alone when it has to understand a scene, estimate where it is, decide a movement, manipulate an object or coordinate with sensors and machines.
In our Laboratory we work on computer vision, perception, sensor fusion, state estimation, robotic navigation, motion planning, manipulation, control and edge computing. MagdalenOS, MontojOS and Visort allow us to research these capabilities on real physical systems.
With IoRT we also explore how to join robots, sensors, cameras, models, agents and applications in a single operational architecture. And through our spin-off Utobot, this experience connects with applications in service robotics, logistics and automation.
We are especially interested in the frontier where a digital decision has a physical consequence and the system then needs to observe what actually happened.
09 — Building together
We listen to the work from different positions: those who carry it out, those who take decisions, those who maintain the systems and those who answer for their consequences. We go through the ordinary cases and the ones that put the process to the test; we bring together their perspectives and make visible the knowledge that today lives spread out.
With that understanding we formulate together what deserves to change, what is worth preserving and where intelligence can contribute. The proposal takes shape through observable objectives, limits, responsibilities and a first complete journey that the team can examine and put to the test.
We then build on shared artefacts: sources, knowledge models, decisions, interfaces, prototypes, traces and evaluations. Each iteration incorporates the experience of the people who will use the system and the evidence of what happens when it starts working.
A person at 3.14 will reply.