Geometry locked in closed files
Parametric intelligence sits inside models that only open one project at a time. The solution to your next problem is in a file no one will think to open.
Sovereign knowledge infrastructure for enterprise AECPatent pending · USPTO App. 30/056,415
Your data is your future AI.
Tomorrow's answers will be found in data you start accumulating todayThis is what Nexus Hestia is here for.
Sovereign knowledge infrastructure: recover design precedent, eliminate re-work, preserve IP without cloud vendor lock-in.
Step into your unfair advantage
PROLOGUE
You signed up to build, not to triage a Tuesday fire drill
08:47 — Tuesday morningYour buried work isn't history. It's your live unfair advantage.
1992 · archived projects2022 · the unlocked answer2026 · today
1992 — Sovereign SecurityHover the ticks to explore
1992 Project: US Embassy in Egypt — Sovereign Security & Structural Specification [Cairo, Egypt]
Archive timeline · 1992 to 2026
01 — What is being lost
Your firm already knows the answer.
It's in a project you finished four years ago. Every design office sits on a goldmine of past drawings, models and details, but it lives in a digital graveyard of folders, old hard drives and personal laptops — and it leaves the building when the people who made it do.
11.5hevery week
That is the average time construction managers and executives spend researching and analyzing data. The report finds much of that time is spent inefficiently. Autodesk's summary puts BIM/VDC and preconstruction roles higher still: 12.3 and 13.4 hours.
Deloitte Access Economics for Autodesk, The State of Data Capabilities in Construction, 2023 — a survey of 1,275 construction leaders across 12 countries.
Photos no one can search
Thousands of site photos, renders and drone captures per project, with no meaning attached to any of them beyond a filename.
Knowledge that retires
When a senior director leaves, the reasoning behind every detail leaves with them. What remains is the drawing, without the why.
When your most experienced architect leaves, do they take 15 years of project knowledge?
You have thousands of photos from every project. None are searchable.
What sits unread in a typical practice
◇BIM files
▤Drawing archives
◉Site photos & renderings
▧Specs, RFIs & contracts
⬡Material libraries
◈Design decisions in email
02 — The turn
Two architectures for enterprise AI
Public AI tools can't help, because no firm can risk exposing confidential blueprints and proprietary design secrets to an outside company. The difference isn't features — it's where your data's boundary sits. Toggle between the two.
03 — How the memory works
A permanent vault, and a private apprentice.
Explained without the jargon. Six ideas, in the order a practice meets them.
A goldmine in a digital graveyard.
Decades of drawings, models and details are scattered across folders, old drives and personal laptops. Public AI tools can't touch them, because no firm can risk exposing confidential blueprints and proprietary design secrets to an outside tech company.
Nothing is ever overwritten or lost.
The smart data lake is the firm's indestructible collective memory: every drawing, detail and note recorded in an append-only ledger, so nothing is ever accidentally overwritten or lost.
Not a public chatbot.
Each firm gets its own compact, private AI assistant that lives inside the firm's own setup, learns how its people draw and organize work, and keeps every trade secret in-house.
Say what you remember. Get the detail.
Plain language in, the exact element out.
“Show me the cantilever detail we engineered for the 2022 mixed-use project that passed city inspection.”
A hardened enclave, not a server closet.
Architects design spaces; they shouldn't have to run IT departments. A security-hardened, isolated enclave gives firms strong protection and access across offices or home setups, without buying servers or managing hardware.
Grows with your tools. No rip-and-replace.
Because the archive itself is clean and permanent, the firm owns its foundation forever. When a better AI model arrives years from now, it plugs in as a new adapter, with nothing to migrate or put at risk.
Phase 1 · Unify
From chaos…
Drawings, models, photos, contracts and correspondence are brought into one governed archive. Integration and governance come first, because intelligence is only as good as its memory.
Phase 2 · Activate
…to clarity.
