Research has depth. A meeting has twenty minutes. Give this a customer name and a website and it researches the account once — the dossier, the buying committee, the commercial diagnosis — and keeps it. Every deck you then build for them is briefed in two fields and argues from that same record: a real, editable PowerPoint with speaker notes, photography from the customer’s own published pages, and every claim traceable to a source. Then you revise it by talking to it.
The pipeline
Stages declare what they depend on, not where they sit, so the engine runs everything it can at the same time. Each one has a schema, a cache key, a retry policy and a progress stream — which is why an interrupted run resumes instead of starting over, and why re-running one stage invalidates its dependents and nothing else.
20 of 20 stages
Deck ready
Every stage checkpointed, every claim sourced.
An illustrative run. Pacing follows each stage’s declared cost weight, and the ring is the same runProgress() the live generation screen calls.
No shared inputs — the engine starts all three at once.
The Director’s gate. Nothing downstream may contradict the blueprint.
Copy and visual direction run concurrently across every slide.
What it actually does
None of this is “AI writes your slides”. Each of these is a constraint the system enforces on itself — and each one is why the output survives contact with a real meeting.
Company, leadership, news, initiatives, tech stack, stakeholders, likely objections and competitors — gathered before a word is written, and cited. The Content Writer may only assert what research or your own collateral supports, and it flags anything it had to reach for.
Citations travel with the slide plan · unsupported claims surface in QA
The Presentation Director fixes audience, story framework, emotional beat, rhythm and call to action first. Every downstream agent is bound by it — which is the difference between a deck and a pile of slides.
Nothing downstream may contradict the blueprint
Canonical JSON is the source of truth and PowerPoint is emitted from it — with speaker notes, on-brand typography and charts composed from native shapes, so every bar stays selectable and recolourable in PowerPoint.
Deterministic layout on a 13.333 × 7.5in canvas — identical every run
“Make this slide more executive.” “Add ROI statistics.” “Use the same style as slide 5.” The Director classifies the request, runs only the agents it actually touches, and emits object-level patches — validated against their operation and rejected with a reason rather than half-applied. The rest of the deck is left alone.
Route → patch → re-render one thumbnail · never a regeneration
Each slide carries a revision counter, and each rendered image records the revision it came from. When the two diverge the viewer draws the slide itself instead of showing you a stale picture.
One edit, one 640px render — not twenty
Upload decks, PDFs, case studies and spreadsheets. The Knowledge Agent extracts services, messaging, statistics, success stories and differentiators, indexes them, and retrieval ranks them above anything scraped from the web — because a customer case study is ground truth and a web page is a hypothesis.
Retrieval records the authority boost it applied, so the ranking is visible
Imagery is sourced before it is synthesised — the brand’s own product shots, then pictures harvested from the customer’s published pages, each captioned with where it came from. Only illustrative images are ever generated, and never a map or a chart: a fabricated fact is worse than a missing picture.
Harvested images carry their source hostname in the caption
The finished deck is read back as pixels — every slide, plus a contact sheet of the whole deck — by a quality review and a ten-expert panel. What a machine can settle is settled: word budgets, contrast, duplicate headlines, missing imagery, unfilled placeholders. A pass that makes the deck worse is discarded and the previous version restored, so what reaches you is judgement calls, never chores.
Deterministic findings are re-measured after the fix, not taken on trust
The agents
Not one prompt wearing eleven hats. Each agent has a brief, a schema it must satisfy and a stage it owns — and each is routed to the model its work deserves, so strategy gets depth and per-slide copy gets speed.
Agent 01
The Blueprint: purpose, audience, story framework, emotional journey, rhythm and call to action. Nothing else starts until it exists. It also routes every revision request and keeps the presentation’s memory, so the same correction is never needed twice.
Agent 02
Company, leadership, news, initiatives, tech stack, digital maturity, stakeholders, objections and competitors — with citations.
Agent 03
Extracts a design system from a live website: colours by role, typography with PowerPoint-safe fallbacks, logo usage, voice and layout density.
Agent 04
Reads uploaded PPTX, PDF, DOCX and XLSX and extracts services, messaging, statistics, success stories, quotes and differentiators.
Agent 05
Pain points with a cost of inaction, the value proposition, differentiators with proof, objection handling and quantified outcomes.
Agent 06
The outline, then a plan per slide: purpose, key message, audience, layout, visual type, word budget and the facts it may assert. No copy.
Agent 07
Headlines, callouts, statistics, quotes and speaker notes at consulting-grade concision — and it flags anything it had to assert without support.
Agent 08
Chooses the diagram, chart, photograph, table or timeline and states why. Never invents chart data.
Agent 09
Turns plan, copy and visual direction into canonical presentation JSON through the deterministic layout engine, then emits the .pptx and the thumbnails.
Agent 10
Duplication, weak messaging, overloaded slides, branding drift, contrast, alignment, missing assets, flow and unsupported claims — model review plus deterministic checks.
Agent 11
Emits object-level patches — never a regenerated deck — and applies the auto-fixable findings QA raised.
Where the rail meets the road
Everything runs end to end on a deterministic offline provider — no key, no network, no cost — so you can see the whole machine work before you point it at a real model. Switch one environment variable when you want the real thing.
Built at Glīd Technologies — veteran-owned. Cleaner, safer and smarter freight systems, and the tooling that sells them.
Road \ Rail Synchronized