Internal System · Central AI Core

Berlin

Berlin is our central, locally operated AI core – a tool-using research and simulation system. It combines a large local language model with source work, a memory with provenance, controlled experiments and a simulation layer – and orchestrates the other systems. All on our own hardware, without any cloud.

Start a project All internal systems

Principles

How Berlin thinks

Model evidence, not proof of reality

A simulation is an experiment in a model, not proof of reality. Berlin labels results explicitly as model evidence – not as a proven fact.

Traceable from start to end

Every research run has clear phases, fixed artifacts and checked provenance. Inputs, results and checksums are stored so others can repeat and review the run.

Controlled, not unleashed

Tools are released per task, state-changing actions require confirmation, code runs isolated in a sandbox. Risky or real-world actions need a human release.

Capabilities

What Berlin can do

Tool-assisted research

Sources & evidence

Berlin searches, reads and assesses sources, links them with its own knowledge and uses Troja as a fact-checking node.

Memory with provenance

Facts & conflicts

A local memory stores facts with provenance and validity. Contradictions are not silently overwritten but kept and resolved.

Controlled experiments

Experiment contract

Every experiment needs a hypothesis, falsifier, baseline, metrics with units, seeds, repetitions and stop criteria – defined in advance.

Simulation across domains

Physics to chemistry

From quantum chemistry through molecular dynamics and fluid mechanics to battery and climate models – established scientific solvers, invoked in a controlled way.

Multimodal analysis

Speech, image, document

Transcription, speaker separation, OCR as well as image and document analysis run locally – with a clear separation of content, algorithmic attribution and confirmed identity.

Memory-aware control

Unload & resume

For heavy computations Berlin pauses the main model, runs the job and restores the previous state exactly.

Research mode

Eight phases, one auditable run

A long run separates research into clearly bounded phases – each with its own tasks, permitted tools and artifacts. That keeps a whole reasoning process traceable.

Scope

Frame the question and boundaries.

Sources

Gather and sift sources.

Evidence

Assess and organize evidence.

Hypotheses

Turn claims into testable statements.

Experiment

Run controlled simulations.

Review

Check results and consistency.

Report

Summarize findings in a structured way.

Audit

Hash the artifacts and seal the run.

Reproducibility

Every result is traceable

Each simulation run produces input, result, environment data and a manifest with SHA-256 checksums. A result can thus be traced unambiguously to its inputs and artifacts.

Fixed solver versions and documented seeds
Complete boundary conditions, units and convergence criteria
Input, result and environment as separate artifacts
A SHA-256 manifest binds every result to its input
Results marked supported, falsified or inconclusive – always within the model
An independent review path can flag consistency errors

Reproducibility is more than file integrity – it is the condition for a simulation to become reliable evidence at all.

In the Network

The core everything groups around

Berlin is the central AI core of our network. Troja provides evidence and fact-checking, Atlas structured data, VARDR secures decisions – and Berlin orchestrates. It runs around the clock on our own hardware in Germany, without any cloud.

All internal systems →

Berlin is a local research and simulation prototype. Simulation results are model evidence and do not replace a real measurement or professional review.

Contact

Berlin for your project

Do you have a research, analysis or simulation question that should be handled locally and traceably? Tell us what you are working on.

kontakt@45technologies.de