Q4 2026 SUMMIT CIRCUIT 9 cities · 40 seats each · San Francisco opens Oct 31 View summits →
Stage-X Space Open a Dispatch
OPEN SOURCE HARNESS · MIT & APACHE 2.0

Forward deploy any frontier model into any operator.

Bridging research labs and operational reality. We embed, integrate and optimize frontier AI systems inside the infrastructure you already run — air-gapped, on-prem or in your own VPC.

Open a Dispatch View the runtime
$ pip install stagex-runtime▌
FIG.01 · DEPOT → INFERENCE PATH LIVE
AIR-GAP PERIMETER · EGRESS drop_all DEPOTisolated · no WAN INGESTread-only OPC-UA HYBRID INDEXBM25 + HNSW vLLM / FP8on-prem inference VERIFY GATEconfidence floor AUDITappend-only stream DEPOT OPShuman signs off WAN EGRESSdiscarded at kernel
+25%
Fleet mission readiness, 90 days post-cutover
< 4 min
Part requisition, down from 72 hours
0 bytes
Classified data egress, enforced at the kernel
< 80ms
Cold-start TTFT inside a SCIF air gap
LIVE TRACE
P99 JITTER11.8ms COLD TTFT74ms EGRESS0 B · drop_all FLEET READINESS+25% REQUISITION3m 48s AUDIT STREAMAPPEND-ONLY OPC-UASECURE CHANNEL SAP RFC500 jobs / 15 min UPTIME99.2% P99 JITTER11.8ms COLD TTFT74ms EGRESS0 B · drop_all FLEET READINESS+25% REQUISITION3m 48s AUDIT STREAMAPPEND-ONLY OPC-UASECURE CHANNEL SAP RFC500 jobs / 15 min UPTIME99.2%
01 / ENTERPRISE PROVEN
Deployed inside the production environments of mission-critical operators.
Shell GLOBAL ENERGY
U.S. Air Force DEFENSE & AEROSPACE
Baker Hughes ENERGY TECHNOLOGY
Con Edison POWER & UTILITIES
Holcim BUILDING MATERIALS
U.S. DoD JOINT STRIKE FIGHTER F-35
Koch Industries ADVANCED MANUFACTURING
Duke Energy POWER & UTILITIES
Shell GLOBAL ENERGY
U.S. Air Force DEFENSE & AEROSPACE
Baker Hughes ENERGY TECHNOLOGY
Con Edison POWER & UTILITIES
Holcim BUILDING MATERIALS
U.S. DoD JOINT STRIKE FIGHTER F-35
Koch Industries ADVANCED MANUFACTURING
Duke Energy POWER & UTILITIES
02 / ENGINEERING PHILOSOPHY

Built for the system.
Committed to the ship.

We operate like SWAT teams for software infrastructure. Engineers embed with your team, write production code, stress-test the environment and harden it far past the baseline.

01

No slide decks. No advisory handoffs.

Engineers embed with your engineering team and ship production code from week one.

02

Stress-tested, then hardened.

We break it before it breaks production — fuzzing, egress pinning and rehearsed rollbacks.

03

Paid for outcomes, not outputs.

Outcome-guaranteed deployment contracts measured at your serving boundary.

Aircraft maintenance hangar
DEFENSE & AEROSPACE
Air-gapped depots
Industrial refinery at dusk
ENERGY & PROCESS
Refineries & plant floors
Server racks in a secure data center
SOVEREIGN COMPUTE
On-prem inference
Operators in a grid control room
POWER & UTILITIES
Grid operations
03 / FDE FIELD REPORTS

Frontier AI fails at the integration boundary. That’s where we work.

Not on reasoning capability — on air-gapped VPCs, legacy SCADA and ERP protocols, and dirty telemetry.

Fighter aircraft under maintenance in a depot hangar at night
FIELD MISSION 01
U.S. AIR FORCE · F-35 JOINT PROGRAM OFFICE

Air-gapped fleet readiness & predictive depot supply

vLLM / FP8 Local GGUF Zero-Egress PrivateLink OPC-UA Bridge
01 SITUATION

Depots lost weeks per turnaround. Technical orders spanned 40 years across three incompatible document systems, logs were hand-keyed from paper, and the facility sat behind an air gap.

02 BOTTLENECK

Retrieval, not inference. Technicians lost 30–70 minutes to search over scanned PDFs with no semantic layer and no route to a hosted vector store from inside the SCIF.

03 ARCHITECTURAL FIX

Everything moved inside the perimeter: FP8 GGUF weights on local vLLM, a hybrid BM25 + HNSW index on depot storage, and egress pinned to drop_all at the kernel.

