A working study in AI-supported judgment for consultants who run programmes

Your work, structured.
Your data, yours.

From Greek εὐβουλία — "good deliberation". A workspace where conversations become structured memory: risks classified, decisions versioned, stakeholders mapped, data requests documented. Chat with your programme — your programme listens back. Built in EU. Built and used in real programme work — shown here on the fictional HORIZON case.

See it in action → Follow the build

Working reference implementation · Built in EU · No training on your data · Not a product — a practice

5risk dimensions auto-scored
27structured AI tools
EUcompliance by default
0training on your data
The problem

Four AIs on your desk.
None of them remember your programme.

You use ChatGPT, Claude, Copilot, and Gemini every day. Each chat starts fresh. Each answer lives in a conversation you'll never find again. When something needs to land in the risk register, you copy-paste it yourself. You are the glue.

ChatGPT forgets the programme
You re-explain who Anna is, what the legacy system does, and why Q3 matters — every single chat. Memory features now exist, but the memory lives on Anthropic's / OpenAI's / Google's servers, not yours.
Claude / Copilot read but don't write back
They summarize the meeting transcript. They don't extract risks, classify them, link them to decisions, or version them when reality changes. The structured graph never builds up.
"Memory" features = vendor lock-in
Anthropic's "infinity memory" and OpenAI's persistent context live at the vendor. When prices change, when models drift, when a client says "no Anthropic" — your memory is held hostage.
Domain ignorance
A fat-tail risk, a reference-class forecast, a decision option, a stakeholder map — generic AIs don't know what these are as structures. They write them as paragraphs that scroll past in the next session.
The product — real screens

Not mockups. The working system.

A walkthrough of HORIZON — a fictional but fully worked engagement (PE-backed insurance core modernization, 420 MDKK, 36 months, 4 fat-tail risks). Every screen below is the actual product running on the HORIZON case. Nothing is staged beyond the case itself.

Eubulia document extraction: a Q3 board deck open beside 71 proposed entities — risks, concerns, decisions, goals and milestones — each classified and awaiting user approval
01 · From board pack to structure. One click on a real Q3 board deck proposes 71 entities — 13 risks, 7 concerns, 3 decisions, 5 goals, 11 milestones — classified by uncertainty type and fat-tail exposure. Nothing enters the register without your approval. Consent-first, always.
Eubulia risk register: 19 HORIZON risks with area, context, source document reference, dimension scoring, status and project columns
02 · A register that remembers where everything came from. Every risk carries its source document, its 5-dimension scoring profile and its status — versioned in a bitemporal graph. Ask "what did we know in November?" and the data model can answer.
Eubulia chat running a programme review of HORIZON next to the risk register: milestones, budget position and critical observations grounded in the register
03 · Chat with your programme — it answers from the register. "Run a programme review" returns the actual state: 7 of 11 milestones on plan, 12 MDKK reserve remaining, three concerns flagged with reasoning — every claim grounded in structured data, not a model's vague recollection.
Eubulia settings: provider-agnostic model choice with Claude, Mistral EU, OpenAI and Gemini, an allowlist, GDPR mode and web-search controls
04 · Your choice of model — EU when it matters. Claude, Mistral (EU), GPT or Gemini, interchangeable under one router. GDPR mode is a global lock: EU-only blocks non-EU providers automatically. Sovereignty as a setting, not a slide.
How Eubulia works

Three things nothing else does.

Not features. Capabilities that emerge from a structured database, a bitemporal data model, and a domain built specifically for management consulting.

