Today I get to share something the team has been quietly proud of for a while: Concorde Systems is officially a member of NVIDIA Inception, NVIDIA's global program for startups building the next generation of AI.
For a company whose whole thesis is "your plant's knowledge should live in one graph — wherever your data is allowed to live", this matters far beyond the badge. Here's what it changes under the hood, and what it means if you run a plant.
Why Inception, and why now
Heavy industry has a constraint most AI startups never face: the data cannot leave. SCADA exports, P&IDs, batch records, downtime logs — for many of our customers these are regulated, contractual, or simply too sensitive to ship to someone else's cloud.
So from day one we built FOG, our deployment layer, around a simple promise: cloud, private cloud, on-premises, or air-gapped edge — same platform, same graph. The hard part of that promise is inference. Running document understanding and GraphRAG at plant scale, locally, needs serious acceleration.
That's exactly what the Inception partnership gives us structured access to: NVIDIA's accelerated-computing stack, engineering resources, and validated deployment patterns for enterprise AI.
The stack behind DECODER and ASSIST
A quick tour of the infrastructure we run today, and where NVIDIA fits:
- Ingestion (DECODER). PDF manuals, Excel logs, CAD drawings, SCADA exports, and site photos go through a GPU-accelerated parsing and extraction pipeline. Vision-language models handle scanned documents and drawings; table and layout models recover structure that plain OCR destroys.
- Inference serving. We package our models as NVIDIA NIM microservices — containerized, GPU-optimized inference for the LLMs, VLMs, and embedding models the platform uses. The same containers run in our managed cloud and inside a customer's data center, which is what keeps FOG's "same platform everywhere" promise honest.
- Retrieval & embeddings. Multimodal extraction and embedding runs on NVIDIA NeMo Retriever microservices, feeding the knowledge graph that ASSIST traverses. Answers stay grounded because retrieval — not free generation — is the gate: if the graph returns nothing, ASSIST refuses.
- The graph itself. Entities, equipment, procedures, sensors, and documents live in a property graph, with live metrics streaming in via TimescaleDB from plant historians and IoT gateways. HORIZON's roll-ups and forecasts read from the same graph, so the numbers your board sees trace back to the same sources your technicians cite.
- Deployment (FOG). Kubernetes-orchestrated, with connectors, scheduling, observability, and error handling in one layer — deployable to a VPC, bare metal, or a fully air-gapped rack next to the line it serves.
What this means if you run a plant
Three practical consequences:
- On-prem AI is now a first-class option, not a compromise. The same GraphRAG answers, citations included, without a single byte leaving your network.
- Faster pilots. GPU-accelerated ingestion means the "your files become a live graph" moment in our rollout happens in the first sessions, on your own documents — not after months of integration.
- A roadmap with headroom. As NVIDIA's inference stack improves, DECODER's ingestion and ASSIST's traversal get faster on the same hardware — including the constrained edge boxes that live on factory floors.
We didn't join Inception to put a logo on a slide. We joined because our customers' most valuable knowledge is trapped in formats and firewalls that ordinary AI can't reach — and this is the stack that reaches it.
What's next
Over the coming quarters we'll publish deeper engineering notes: benchmarks from our ingestion pipeline, how we schedule mixed LLM/VLM workloads on shared GPUs at the edge, and what air-gapped GraphRAG looks like in production.
If you want to see the stack against your own data — a real P&ID, a real downtime log — that's exactly what our 90-day pilot is for. Book a demo and we'll bring the graph to you.
Andreas Krisdianto is Co-Founder & CEO of Concorde Systems (PT. Maestro Imperium Teknologi), building operational intelligence for heavy industry from Bandung, Indonesia.