A private assistant, tuned to your practice, searches by shape, by narrative and by image, and reasons over your own standards, your own details and your own past decisions.
Your unified institutional memory
step 1 of 4Interactive pipeline
Select any block to see how data moves through your sovereign infrastructure. On a pointer device, hover a source to watch it flow in.
Reading RVT-0442 · East Wing structural set …
04 — Under the hood
The Sovereign Stack
Nexus Hestia develops sovereign, anti-fragile data infrastructure that transforms inert enterprise archives into living, adaptive intelligence: an append-only, immutable smart data lake paired with industry-specific adapters, and hyper-localized small language models tuned by local reinforcement learning. Four layers, bottom to top — open one to see the mechanism, not just the name.
-
The underlying data lake is append-only, preserving an uncorrupted ground truth of all organizational drawings, spatial models, contracts, sensory data and documentation. Nothing already written is ever edited or deleted — only added to.
Simulate an ingest — watch the ledger, not the past, change
2024-03-11 · RVT-0442 · East Wing structural set2024-07-02 · CTR-1180 · Steel supply contract2025-01-19 · IMG-3391 · Site walkthrough capture
Every ingest, merge, model call and human decision, source-tagged. Nothing is deleted; superseded entries are archived, not overwritten. (Sample data.) -
Tailored adapters sit between raw data streams and active workflows, natively embedding specialized SLMs directly into client infrastructure to ensure complete IP isolation and zero external leakage.
Raw streams never reach a workflow unadapted
BIM / RevitContractsSensor dataADAPTERisolation boundaryDesign reviewRFI responseField lookup -
SLMs are positioned locally, operating on a lightweight compute footprint while using an on-device reinforcement learning layer that continuously adapts to individual user behavior, spatial workflows and firm conventions.
Simulate usage — the model tunes to this firm, on this device
fit to firm: 22% -
Instead of static file storage, documents and spatial models become interactive, continuously indexed assets that can be queried, executed and referenced in real time.
Try a query against the live index (illustrative)
—
Four message pillars
- IP Preservation — institutional amnesia and re-work cost 10–15% of project value.
- Sovereignty as non-negotiable — "Your knowledge. Your infrastructure. Your control."
- Domain-specific intelligence over generic LLMs — generic models "hallucinate over technical data" and lack geometric reasoning.
- Architectural extensibility — the adapter model lets firms upgrade AI models later without migrating or exposing their core archive.
05 — Proof
Shown, not claimed.
Two ways to ask your archive a question, and one layer that watches over both. These are the working interfaces — hover an image to inspect it.
Currently shown, not claimed. This is where we start. Each product here is an adapter on the same sovereign archive, and we will build more.
Find the detail, not the filename.
Local vision models index every image, rendering and site record behind your firewall. Type what you remember — “hotel”, a material, half a word — and the archive answers with meaning, not with filenames.
- Every image read where it sits. Nothing is exported to a cloud.
- Your own taxonomy — Interior, Exterior, Site — not a generic one.
- A human stays in the loop: edit any description and the memory learns your wording.
Product screenshots · sample archive
Ask in plain language. Get back the exact Revit element.
Native Revit and IFC in one catalogue. Geometric, narrative and visual signals are read together and fused into one ranked answer, so you can find a detail by its shape, its story or its look — not by guessing the file name.
- Topology parsed from the model: project, storey, room, element.
- Room-bounding and element parameters searchable in plain words.
- Results link back to the original element, not a screenshot of it.
Try a question (illustrative)
—
BIM graph explorer — five stages from one model
Product screenshots · sample model data
The layer above both searches.
In active deployment · Patent pendingAthena minds the timeline, draws your attention, and remembers on your behalf. She sits above search: managing timelines, alerting leadership, and linking project insights to each other automatically.
- Search answers a question you thought to ask. Athena notices the ones you didn't.
- Runs inside the same sovereign boundary as the archive it reads.