01 / VERTICAL DEPTH
Frameworks applied, not described
Wardley Maps, SWOT, Engagement Strategy, Systems Lens, Three-Groups Benefits — Eubulia doesn't tell you about them. It runs them structurally on your workspace data and renders output as Strategic Wall tiles, canvas views, and exportable slides.
→ "Apply Wardley to the HORIZON programme" → live map, persisted, exportable
02 / DATA GRAPH
A relations graph that compounds
Every meeting, document, person, risk, decision, milestone — stored with foreign keys, not just embeddings. Bitemporal: every version of every fact is preserved with when we knew it and when it became true.
→ "What did we know about the vendor risk on Nov 15?" → exact answer
03 / DOMAIN INTELLIGENCE
Risk classification built for consultants
Two-layer risk model: worrying facts (Layer 1, observable) → risks (Layer 2, classified by uncertainty type, epistemic confidence, fat-tail flag, prerequisite flag). Auto-scored on 5 scenario dimensions inspired by Real Options.
→ Steering group asks: "Why is this fat-tail?" → trace it to the meeting, the speaker, the date
Capabilities
Eight things Eubulia does that nothing else does together.
01
Frameworks applied
Eubulia runs frameworks structurally on your workspace data — not as templates, not as descriptions. Output persists as Strategic Wall tiles, canvas views, and exportable slides.
WardleySWOTEngagement StrategyRepenningThree-Groups
02
Goal architecture
Goal narrative as anchor, milestones as time-bound commitments, benefits realization tracking — so what was promised in kick-off is auditable when steering group asks.
goalsmilestonesbenefits
03
Stakeholder graph
Auto-extracted from meeting transcripts. Who decides, who blocks, who's silent. Roles versioned across the engagement — when did Anna become "the bottleneck" versus "the champion"?
peoplerolesversioned
04
Risk topology
Two-layer model: worrying facts (observable) → risks (classified by uncertainty type, epistemic confidence, fat-tail flag, prerequisite flag). No premature classification.
5-dim scoringfat-tailprerequisite
Show 4 more capabilities
05
Scenario landscape
Compare decision options on 5 risk dimensions inspired by Real Options. Cascade detection between fat-tails. Board-ready in one click — branded PowerPoint, PDF, or HTML.
option scoringcascade detectionslide export
06
Hidden insight extraction
From transcripts, documents, conversations. Surfaces worries before they become crises — the off-hand remark in meeting 14 that contradicts the board narrative in week 22. Every insight traceable to source.
meetingscross-referenceprovenance
07
Challenge mode
Devil's advocate before steering group. Hidden assumptions, critical breakpoints, missing information, internal conflicts — surfaced from your own workspace history, not generic prompting.
self-critiquenarrative-testpre-mortem
08
Build your own skills
Skills are declarative Markdown workflows with YAML frontmatter (triggers, tools, context injection). Add new analytical lenses without writing code. Self-improving via embedded evaluation.
Markdown skillsYAML frontmattermarketplace (planned)
Comparison

vs. ChatGPT, Claude, Cowork, Notion+AI

An honest comparison. We left the cells blank where the competitor genuinely matches us.

Capability
ChatGPT
Claude / Cowork
Notion + AI
Eubulia
Memory across sessions
Per-chat
Per-project
Manual
✓ Bitemporal
"What did we know on Nov 15?"
✕
✕
✕
✓
Risk classification (5-dim scoring)
✕
✕
✕
✓
Frameworks applied to your data
Mentions
Knows them
Templates
✓ Structural
Stakeholder graph from meetings
✕
✕
Manual
✓ Auto
Branded board slides (one click)
✕
✕
✕
✓
EU data processing (default)
Optional
Optional
US
✓ EU default
Per-client hard-delete (GDPR Art. 17)
Generic
Generic
Manual
✓ One button
Built by a consultant
✕
✕
✕
✓
EU-hosted · no training on your data
✕ US
✕ US
✕ US
✓ EU-default
Your data lives in your own DB
✕ Vendor
✕ Vendor
✕ Vendor
✓ Your Postgres
Swap LLM provider freely
✕
✕
✕
✓ Any via LiteLLM
Trust

Built for consultants who handle real client data.

