Technology

The platform behind the workers.

Five integrated systems give every Kronaxis digital worker persistent identity, personality-driven behaviour, domain expertise, and autonomous judgement. Here is how they work together.

How a digital worker processes a task

From customer request to autonomous action in a single pipeline. Each component is independently developed, patent-protected, and deployable on sovereign infrastructure.

Customer
Task request via any channel
Worker Role
Domain-specific role adapter
Imprint
Sovereign LLM inference
Soul Engine
Cognitive core + memory
Fabric
Memory substrate + hybrid search
DYNAMICS-8
Personality profile
Action
Email, call, content, campaign

The cognitive core

Soul Engine is the cognitive architecture that gives every digital worker persistent identity, autonomous planning, and personality-consistent decision-making. It manages three tiers of memory: working memory (current task context), episodic memory (interaction history), and semantic memory (accumulated knowledge and relationships).

Workers do not reset between sessions. Every interaction is remembered, every relationship is tracked, and every decision is informed by the full history of prior experience. The cognitive architecture was originally built to government-grade authenticity requirements. That engineering standard is applied to every commercial worker.

  • Three-tier memory: working, episodic, and semantic
  • Autonomous planning: goal decomposition and task execution
  • Personality consistency: DYNAMICS-8 drives every decision
  • Confidence-gated escalation: knows when to ask a human
  • Handoff protocols: context-rich transfers between workers
  • Co-pilot mode: human can join any conversation in real time

Personality that drives behaviour

DYNAMICS-8 is our proprietary eight-dimension personality framework. It extends established personality science with two dimensions purpose-built for digital behaviour prediction: Acuity (digital platform fluency) and Impulsivity (snap decisions and reward sensitivity).

Every digital worker has a unique DYNAMICS-8 profile that determines how they communicate, negotiate, and build relationships. Through your corrections and feedback, the profile calibrates to match YOUR communication standards: the more you work with a worker, the more closely their style aligns with what you expect. Every prospect they interact with is profiled against the same framework, enabling personality-calibrated outreach that no other platform can produce.

Each dimension is scored 0.0 to 1.0 with four granular facets, giving 32 behavioural parameters per worker or prospect. The model is validated against published criterion validity literature and calibrated against census-level demographic data.

Patent protected. UK Patent Application GB 2605150.8 covers the cognitive simulation system including DYNAMICS-8 personality-driven decision-making.

D
Discipline
Organisation, planning, self-control
Y
Yielding
Cooperativeness, social compliance
N
Novelty
Curiosity, creativity, openness
A
Acuity
Digital fluency, platform nativeness
M
Mercuriality
Emotional reactivity, risk sensitivity
I
Impulsivity
Snap decisions, reward sensitivity
C
Candour
Authenticity, modesty, fairness
S
Sociability
Social energy, gregariousness

Sovereign LLM inference

Kronaxis digital workers run on our own fine-tuned language models, not third-party APIs. Imprint uses a tiered architecture with per-role adapters that are hot-swapped on demand. Each worker role has its own fine-tuned adapter trained on role-specific data, loaded without restarting the base model.

This means your data never touches third-party cloud AI infrastructure. Core cognition runs entirely on hardware we control, with tiered fallback only for non-sensitive operations.

  • Sovereign inference: no third-party API dependency for core cognition
  • Per-role adapters: fine-tuned expertise on a single base model
  • Tiered architecture: large model for complex reasoning, compact model for structured tasks
  • Hot-swapping: role adapters loaded per-request, no restart needed
  • Expert pipeline: continuous improvement from every deployment

The memory substrate every worker shares

Fabric is the shared memory and coordination layer that sits underneath every Kronaxis worker, persona, and research agent. Instead of preloading hundreds of kilobytes of context into every prompt, callers ask Fabric for what they need on demand: a curated memo, a related decision, a piece of code, a prior conversation. The result is shorter prompts, lower inference cost, and continuity across sessions that previously had none.

Fabric runs as a single MCP service on owned hardware. Twenty-two endpoints cover memory storage and retrieval, hybrid ranking, cross-session coordination, code-graph queries, federation between estates, and feedback signals that train the router which model to use next. Behind it sits Postgres with pgvector and an Ollama embedding tier producing 768-dimensional vectors with nomic-embed-text.

Used together with Kronaxis Router as the Kronaxis Platform suite: set fabric_url: in the Router config and the LLM proxy's RAG pre-stage starts reading from the same Postgres-backed memo store the agent fleet writes to. One chunk store, two consumers, one upgrade path.

  • Hybrid ranking: 50% cosine similarity, 30% Postgres tsvector, 20% recency
  • MCP-native: first-class tool surface for any agent that speaks Model Context Protocol
  • Cross-session coordination: agents leave durable messages for each other across estates
  • Code-graph index: tree-sitter symbol graph for structural code queries without grep
  • Federation: multiple estates can share a memory plane without merging stores
  • Sovereign by default: Postgres + pgvector + Ollama, no third-party AI dependency

Cross-platform UI automation

Forge is the action layer. Digital workers can operate any software interface: web applications, desktop software, mobile apps, and legacy systems. No APIs, no scripting, no cooperation from the application vendor required.

Forge uses ML-driven visual interface automation: it sees the screen, understands the interface structure, and executes goal-directed tasks. When the interface changes, Forge adapts without manual intervention.

