Every week, I’m opening up the build of Arthur — what we’ve shipped, what we’ve fixed, and what we’ve learnt.
No theatre. No hype. Just the real work of turning Arthur into a mentor, co-founder and team founders can rely on.
If you’d like to get involved, sign up for early access or join the community.
This week was about something simple but hard:
taking all the “cool AI demos” and turning them into opinionated, reusable engines that investors would actually use.
Video pitches, pitch decks, admin controls, and Google Docs output all moved forward — and the architecture to sell this as a white-label product started to take shape.
Arthur’s world is getting sharper
Arthur is now moving beyond “smart chat” and into structured evaluation engines that can:
watch your pitch
read your deck
score what matters
surface the top things you must fix
then hand you straight into Arthur to actually do the work
Think: intake → diagnosis → “fix these now” → execution with Arthur.
✅ What we shipped
🎥 Danger Room video engine — first slice live
The first version of the 2-minute pitch analysis engine is now live as its own microservice.
Current flow
upload (or record) a pitch
store securely in S3
analyse via AWS Rekognition + Transcribe
pass signals into our agent stack mapped to the Fusion42 pitch framework
return a structured JSON dataset Arthur can use everywhere
Already extracting
who’s on screen and when
presence & framing consistency
movement & scene changes
full transcript with timestamps
delivery patterns (pace, pauses, fillers)
narrative beats mapped to investor logic
The transcript + delivery data now feed the content engine, so Arthur can connect how you pitch with what you said.
From this, Arthur gives structured pointers and the top 5 investor questions your pitch will trigger — a direct stress-test of your investor readiness.
Built to power
Danger Room → Pitch Kung Fu → FUSED
accelerators & incubators, VCs & CVCs
government programmes
external pitch competitions
Next
Turn these signals into an investor-grade pitch brief with a brutal fix-first list founders can iterate on session after session.
📊 Pitch deck engine — from prototype → product
The pitch deck analysis engine has moved from a cool prototype to a real product surface.
Arthur can now read a full deck and evaluate it like an early-stage investor — across four lenses and eleven signals.
Current flow
upload a deck (PDF/PPT)
extract structure and slide intent
map content to the Fusion42 pitch framework
score each signal
return a structured report with grades, fixes and next steps
Founders get
overall grade
signal-by-signal score
top three fixes
an explicit fix-first list
eight questions your deck will trigger in a meeting
and then we go through them with Arthur and fix them
Why it matters
Founders rarely know what investors are actually scoring when they look at a deck. PitchDeck Intel makes the scoring explicit — and gives a path to close the gaps fast.
Built to power
Danger Room → Pitch Kung Fu → FUSED
accelerator intake
cohort prep
investor readiness flows
Next
Link deck + video into one readiness score — content + delivery + presence.
📄 Google Docs report generation — from static → living docs
The report generation flow is now wired end-to-end in the prototype.
Today:
one-page documents generated from a request
returned as editable docs
OAuth integration into Google Docs is in progress so that founders get live, editable documents in their own workspace, ready for comments and track-changes.
Once stable, the same pattern applies to:
Google Sheets → KPIs, funnels, financials
Google Slides → pitch revisions, investor updates
The bridge between Arthur’s structured analysis and where founders actually work.
🌐 Website analysis — engine in progress
The next microservice on the bench is the website engine.
Flow
founder drops in a URL
engine runs deep research on the business behind it for sales signals + customer intel
pulls this into the same pipeline as video + deck
optional SEO + Generative Engine Optimisation (GEO) checks
Output is written for founders — “fix your positioning”, not a dry technical SEO audit.
Longer-term: a simple “Is your GEO working?” check-up any startup or agency can plug into.
🧩 Jiva integration — wiring in the SDK
Alongside the microservices we’ve been wiring Arthur into the Jiva SDK — and it’s worth saying why.
Jiva’s core IP is a “fusion” model architecture — originally built for fusing multiple medical models into a single agent that can reason over messy, multi-source data and still give one clean answer.
We’re using the same idea for founders.
Arthur is fusing:
video signals (delivery, presence, narrative)
pitch deck logic
website + GEO + sales signalslater: product, financial and traction data
Instead of bolting together 20 brittle prompts, Jiva gives us:
a structured agent stack
consistent schemas for plans, briefs and reports
routing + cost control so Fuel Units map to real inference usage
We are using Jiva.ai’s Fusion technology to combine multiple domain-specific models into a single fusion agent capable of multi-source inputs and a unified prediction for startup readiness — aligning with US20240071062 (Model Fusion System).
In short
Arthur isn’t a string of oversized prompts taped together — it’s a fused network of specialist agents.
A small workforce of agents, each running the best model for its job, all cross-communicating to decide what matters most.
🚧 What’s next
Next sprint is about tightening the loop between founder → signal → asset.
Near-term track:
plug Transcribe into the Danger Room engine
extend the Website engine so one URL feeds positioning + GEO into the same stack
finish OAuth + dynamic Google Docs storage
wire the founder account area into these microservices
move Stripe from “planning” to live billing
design the dual admin model:
• Fusion42 control plane (rules, prompts, cohorts, global data)
• Partner / cohort admin for white-label customers
In parallel, we’ll start mapping how these engines roll up into one investor-grade view of a startup: pitch video, deck, website and traction all feeding the same agent stack.
We’ve gone from skateboards to Ferraris in how fast you can build with AI.
The hard part now is making sure Arthur isn’t just fast, but trustworthy, reusable and investor-grade — and that’s the line we’re holding the build to.
FOR THE ❤️ OF STARTUPS
Help steer what we build next
I’m not building Arthur in a vacuum.
If there’s a workflow, bottleneck or ugly bit of founder work you want Arthur to take off your plate, tell me.
What you’re trying to do
What slows you down
What “done” would look like if Arthur nailed it
These go straight into the build plan for the next sprints.