Illustrative — what Athena surfaces on a Tuesday
Similar conflict solved before
The transfer-beam pour sequence on this project matches a podium slab you resolved in 2021.
08:52 · linked to RFI threadMilestone drifting
Curtain wall submittal is two weeks behind the pattern of the last three comparable projects.
09:15 · flagged for the project directorKnowledge at risk
Three details on this job were resolved by one person who is leaving in June. Capture the reasoning now?
10:03 · offered to the practice lead
Illustrative cards left · photograph of the working interface right
06 — One Tuesday, one question
The contractor is asking about a slab you drew in March.
Scroll. The whole exchange takes a few seconds.
11:52 · The ask
From: site superintendent
Subject: Podium slab — pour sequence at the transfer beams
Subject: Podium slab — pour sequence at the transfer beams
“Level 3 podium pour sequence is clashing with the transfer beams at gridline C. Did we solve this on anything before? I need to confirm with the crew by 1pm.”
11:52 · Now you have ATHENA ask her. Start with whatever you remember
podium slab pour sequence conflict at transfer beams
11:53 · The memory answers
▲ THIS PROJECT · 2026Roof slab assembly & parapet detail
Related assembly, not a sequencing precedent.
▼ SOLVED 2021 · LEVEL 2 PODIUMPodium slab sequencing conflict — transfer beam pour order
Resolution signed off by structure. RFI thread, revised sequence drawing and the original element, all linked.
11:53 · Read it across
Drag the cut line. Left of it, today's drawing; right of it, the 2021 resolution. The sequence that worked is the one that moved the joint off the beam.
2026 · this project
2021 · resolved
Schematic. Not to scale.
11:53 · Reply sent
Nothing was retyped.
Nothing left the building.
Four other things needed you before lunch. This is what it feels like when your firm remembers.
07 — Sovereignty
Your knowledge. Your infrastructure. Your control.
Total data sovereignty
Your archive is read on infrastructure you control. Your work does not become anyone's training data, and the boundary the model operates inside is one you set.
Memory architecture
Multi-decade archives are normalized into one governed index that the practice owns permanently — drawings, models, images, correspondence and tacit knowledge in a single system, not five separate ones.
No cloud lock-in
You retain ownership of the index, the models and the infrastructure. If we ever part ways, every piece of intelligence stays with your practice.
Sovereign deployment
Nexus Hestia runs inside a boundary your firm controls: a security-hardened, isolated enclave dedicated to your practice. Adoption begins with one project archive. No reorganization of twenty years of legacy systems is required.
08 — Why this is hard to copy
Domain-specific by design.
Generic models hallucinate over technical data and have no sense of geometry. Nexus Hestia is built by architects, for architects, and it shows in what it isn't.
Not a public chatbot
Your own private assistant
Each firm gets its own compact model that never sends blueprints outside the firm and learns only your conventions.
Not a filename hunt
Finds by shape and intent
Geometric search recovers design intent from the model itself — by shape, not by guessing the file name.
Not a walled garden
Independent of any one ecosystem
Works across your tools and grows with them. Customers retain ownership of all infrastructure and intelligence.
01
We speak architect
SCI-Arc training means fluency in Revit families, spec sections and design culture.
02
We understand affective systems
USC affective computing shapes AI around how designers think, not how engineers assume they do.
03
We teach it
Ties to Southern California architecture schools bring both talent and a route to the next generation of practice.
04
We have working products
Deployed and running today. Not a future promise.
SCI-Arc training + USC Affective Computing + proven AI leadership = an intersection no competitor can quickly replicate.
09 — Why now, and where we are
The market has caught up with the problem.
- Open data formats have stabilized, so an archive no longer has to live inside one vendor's system.
- The industry recognizes the problem, and no dominant answer exists.
What the industry captures, and what it uses
The average construction business captures 11 of 13 types of data. It uses 3 to make decisions.