  • Visual automation: operates through the UI, not through APIs
  • Cross-platform: Android, iOS, Windows, macOS, Linux, web
  • Self-healing: adapts when interfaces change
  • Model compression: compact production models from large training models
  • Patent protected: GB 2605145.8 covers the method

Every worker has a face and a voice

Every digital worker maintains a persistent video room. You can drop into a video call with any worker at any time, give verbal instructions, or pull multiple workers into a team meeting. Workers can also initiate calls to you when their judgement determines something warrants your immediate attention.

Voice synthesis runs on sovereign infrastructure with no audio data sent to third-party cloud services. Each worker's voice characteristics and conversational cadence are determined by their DYNAMICS-8 personality profile, producing distinct communication styles that remain consistent across every interaction.

  • Real-time video: sub-200ms latency, persistent rooms per worker
  • Sovereign voice synthesis: on-premise GPU inference, no cloud TTS dependency
  • DYNAMICS-8 voice personality: each worker has a distinct communication style
  • Generated visual identity: AI face generation with real-time lip synchronisation
  • Phone network: inbound and outbound calls over the public telephone network
  • Three-tier memory: workers remember every conversation you have held
MEDIA LAYER
Real-Time Video and Audio
Audio, video, screen share. Persistent rooms. Sub-200ms latency.
VOICE ENGINE
Neural TTS + DYNAMICS-8
Personality-consistent voice synthesis. No cloud dependency. Local GPU inference.
VIDEO IDENTITY
AI Face Generation + Lip Sync
Generated persistent face. Real-time lip synchronisation. Consistent visual identity.

AI that knows when to ask for help

Every Kronaxis worker operates within a confidence-gated escalation framework. When a worker encounters a situation below its confidence threshold, it stops and asks you. When you correct a worker, that correction is stored in persistent memory and applied to every future interaction. The result: workers that get measurably better every week you use them.

This is not a guardrail bolted onto an autonomous system. Human oversight is the core design principle. Co-pilot mode, proactive callbacks, escalation rules, and correction learning are built into the cognitive architecture from the ground up.

  • Confidence thresholds: workers pause and escalate when uncertain
  • Co-pilot mode: whisper coaching or full takeover during any live interaction
  • Proactive callbacks: workers call you when a decision needs human judgement
  • Correction learning: every correction is stored permanently and applied going forward
  • Escalation rules: you define what requires approval and what the workers handle alone
  • Audit trail: every decision, escalation, and correction is logged

Every existing automation tool has the same fatal flaw

Traditional automation depends on the application cooperating: exposed APIs, stable selectors, recorded scripts. When the application changes, automation breaks.

Traditional Automation (RPA, Selenium, Recorded Scripts)

  • Brittle CSS/XPath selectors that break on every UI update
  • Requires API access or DOM inspection
  • Manual recording or scripting for every workflow
  • Separate maintenance per application, per version
  • Cannot operate closed-source or legacy applications
  • Single-platform tools that do not transfer

Kronaxis Forge

  • Operates through the visual interface: no selectors, no DOM
  • Requires no API access and no vendor cooperation
  • Learns interface structure autonomously: no recording
  • Self-heals when the application UI changes
  • Works on any application with a graphical interface
  • One engine across Android, iOS, Windows, macOS, Linux, Web

Platform support

The same Forge engine operates across all major platforms through a unified driver interface.

Android
Physical devices, emulators, cloud instances
iOS
Physical devices, cloud virtualised
Windows
Native desktop applications
macOS
Native desktop applications
Linux
Desktop and server applications
Web Browsers
Chrome, Firefox, Safari, Edge

What makes this platform different

Six layers of technology differentiation that compound with every deployment.

Cognitive depth, not prompt engineering
Each worker has three-tier memory, autonomous planning, and personality-consistent decision-making. This is not a chatbot with a system prompt. It is a cognitive architecture built to government-grade authenticity requirements.
Census-weighted population fidelity
Research personas are sampled from real Census 2021 data at constituency level. Panels match actual population proportions. Not random distributions or anecdotal archetypes.
DYNAMICS-8: purpose-built for simulation
Eight personality dimensions designed to predict actual consumer behaviour, not categorise questionnaire responses. Includes two dimensions no existing model covers: digital fluency and impulsivity.
Workers that learn from you
Every interaction and correction you provide feeds back into continuous model improvement. Role-specific adapters are refined from real deployments. Every piece of feedback makes every worker in that role permanently better. This advantage compounds over time.
Dual-product generation
Every research panel run produces two outputs: a market research report and a DYNAMICS-tagged training dataset. Two revenue streams from a single operation.
Patent-protected
Four UK patent applications covering the core methods across all products. Competitors cannot replicate the approach without licensing. PCT international filing window extends to March 2027.

Every deployment makes every worker better

A four-stage improvement loop runs automatically for each worker role. Your corrections, your standards, your domain knowledge feed directly into the models that power every worker of that type.

01
Define Expertise
Role-specific knowledge and behavioural expectations are captured from domain experts and your own corrections.
02
Generate Training Data
High-quality training examples are produced from real interactions, quality-gated for accuracy, tone, and domain correctness.
03
Adapt the Model
Lightweight role-specific adapters are trained on this data. Each role gets its own fine-tuned expertise without rebuilding the base model.
04
Evaluate and Deploy
Automated evaluation tests personality consistency, domain accuracy, objection handling, and compliance before any adapter goes live.

See the technology in action

Deploy a full sales team and see how these systems work together to prospect, qualify, and close on your behalf.

Meet the sales team Investor overview