Compliance is a precondition for using AI in real programme work at all. Everything you see here runs on EU infrastructure with full data sovereignty — because responsible use isn't a feature, it's the starting point.

EU processing by default
EU infrastructure routing by default for tendering-bound work. Compliance-restricted clients get EU-only and Azure-only fallback automatically. No required US data transfer.
Multi-tenant isolation
Row-level security on 15 core tables. Per-organization isolation on 35 tables. Defense-in-depth, not a checkbox.
Per-client hard-delete
GDPR Article 17 implemented as one button per client. Cascading delete with audit trail. No "submit a ticket" theatre.
Markdown export, anytime
Every chat, every note exports as Obsidian-flavor Markdown. We don't believe in lock-in via format.
Zero training on your data
We do not train models on customer data. Period. If we ever build a domain-specialist model, it will be opt-in and anonymized.
Documents stay in your cloud
Your source files live in your own OneDrive, Google Drive, NextCloud, or local folders. Eubulia indexes the metadata, never copies the files.
Follow the build
Research & build notes · occasional · no spam
Eubulia is not a product you can sign up for — it's a working reference implementation I build and use in real programme work, shown publicly on the fictional HORIZON case. Leave your email to follow the research: what works, what breaks, and what AI-supported judgment actually looks like in practice.

  • Build notes — real decisions, real trade-offs, real dead ends from building it
  • The HORIZON case — a full fictional programme run through structured AI, in the open
  • Judgment, not hype — where AI helps programme leaders, and where it can't replace experience
  • Want it applied to your programme? That's a conversation, not a signup — get in touch
Frequently asked

The questions other consultants ask first.

Can I get access — is this a product?
No — and that's deliberate. Eubulia is a working reference implementation: I build it and use it in my own programme work, and show the results publicly on the fictional HORIZON case. It's research into AI-supported judgment, not a SaaS you can sign up for. Follow the build notes — or if you want the approach applied in your own programme, that's a conversation.
Which LLM does it use?
Your choice. Eubulia is provider-agnostic — Claude, GPT, Mistral, or Gemini, interchangeable under the hood. EU-only routing is available when a client requires it. When prices change or models improve, you switch models; your structured memory stays exactly the same.
Where does my data live, and is it used to train any model?
Your data lives in a bitemporal Postgres graph you can query, export, and audit — EU-hosted. It is never used to train any model. No training on your data, GDPR Article 17 hard-delete built in, Markdown export so there's no lock-in. Your source files stay in your own OneDrive / Google Drive / NextCloud — Eubulia indexes metadata, never copies the files.
How is this different from ChatGPT, Claude, Copilot, or Gemini with memory?
Those tools store memory at the vendor as loose notes on their servers, locked to their tools and pricing. Eubulia stores memory as structured, versioned entities — risks classified, decisions versioned, stakeholders mapped — in a graph you can query and export, portable across any LLM. You can ask "what did we know on November 15th?" and get an answer. That's the difference.
Can I extend it with my own skills and frameworks?
Yes — this is the deliberate edge. Skills are declarative Markdown workflows: an analytical playbook the engine runs against your structured data. Bundled today: document-review, project-intake, project-review, a six-skill context-harvest family, plus framework lenses (Wardley Maps, SWOT, Engagement Strategy, Repenning Systems Lens, Three-Groups Benefits). Your own judgment, encoded as a repeatable skill.
Why "EU by default" — does it actually matter?
For consulting in regulated sectors (healthcare, public, financial) it is increasingly a tendering requirement. Eubulia runs on EU infrastructure by default and can restrict the LLM router to EU-deployed models. Full audit trail, no required US data transfer.

Judgment can't be downloaded. It can be shown.

Eubulia is what AI-supported judgment looks like when someone with 25 years in steering groups builds it for their own practice. Follow the research and build notes — what works, what breaks, and what it means for how consulting changes.

Follow the build → See it in action

Built in EU · No training on your data · Not a product — a practice