MEDIAN RETRIEVAL TIME PER LOOKUP
Before
70 min
After
< 4 min
+25%
Fleet mission readiness rate
< 4 min
Part requisition & resolution
0 bytes
Classified data egress

Technical field log

Five further deployments, in release order. Open a row for the cutover notes.

5 RELEASES · 2025–2026

Streaming consumers were written directly against the PI archive rather than polling it, cutting historian read pressure by an order of magnitude so acoustic models score every channel. The edge cluster writes a scored anomaly, not an alert — operators wake up to a ranked list.

RUNTIME
Edge inference on turbine-mounted nodes
EGRESS
Store-and-forward over leased line
OWNER
Client reliability team
03.B / DEPLOYMENT NOTES & FIELD RUNBOOKS

Cutover reports, written during the cutover.

Internal incident logs, redacted at the client’s request. Hover a report to lift the redaction.

SEV-2 RESOLVED
[ENV-REF: TIER-1 POWER GRID HARMONIC DRIFT]

Root cause analysis & deployment cutover

Harmonic distortion on the 138kV feeder peaked during the overnight EV charging window. Two prior vendors routed inference to a hosted endpoint the compliance boundary forbids. We moved the gateway onto a PrivateLink endpoint inside the client VPC and ran the balancing loop in shadow mode for 40 minutes before letting it write.

CUTOVER
02:14 local · feeder outage window
BOUNDARY
Hosted endpoint → in-VPC PrivateLink, drop_all
VALIDATED
Client grid ops · 18 months AMI replay
SEV-3 RESOLVED
[ENV-REF: DEPOT TURNOVER QUEUE SATURATION]

Retrieval latency & read-hostile document corpus

The depot was retrieval-bound, not compute-bound: forty years of technical orders OCR’d into scanned PDFs. We quantized weights to FP8 GGUF and built a hybrid BM25 + HNSW index on depot storage with egress pinned at the kernel. Maintainers ran the last three rebuilds unsupervised.

CUTOVER
Weekend window · four-hour build
BOUNDARY
Hosted vector store → local hybrid index
VALIDATED
Client depot maintainers
SEV-2 RESOLVED
[ENV-REF: PLANT EDGE CONTROL LOOP INTERLOCK]

Real-time control integration & interlock handover

The kiln runs at 1,400°C and cannot pause for an untested control path. The model reads a read-only OPC-UA subscription and only proposes — the existing interlocks decide. Every proposal was shadowed against lab tolerances first.

CUTOVER
02:40 local · refractory inspection
BOUNDARY
Bidirectional assumption → propose-only writes
VALIDATED
Client process engineers
04 / THE WORKFLOW

An engineering runbook, not a consulting engagement.

Four phases. Each one ends with something running in your own environment.

PHASE 01

RECON

Enumerate every hop between historian and systems of record. Packet capture on the plant floor, a container scan of the exact image digest, one working path through the oldest protocol.

EXIT → one reproducible failure
PHASE 02

FUSION

Write the connector, not the wrapper. Native BAPI/RFC and OPC-UA adapters, Kafka and gRPC fan-in, weights moved into your private repository.

EXIT → repo owned by the client
PHASE 03

HARDENING

Red-team fuzzing against parser and index, egress pinned to drop_all at the kernel, and a rollback drill rehearsed with your on-call in the room.

EXIT → signed artifacts + SBOM
PHASE 04

TRAINING

Paired on-call until your operators hold the runbook, dashboards and incident channel. The handover is the deliverable.

EXIT → client runs the shift
05 / OPERATIONAL ENVELOPE

SLA benchmarks, not marketing percentages.

Every figure is reproducible from your own telemetry once the harness is running. Ask for the raw trace — it is part of the handover.

P99 TELEMETRY JITTER
< 12ms

Across OSIsoft PI streaming consumers, measured at the bridge under full historian write load.

COLD-START INFERENCE
< 80ms

Time to first token on a cold process inside a SCIF. No warm cache, weights resident in local storage.

PROTOCOL COVERAGE
Native

SAP S/4HANA over BAPI/RFC, OPC-UA v1.04 with secure channel, Kafka/gRPC fan-in. No polling shims.

06 / OPEN CORE HARNESS · DUAL-LICENSED

100% open source. Nothing calls home.

The deployment runtime, integration connectors and orchestration harness are fully auditable and self-hostable — whether you serve open weights on your own silicon or bridge a hosted endpoint over a private link.

MIT License

PERMISSIVE

Embed inference pipelines into proprietary software with zero friction.

Apache 2.0 License

ENTERPRISE SAFE

Explicit patent grant and trademark safeguards for legal review.

