The Powdrix platform is a floor-edge inference runtime on NVIDIA Jetson at every printer and cell, GPU inspection servers at CT, a floor orchestrator with a policy engine, specialist agents for every stage of the floor, and an as-built part twin that rehearses each build. Everything an agent does is typed, bounded by the qualified envelope and logged.
Illustrative scenario: PBF-2 · Ti-6Al-4V bracket family, 24-part build, 23 of 24 parts released, one held
Perceive the build, plan the recipe, act inside the envelope, inspect every part, optimise yield and takt, log everything, retrain. The layers below are the design, not a description of a shipped system.
Build-prep and slicer software, printers through OEM control APIs, coaxial and off-axis melt-pool cameras, photodiodes, layer and recoater cameras, thermal sensors, powder lots, CT, vision and metrology, ovens and HIP, robot cells and AM MES. Signals are normalised into one time-aligned record per build and layer.
DeepStream and Holoscan process melt-pool video and sensor streams on Jetson Orin. TensorRT-optimised models classify and predict porosity, lack of fusion, keyholing, spatter and recoat faults; time-series models track thermal history and machine, laser and recoater health. Triton serves them with canary rollout by machine.
The floor orchestrator routes events to specialist agents. Agents retrieve parameter sets, specs and standards through permission-aware retrieval, evaluate the qualified envelope and the policy, ask for approval when required, and write through validated connectors only when the checks pass.
The Omniverse twin simulates thermal fields, a melt-pool proxy, residual stress and distortion for candidate recipes, ranks them against density, geometry, takt and energy, and compares as-run against as-designed as the build progresses. Cosmos and Replicator generate rare defect scenarios for stress tests.
Triton-served 2D and 3D segmentation on L4, L40S or A100 class GPUs at the inspection cell. Weekly training over multimodal build data on DGX or HGX class capacity, a fleet model registry and gated promotion. Federated aggregation across floors is planned and waits on data agreements.
A review console for AM process and certification engineers to approve, correct and give feedback. Corrections train the models. An append-only, assurance-grade quality and qualification audit log records every agent action with model version, lineage and qualification impact from day one.
The core user flow, in order. Outputs are reviewed by an AM process engineer whose corrections train the models.
Every Powdrix run is an ordered, inspectable sequence. This is run_7c2e4b on PBF-2 · Ti-6Al-4V bracket family, 24-part build. Scenario completes in 10 steps with one approval gate. Twin recipe #14 selected, one lack-of-fusion precursor at layer 1,872 cleared by a bounded power trim, CT triage in 11 minutes, 23 of 24 parts released with full genealogy and one held for review.
Floor Orchestrator · Pulled the job from AM MES: 24 Ti-6Al-4V brackets, 60 µm layers, 3,140 layers, density target 99.7%, geometry tolerance ±0.10 mm, AS9100 evidence packet required. Locked the OEM-approved parameter envelope for PBF-2.
Part Twin · Simulated 40 candidate orientation, support and parameter recipes for thermal field, residual stress and distortion, ranked on predicted density, geometry deviation, takt and energy.
Prep-and-Parameter · Recipe #14 released: 32° orientation, lattice supports on two faces, laser power and speed inside the qualified envelope. Predicted distortion 0.06 mm, predicted density 99.8%.
Recoat-and-Powder · Recoater image checked every layer. Powder lot L-2291 on its third reuse, humidity in window. One short-feed at layer 412 corrected with a dosing increase before the next layer fused.
Melt-and-Layer · Melt-pool, photodiode and thermal streams fused at 120 fps. Lack-of-fusion risk rose to 0.61 on part 17 at layer 1,872 and was cleared by a 3% power trim inside the envelope within four layers.
Thermal-and-Distortion · Interlayer temperature on the tall brackets ran 14 °C above the twin's prediction from layer 2,300. A 1.5 s interlayer dwell held the thermal field inside the predicted band through the top section.
Floor Orchestrator · Melt-and-Layer proposed a hatch-spacing change on part 17 beyond the qualified envelope. Held for the AM process engineer, who declined it and chose to flag part 17 for CT priority instead. Decision logged.
Robot-and-Handling · Plate unloaded by the handling cell on build complete, depowdered, and parts transferred to stress relief. Powder returned to the sieve with lot L-2291 reuse count incremented in genealogy.
Defect-and-Inspect · 24 CT volumes segmented for pores, cracks and lack of fusion in 11 minutes. 23 parts graded accept against the acceptance criteria. Part 17 shows a 180 µm indication at layer 1,870 and is routed to the inspector for disposition.
