About
NODAVX is the atlas; SPECTRA Research is the workshop
NODAVX is a spatial intelligence platform built by SPECTRA Research Pte Ltd in Singapore. It exists because operational dashboards keep failing the same way: they display a number nobody can question, and the questioning happens afterwards, in a spreadsheet, by someone who no longer trusts the dashboard.
The name
NODAVX, spelled out
A name you can take apart is a name you can hold accountable. Each part is a layer that has to work on its own.
Graph entities and vertices
Properties, labels and ontology
Live 3D Atlas and spatial representations
Agentic AI, graph reasoning and digital twins
Working names inside the codebase are older than this domain: the npm scope is @atlas/*, DOM events are atlas:*, and the database namespace is spectra / database atlas. They were left in place deliberately — renaming identifiers is how a tested system quietly loses its tests.
Position
Local-first, on purpose, and it costs something
Choosing that the whole system runs on one machine is a product decision with a trade-off attached, and the trade-off is the part most vendors leave out of the sentence.
What it buys
The reason to accept the constraint
Data never has to leave the building to be useful, so regulated or classified material can be modelled at all. Inference is a local model on the same machine — the published figure for cloud calls in this build is 0, and the offline test cold-boots the whole application with the network blocked.
What it costs
What we will not pretend
Throughput is one machine’s throughput. There is no horizontal scale, no multi-region failover, and no answer yet about what happens at a hundred concurrent analysts. Scaling out would mean introducing a cache, and a cache is the thing this architecture refuses.
Who it is for
And who should look elsewhere
A team that owns one hard operational question and can name the decision it makes weekly from it. If the requirement is a managed BI tool over every system in the company, a conventional warehouse stack is the honest answer today, not this.
Provenance
Where every number on this site comes from
A build script reads the running store and the validation documents and writes the figures into these pages. Nobody typed a count into a marketing sentence.
If the store is unreachable, or if the two count paths disagree, or if a document stops stating a figure in the expected form, the build stops with an error instead of publishing yesterday’s number. An out-of-date honest site still has to say so.
This site makes zero requests outside nodavx.com: fonts, icons, imagery and the search index are all served from this origin. Search is a build-time index of these seven pages, so there is no backend to query and nothing to phone home to.
Not proven
Six things this build does not claim
Read this section before the performance table. It is the same list we would want if we were buying.
One Mac, not a fleet
Every latency, frame-rate and throughput figure on this site was measured on a single Apple Silicon Mac with the store, API, inference runtime and browser on it. Nothing has been measured across a network boundary or under a concurrent load test with many users.
Single-user loopback auth
Sessions are a signed cookie issued to one local user. Role enforcement and cross-workspace scope refusal are implemented and asserted by tests, but there is no identity provider, no team onboarding and no audit-workflow product.
Deterministic synthetic data
Both packs are seeded from ATLAS_SEED=20261016. That is what makes the numbers reproducible, and it is also why no figure here is a claim about your business. Entity counts are honest; the vessels are not real vessels.
Answer quality is not graded
Citations are counted and latency is measured, and the copilot refuses to render output it cannot parse. There is no eval set scoring whether an answer was good, and vector search is off until an embedding model is verified locally.
No live telemetry
Positions are asserted in the seed, not received. There is no AIS receiver, ERP connector, historian feed or message bus ingest in this build — arrivals enter through a validated ingest path that a person or script calls.
SurrealDB licence terms
The server is BUSL-1.1 and every other dependency is MIT or Apache-2.0, read from installed manifests rather than websites. Before commercial deployment at scale, confirm the current SurrealDB terms independently rather than trusting this sentence.
Contact
One address, and it is monitored
No form, no CRM, no newsletter. Mail reaches a person at SPECTRA Research.
[email protected]
demos, pilots, technical questions
Tell us the decision and the tables you have today. A CSV is enough to start; an ERP extract is better. If you want a pilot, we will ask for one decision, three tables, and the formula you currently compute by hand.
How this site is served
so you can check it yourself
A static origin on the same Mac as the atlas, fronted by a Cloudflare Tunnel — no web server exposed to the internet, no CDN copy of your data, no third-party analytics. app.nodavx.com reaches the application on the same machine, and visitors arriving over the public URL are read-only by design.
apex → localhost:8901 · app → localhost:8897 · origin HTTP, TLS at the edge