Zero Vendor Lock-In

SOVEREIGNTY

Swap models freely; run on-prem, air-gapped or in private cloud with no platform fees.

harness.py 
from stage_x import Runtime, AirGapEnclave

harness = Runtime(
    enclave=AirGapEnclave.from_env(),
    egress_policy="drop_all",
    audit_stream="syslog:///var/log/stagex.audit",
)

harness.bridge(
    protocol="opc-ua",
    endpoint="opc.tcp://10.20.0.12:4840",
    write_mode="propose_only",
)

harness.index(corpus="./technical-orders", mode="hybrid")▌
$ pip install stagex-runtime github.com/stage-x/runtime
07 / TECHNICAL ARCHITECTURE BRIEFS

Written after the incident.

Each brief documents the constraint that broke a deployment, the mechanism that fixed it and the measurement that proves it — including where results missed the target.

BRIEF-001 AIR-GAPPED · 2026

Sovereign Weights: Quantized 405B-class models in zero-egress environments

Standing up 405B-class models on in-house GPU nodes with no cloud egress — and why the first month went into FP8 layouts and scheduler defaults.

25 pp · Tech Report READ PDF →
BRIEF-002 HIGH-THROUGHPUT · 2025

Petabyte-scale telemetry fusion at rotating-asset fleets

Streaming vibrational and thermal waveforms into PI and SCADA historians for ~1,000 assets, documenting the full spread of lead times.

38 pp · Field Report READ PDF →
BRIEF-003 LATENCY · 2026

Sub-50ms TTFT under extreme concurrent load

Holding a 50 ms first-token target across five million meter streams under a zero-retention constraint.

33 pp · Tech Report READ PDF →
BRIEF-004 RESEARCH · 2026

Fuel-burn control loops in 1,400°C kiln combustion

Early results: single-digit fuel-burn reduction and unplanned stops down roughly a third. Paper in review.

Forthcoming REQUEST ACCESS →
BRIEF-005 HARDENING · 2025

Zero-knowledge clinical reasoning in confidential enclaves

PHI stripped before ingestion, run inside an enclave and cross-checked against current formularies.

26 pp · Tech Report READ PDF →
08 / INTEGRATION MATRIX

Native drivers. No scraping, no shims.

Every row is an integration we have shipped, with the transport and rollback path stated as deployed.

TIER 01 Hyperscaler & Silicon Kernel runtimes & interconnect constraints
HYPERSCALER

Amazon Web Services

→ HTTPS 443 → VPC endpoint ENI ↺ DynamoDB point-in-time restore
HYPERSCALER

Microsoft Azure

→ Private Link, TLS 1.3 ↺ Local enclave checkpoint
HYPERSCALER

Google Cloud

→ gRPC / HTTP2 → PSC ↺ Stateless client
AI SILICON

NVIDIA

→ Triton C++ · FP8 on SM90+ ↺ Refuses unsafe kernel fallback
TIER 02 Platform & Mission-Critical Historian, DCS, PLC & ERP interfaces
DCS

Emerson DeltaV

→ OPC-DA, TCP 135, read-only ↺ Unload the bridge
DCS

Honeywell Experion

→ OPC-UA 1.04.11, plant-CA cert ↺ Last good controller set
PLC

Siemens TIA Portal

→ S7comm / OPC-UA, TCP 102 ↺ No PLC state change
ERP

SAP

→ RFC / BAPI, outbound-only ↺ Rebuild CD view
HISTORIAN

OSIsoft PI

→ AF SDK event stream ↺ Delete derived tag set
DEFENSE

Lockheed Martin

→ Classified enclave, media carry-in ↺ Re-image signed baseline
ENERGY

Baker Hughes

→ OPC-UA, MODBUS/TCP 502 ↺ Idempotent buffer replay
09 / TECHNICAL INTAKE · NETWORK TOPOLOGY FIRST

Open a Dispatch.

We scope from the network boundary outward, not from a discovery call. State three constraints and an engineer who has run this topology before replies with the interfaces we’d bridge and the egress rule we’d enforce.

01 · TARGET NETWORK ZONE
02 · CORE TELEMETRY / INGEST LAYER
03 · EGRESS CONSTRAINT LEVEL
operator-terminal READY
# topology.json
{
"zone": "scif",
"ingest": "opc-ua",
"egress": "drop_all"
}
# transmit
$ curl -s https://api.stage-x.space/v1/dispatch -d @topology.json

No SDR sequence. Your dispatch lands with an engineer, not an account manager.

10 / TECHNICAL BRIEFING

Questions, answered by engineers.

An FDE embeds directly inside your engineering and operations teams to integrate, optimize and harden frontier AI systems inside existing mission-critical infrastructure. Stage-X engineers write production code, remove inference bottlenecks, build telemetry harnesses and are retained for verifiable production outcomes.