Quality-and-Conformance · 23 parts released with the full genealogy: recipe, layer tags, melt-pool events, thermal trace, approvals, post-processing cycles and CT result. Part 17 held with a non-conformance opened. Evidence packet assembled for sign-off.
Every tool invocation, argument and result is written to an append-only, human-readable log, and every reasoning step is expandable. Nothing about a build is hidden from the floor or from the certification engineer.
Plan, thought, action, observation, exactly as the orchestrator would record it.
Run job J-4471, 24 Ti-6Al-4V brackets on PBF-2, to 99.7% density and ±0.10 mm, with an AS9100 evidence packet on every released part.
Simulate the recipe on the twin first. Release only inside the qualified envelope. Watch recoat and melt pool every layer, compensate thermally against the twin, and route anything outside the envelope to the process engineer. Triage every part by CT before release.
Three candidate recipes beat #14 on takt but predicted distortion above 0.08 mm on the tall brackets. Takt is not worth a geometry failure on a qualified part family. Recipe #14 wins on the stated targets.
twin.simulate(candidates=40) returns recipe #14: 32° orientation, lattice supports, power and speed inside the envelope, predicted distortion 0.06 mm.
Lack-of-fusion risk 0.61 on part 17 at layer 1,872. A 3% power trim, inside the envelope, clears it in four layers. The proposed hatch change would leave the envelope, so it goes to a human and is declined.
Flag part 17 for CT priority. Hold interlayer dwell at +1.5 s from layer 2,300 to keep the thermal field inside the twin's band.
CT confirms a 180 µm indication on part 17 at layer 1,870, two layers below the precursor. The other 23 parts grade accept. The precursor, the action and the CT result are now one linked record.
Scenario complete. 23 parts released with full genealogy, part 17 held with a non-conformance, CT triage in 11 minutes, no unapproved parameter left the envelope.
Two agents will want the same machine or the same parameter. The orchestrator arbitrates on the job goal and the qualification rules, not on who asked first, and the handoff is logged like any other step.
Yield-and-Takt wants to raise scan speed on the build to recover takt lost to the recoat correction at layer 412.
Melt-and-Layer wants to trim laser power on part 17 to clear a lack-of-fusion precursor at layer 1,872.
Lack-of-fusion risk 0.61 on a qualified part outranks 1.1 hours of takt under the job goal. Melt-and-Layer wins the parameter for four layers; the speed change is deferred.
Yield-and-Takt recovers the takt by re-sequencing the oven queue with cuOpt instead. Density prediction holds at 99.8%.
Illustrative scenario.
A qualified floor does not go from manual to unattended in one step. Powdrix makes the level explicit, auditable and reversible at any time, and human approval for anything that affects qualification stays mandatory until validation exists. The first release runs shadow mode, then assist mode, then graduated autonomy.
| Level | What the agent does | What the human does | Where it sits in a pilot |
|---|---|---|---|
| L1 · Advisory | Predicts and recommends, with citations; nothing is written | Enters every change manually | Shadow mode on one machine, first weeks of a pilot |
| L2 · Supervised | Proposes a parameter, thermal or inspection move; it executes on approval | Approves each move in the review console | Assist mode once accuracy is measured against CT |
| L3 · Bounded | Writes inside the OEM-approved parameter envelope and per-tag limits | Approves anything that leaves the envelope or affects qualification | Graduated autonomy on low-risk loops, if the floor is satisfied |
| L4 · Unattended | Runs the envelope without prompting | Sets the envelope; reviews the shift record | Planned; only after validation on the part family |
A build is thousands of layers and a day of machine time. Perception has to be local, deterministic and fast, CT has to be triaged in minutes, and the twin has to run before the build. Powdrix runs GPU inference at the printer edge and keeps training, simulation and optimisation in the cloud or on-prem.
One Jetson node per printer for 30 to 240 fps melt-pool, layer and recoater streams with photodiode and thermal fusion. Sub-250 ms alerting and 1 to 2 s bounded recommendations are the design targets. Jetson Thor is planned for multi-laser machines.
2D and 3D segmentation models for porosity, cracks, lack of fusion, surface defects and metrology comparison, served per inspection cell. Sub-15-minute CT triage per build is the design target.
Defect, melt-pool-instability, build-failure and AM reasoning models trained weekly over multimodal build data, with gated promotion from shadow to assist to bounded. Dedicated capacity is planned.
