Client work · Germany · 2026 — now

Daphne

Bring clarity to long-form conversations.

A cognitive companion that turns documents, recordings and meetings into a reasoning space you can inspect, question and revise.

Human AI

(The brief)

Strategy teams drown in PDFs, recordings and meeting transcripts. They don't need another summary. They need to see the evidence, the risks, the options, the gaps and the tensions, and stay in the driver's seat while an AI helps them think.

01 · Messy sources in

Everything the team already has.

02 · A reasoning space out

Evidence, risks, options, gaps and tensions.

Strategy_2030.pdf

84 pages

Board meeting

54 min recording

Market scan

URL

Expert interview

Transcript

Q3 review deck

32 slides

Regulation notes.pdf

12 pages

Evidence

Risk: supplier lock-in

Option: phased rollout

Gap: no customer data

Tension: speed vs. compliance

(The product)

The Cognitive Workbench

Pick a job: Challenge a strategy before committing, Map sources into risks and options, or Build. An intake agent asks clarifying questions with real options, then the artifact streams onto the canvas and every version is kept.

Workbench · Market entry DACH · v3Saved

Evidence · Daphne

DACH SMEs cite onboarding cost

Option · you

Phased rollout via partners

Risk · Daphne

Supplier lock-in

Gap · Daphne

No churn data for Austria

Tension · you

Speed vs. compliance

Ask Daphne

Where does the phased rollout conflict with compliance?

Two places: partner data sharing in phase one, and the audit trail the board asked for. Both come from the regulation notes, pages 4 and 9.

Daphne can make mistakes. Check important info.

(Live builds)

Artifacts stream in, one operation at a time.

POST /api/sessions/[id]/builds · NDJSON

{"type":"build_started","artifact":"reasoning_map"}

{"type":"baseline","version":3,"nodes":30}

{"type":"thinking","note":"Group risks by stakeholder"}

{"type":"operation","op":"add","node":"Tension: speed vs. compliance"}

{"type":"operation","op":"update","node":"Risk: supplier lock-in"}

{"type":"operation","op":"add","node":"Gap: churn data, Austria"}

{"type":"operation","op":"remove","node":"Duplicate: cost overrun"}

{"type":"saved","version":4,"nodes":31}

Every version kept

v1v2v3v4

(The bug that shaped it)

A 30-node map came back with 4.

One routine update quietly dropped 26 nodes nobody asked to remove. The model had rewritten the whole map and forgotten most of it.

(The fix)

Deltas only. And a gate that refuses to save.

Revisions now start from the saved version and the model only emits changes: add, update, remove. If a build loses nodes it wasn't told to remove, it fails loudly instead of saving.

v3 → v430 → 4 nodes30 → 31 nodes ✓

(Before → after)

Faster, and right more often.

I replaced a 12-node LangGraph pipeline with one native tool loop on gpt-5.4, moved builds into the server process, and rewired retrieval. Then measured it against a 30-case eval that runs in CI.

Eval pass rate (30 cases)

13/30

30/30

Median agent latency

32.3s

4.8s

Build path

223s

105s

Upload to chat-ready

~19s

~8s

Retrieval

8.9s

1.2s

(Six ways to think)

Six artifacts you can build.

Each one is a method pack: its own builder, its own quality gate, and its own view. Eleven more are designed and waiting.

Reasoning map

Evidence, risks, options, gaps and tensions, linked.

3D knowledge space

Organic clusters, bridges and tensions in three.js.

Flow map

How a decision moves through people and time.

Slide deck

Present mode, embedded artifacts, generated illustrations.

Robustness check

One strategy against several futures at once.

QMA

Quadruple materiality: outside-in and inside-out, risks and opportunities.

(Voice)

It joins the meeting. Visibly.

A Recall.ai bot joins Teams, Zoom or Meet as “Daphne Notes (recording)”, because in Germany recording has to be obvious. Speakers can be renamed and the transcript lands in the library. Then you can talk to the map itself: Ask Daphne runs on OpenAI Realtime and re-grounds whenever you switch view.

Mia

Anna

Jonas

Daphne Notes (recording)

Anna If we go through partners, who owns the customer data?

Jonas That's the tension. Put it on the map.

Daphne Added: Tension · data ownership vs. partner speed.