VISION_01Available
DEEPGUARD
A local-first safety layer for AI applications, model runtimes, and tool calls. Adversarial input filtering, confidence-gated output, runtime failure recovery, and a hash-chained audit trail for every decision.
Platform direction
What we're building beyond REFLEX and SPARK.
The broader Gradynt platform vision is a connected industrial AI environment for monitoring, diagnostics, safeguards, workflows, engineering knowledge, and operational data. DeepGuard and IntentCAD are available now. AMIMS is an active development program seeking funding for hardware procurement, testing, and certification. FUSE and FORGE remain directional.
PLATFORM_DIRECTION
Now
Near
Next
Later
Future platform map
REFLEX
Pilot now
SPARK
Pilot now
6
Platform concepts
MonitorREFLEX
DiagnoseSPARK
ExpandVision
VISION_02Orchestration
FUSE
A modular operating layer for connecting monitoring, diagnostics, evidence, applications, and review workflows.
VISION_03Custom AI
FORGE
A path for specialized local models, retrieval systems, and controlled AI deployments around proprietary knowledge.
VISION_04Active development
AMIMS
Late-stage automated, non-destructive insulation resistance testing for ROV umbilicals — run on deck after every dive to build a cable health trend profile. Designed for certification.
VISION_05Available
IntentCAD
Intent-first CAD: a versioned design contract carries the engineering constraints, assertions, and manufacturing context. Geometry is generated from the contract, measured independently, and evaluated against the stated requirements. STEP, STL, and DXF export with cryptographic evidence.
VISION_06Active development
Inventory Copilot
A voice-first, offline-first inventory capture app for field teams—count by talking, barcode, or NFC, then export. Establishes the foundation for CRATE, an offline inventory control system designed to integrate with the broader Gradynt ecosystem.
AMIMS / Active development
Move insulation management from periodic checks to predictive maintenance.
The Automated Matrixed Insulation Monitoring System (AMIMS) turns a periodic, reactive insulation check into a predictive maintenance signal. The hardware and software system automates non-destructive insulation resistance testing across the full conductor matrix of an ROV umbilical—on deck, after every dive. Each automated test sequence feeds a high-resolution trend profile that gives operators advance visibility into gradual degradation, before it escalates into an umbilical fault, unplanned ROV downtime, or asset loss.
Physical safety interlock
Purpose-built for de-energized, on-deck equipment.
AMIMS performs insulation resistance testing on umbilicals and cables only when they are out of the water, out of service, and de-energized between dives. A physical hardware interlock—not firmware or software—makes this operating boundary inherent to the system: an energized asset cannot be tested.
- Testing is confined to the between-dive maintenance window, after the asset has been recovered to deck, isolated, and taken out of service.
- The dedicated hardware interlock prevents the test circuit from engaging unless the asset is de-energized, independent of software state or operator action.
Development stageLate-stage development, with substantial engineering and design complete. Hardware has not yet been procured, and no units are in production.
Certification targetsCE marking and DNV Type Approval are planned certification paths. They will be pursued only after hardware procurement and testing are complete.
Funding opportunityGradynt is actively seeking investment and development partners to fund hardware procurement, final build, and full testing.
Standards basisIEEE 43-2013, IEC 60228, IEC 61010-1 CAT III, DNV-ST-E407, IMCA M 166, and IEC 61326-1.
Target marketsWork-class ROV operators, EPCI contractors, DSV fleet operators, offshore operating companies, and offshore wind and trenching operations.
AMIMS_01Trend intelligence
Five-tier alarm architecture
INFO, WATCH, ADVISORY, WARNING, and CRITICAL states convert insulation trends into graduated operational priorities. Instead of a single pass/fail snapshot, teams receive earlier, proportional warning as cable health deteriorates.
AMIMS_02Integrated variant
AMIMS Integrated
Links cable-health trends to REFLEX anomaly detection and SPARK differential diagnosis. Evidence-gated escalation and TDR trace correlation turn emerging degradation into a traceable investigation path.
AMIMS_03Standalone variant
AMIMS Standalone (SaaS)
A dedicated cloud portal for continuous insulation monitoring, trend visibility, and maintenance evidence—without requiring REFLEX, SPARK, or the broader Gradynt platform.
DeepGuard / Available
Put enforceable safety controls between industrial AI and the real world.
Industrial AI can draft a report, recommend an action, or call a tool in seconds. The risk is that a manipulated request, failing model, or low-confidence answer can move just as quickly. DeepGuard adds a control layer around that path. It filters adversarial input before execution, monitors the AI while it runs, and checks confidence before releasing the result. Each decision is written to a tamper-evident audit trail so teams can see what was allowed, blocked, or escalated and why.
Controlled decision path
Inspect the request. Watch execution. Gate the result.