- 13types of data a business can hold
- 11captured
- 8captured and analysed
- 3used to make decisions
Share of businesses, by type of data
- Captured, analysed and used for decisions
- Captured and analysed
- Not captured, or captured but not analysed
- Geographic Information System 45%
- Change orders and RFIs 44%
- Machine and equipment data 42%
- Communication data 42%
- Permits and regulatory data 40%
- Supplier and subcontractor data 37%
- Client information 37%
- Project progress data 36%
- Safety and incident data 36%
- Building materials data 34%
- Project information 34%
- Labour data 34%
- Financial data 28%
Sorted by the share left unused, highest first. Change orders and RFIs: 44% of businesses do not capture them, or capture them without analysing them. The records exist somewhere in the practice; what is missing is a memory that reads them.
View as a table
| Type of data | Used for decisions | Captured and analysed | Not captured, or not analysed |
|---|---|---|---|
| Geographic Information System | 21% | 34% | 45% |
| Change orders and RFIs | 19% | 37% | 44% |
| Machine and equipment data | 21% | 37% | 42% |
| Communication data | 21% | 37% | 42% |
| Permits and regulatory data | 26% | 34% | 40% |
| Supplier and subcontractor data | 26% | 37% | 37% |
| Client information | 24% | 38% | 37% |
| Project progress data | 27% | 37% | 36% |
| Safety and incident data | 25% | 39% | 36% |
| Building materials data | 30% | 36% | 34% |
| Project information | 27% | 39% | 34% |
| Labour data | 26% | 40% | 34% |
| Financial data | 33% | 39% | 28% |
Source: Deloitte Access Economics for Autodesk, The State of Data Capabilities in Construction (2023), Deloitte Construction Data Survey of 1,275 construction leaders, p. 13. Figures as reported; rows may not sum to 100% because of rounding.
Where we are
Working, deploying, next.
An honest status of the story so far.
Working today
- Image Index & SearchRunning against a real archive, entirely inside the firm's boundary.
- BIM Index & SearchReading native .RVT and .IFC in one catalogue, with topology parsed from the model.
In active deployment
- AthenaThe intelligence layer above both searches. Patent pending.
Next
- First project archivesOne archive per practice, one folder, one model. No reorganization of legacy systems.
- Assistants tuned to each practiceLocal reinforcement learning that adapts to a firm's own conventions, inside its own enclave.
AS OF SEPTEMBER 2026
10 — Who is behind it
Built by practicing architects and AI engineers.
We solve institutional knowledge: decades inside architecture practice, and the engineering to turn that insight into a system that works in production.
Natasha Bajc
Trained in architecture, urbanism and computer science, Natasha has directed design and technology teams across three continents. She sets the direction: an architect's accumulated knowledge becomes an active, sovereign partner in the work.
M.Arch Algorithmic Design · SCI-Arc
M.Eng Architecture & Urbanism · Univ. of Belgrade
Affective Computing · USC
Exhibited: Venice Architecture Biennale · Beijing International Art Biennale
Jovan Njegić
Jovan turns the idea into a system that works in production, and proves it with measurements. He has built private language-model systems for organizations in international policy and diplomacy, and has taught data science and AI for almost two decades.
PhD Statistics & Data Science
Master's in Finance
Founded the Centre for AI, Business School of Novi Sad
Evan Cerasoli
Evan turns the idea into a story the market understands. With over thirty years in marketing across media, art, and technology, he spent thirteen years as a principal strategist and producer behind Toyota's Scion lifestyle marketing, and led the art and curation for it.
B.S. Telecommunications & Film
Brand campaigns for: Nike, Google, Adobe, Microsoft, LVMH, Adidas, Xbox, HBO
Founder, Kerosene · media & culture consultancy
12:00 — Before lunch
Bring one project. We'll show you what's inside it.
The consultation
A consultation takes 90 minutes.
We will evaluate your practice's archive readiness and outline clear deployment steps. Bring one archive, one folder, one model — your own.