DeepGuard evaluates an AI interaction at the points where risk enters and where consequences leave. Suspicious or policy-breaking requests can be stopped before they reach the model. During execution, failure signals can trigger a defined fallback or safe default. Before any answer or action is released, confidence rules determine whether it may proceed, needs operator review, or must be withheld.
- Adversarial input filtering catches prompt injection, instruction bypass attempts, and other manipulative input before it can steer the AI.
- Runtime monitoring detects failures and routes the interaction to an approved fallback instead of passing an unreliable result downstream.
- Confidence is assessed from observable response signals, not simply from how certain the model says it is.
- Operator-defined policies set the release threshold for each use case, allowing stricter controls where the consequence of error is higher.
AvailabilityAvailable today as a safety layer for industrial AI applications, model runtimes, and controlled tool use.
Protection pathIncoming request inspection, runtime monitoring and fallback, confidence-gated output, and a hash-chained record of each decision.
Deployment modelLocal-first. DeepGuard can operate without internet access, cloud APIs, or required outbound data flow, keeping sensitive prompts and results inside the operating boundary.
What it is notDeepGuard is not a safety certification and does not make an AI model infallible. It adds controls and evidence around decisions that still remain under operator responsibility.
DEEPGUARD_01Audit
Tamper-evident decision history
Each safety decision records the relevant input fingerprint, outcome, confidence, policy, and time in a hash-chained ledger. Reviewers can trace how a result was handled and detect if the record has been altered.
DEEPGUARD_02Policy
Confidence-gated release
A plausible answer is not automatically a safe answer. DeepGuard applies use-case-specific confidence thresholds, then releases, escalates, or withholds the result before it can influence an operational decision.
DEEPGUARD_03Control
Local-first operation
Safety checks run close to the application and its data. Teams can keep raw prompts, model outputs, and operating context within their own environment while retaining the records needed for review.
IntentCAD / Available
Turn engineering requirements into CAD you can prove.
A CAD model can show what was drawn without proving why it is correct. IntentCAD puts dimensions, tolerances, functional constraints, materials, interfaces, construction instructions, and executable checks in one versioned design contract. That contract is the design record. IntentCAD builds the geometry, measures the finished solid independently, and checks the result against the stated requirements. It is available now for teams that need a clearer, safer path from engineering intent to manufacturing export.
Build, measure, prove
The contract defines the part. The finished solid has to prove it.
IntentCAD never assumes that requested geometry came out correctly. After each build, an independent measurement layer inspects the actual solid and tests it against the contract's assertions. If a 3 mm fillet measures 2.8 mm, the result says 2.8 mm. The plan cannot grade its own work.
- Create boxes, cylinders, pockets, slots, fillets, chamfers, extrudes, revolves, counterbore and countersink holes, mirrors, and linear or polar patterns from the controlled record.
- Check envelope and feature dimensions, hole count and diameter, wall thickness, spacing, edge distance, orientation, topology, volume, mass, and required exports.
- Specify material density as a typed value with its source citation and normalization rules, so mass calculations retain their basis.
- Run geometry builds in contained processes with resource limits, protecting the host if a model fails or an operation runs away.
AvailabilityAvailable today as a working, intent-first CAD system for requirements-driven design and verifiable engineering output.
Source of truthThe versioned design contract is the design record, not a companion specification. It keeps requirements, materials, interfaces, assertions, and construction instructions together.
Change controlChanges follow a propose, validate, apply workflow. IntentCAD checks constraints, flags weakened requirements such as a looser tolerance or thinner wall, and applies only changes that pass validation.
AI accessA local MCP server lets AI assistants read contracts, propose and validate changes, run builds, and verify evidence without bypassing the contract or measurement layer.
OutputsMeasured assertion results, built-model snapshots, and STEP, STL, and DXF exports. Each artifact carries a SHA-256 hash for verification.
What it is notIntentCAD does not certify that a part can be manufactured and does not replace engineering review. It makes intent and measured evidence easier to inspect before an engineer approves the design.
INTENTCAD_01Change control
Change the design without weakening it by accident
Each proposed edit is an atomic changeset that must validate before it can be applied. IntentCAD checks whether the new design still meets its constraints and calls out requirement drift, including relaxed tolerances and reduced wall thickness.
INTENTCAD_02Evidence
Trace every export back to its requirements
Measurements, assertion results, model snapshots, and exports are linked by SHA-256 hashes. A reviewer can verify which geometry was checked, which requirements passed, and whether an artifact changed after the evidence was recorded.
INTENTCAD_03MCP
Bring AI into a controlled engineering workflow
AI assistants can inspect the design record, propose changes, validate them, run the build pipeline, and verify the resulting evidence through a local MCP server. They work through the same controls as the engineer instead of going around them.
Roadmap conversation
Contact Gradynt
Start with REFLEX, SPARK, or an implementation service.
Platform direction is useful when it clarifies where your roadmap might connect, but the practical starting point should be a real asset, diagnostic workflow, or AI adoption problem.