Thermal, melt-pool-proxy, residual-stress and distortion simulation for recipe ranking, plus robot-cell and queue simulation. Modulus surrogates are planned so sweeps run at production cadence.
Rare pore, lack-of-fusion, spatter, recoat-streak and distortion scenarios generated and validated against real CT and in-situ data before promotion. 1,000 to 10,000 per qualified part family is the initial target.
Build nesting, fleet scheduling across machines, ovens, HIP, inspection and robots, energy windows, and GPU analytics over genealogy tables.
Plate, powder and part handling policies trained in simulation and run on Jetson in the cell. NIM packages perception, prediction and reasoning services behind the floor orchestrator.
Powdrix reads and writes through the systems already on the floor. No rip and replace, no parallel historian, no new slicer to learn. Vendor names are examples of systems found on additive floors; connector availability is confirmed per site.
Build-preparation and slicer software (Materialise, Autodesk Netfabb, Oqton, Siemens NX AM and nTop class tools)
Geometry, orientation, supports and parameter sets in; validated recipes out
Laser and electron-beam PBF, DED and polymer SLS, MJF and FDM machines through OEM control APIs
Machine state and in-situ streams in; bounded parameter writes inside the envelope
Coaxial and off-axis melt-pool cameras, photodiodes, layer and recoater cameras, thermal sensors (EOS EOSTATE, Nikon SLM, Sigma Additive and Velo3D Assure class systems)
Frames and streams fused at the printer edge
CT scanners, surface vision and metrology (Zeiss, Nikon CT, Waygate and Volume Graphics class systems)
Volumes, images and scans in; grading, routing and packets out
Stress-relief and heat-treatment ovens, HIP units, depowdering and support removal (Solukon, AM Solutions and PostProcess class equipment)
Cycle data in; recipe parameters inside the qualified window out
Sieving, lot tracking and reuse records
Lot, reuse count and quality into genealogy
Jetson and Isaac driven handling cells, grippers, transfer rigs
Build-complete and oven signals in; handling tasks out
AM MES and build management (Oqton, Authentise, 3YOURMIND, AMFG and Materialise CO-AM class systems)
Jobs and orders in; genealogy, holds and dispositions back
Powdrix writes to production printers that make qualified parts. Every capability is scoped to the qualified envelope, every write is policy-checked, and every action is written to an append-only audit log the floor owns.
| Item | Scope | Status |
|---|---|---|
| SOC 2 Type I | Cloud control plane | QUEUED Planned, first 6 months |
| SOC 2 Type II | Cloud control plane | QUEUED Planned, 6 to 12 months |
| SSO and RBAC | Console and approvals | QUEUED Planned, 6 to 12 months |
| AS9100 and ISO 13485 assurance packages | Evidence packets and model change control | QUEUED Planned, 6 to 12 months; not a certification claim |
| Standards-cited reasoning | AS9100, ISO 13485, ASTM F42, ISO 52900 | RUNNING In the design, every answer cites its source |
| Per-tenant isolation and on-prem option | Recipes, geometry, models, memory | RUNNING In the design from day one |
The questions operations directors, process engineers and certification engineers ask in the first meeting.
Only inside the OEM-approved parameter envelope, only on an explicit allow-list with per-tag limits, and only at the autonomy level your floor has set. Level 1 is advisory: Powdrix predicts and recommends and a human enters everything. Design partners start there, in shadow mode. Anything that leaves the envelope or affects qualification always waits for a named approver.
Control returns to the printer's own controller and the standing recipe. Powdrix is a supervisory layer on top of the machine controls, monitoring systems and MES you already run, never a replacement for them, so an outage degrades the floor to its current way of running, not to a stop.
Advisory recommendations start from the twin's physics and the part spec in the first weeks. Defect and build-failure prediction improve as layers, CT results and engineer corrections accumulate per part family. The first models are trained on existing in-situ camera data; synthetic rare-defect scenarios augment them only after validation against real CT.
Only if you choose cloud training. Parameter sets, geometry, defect libraries and process memory are isolated per tenant and never used to train another customer's models. Cross-customer learning is federated and IP-isolated and waits on written data agreements. A fully on-prem deployment is available for sensitive producers.
It is designed not to. Validated parameters never change without the sign-off your quality system requires, model changes go through change control, and every action carries model version, data lineage, approval status and qualification impact in the audit log. Powdrix does not claim AS9100 or ISO 13485 certification; assurance packages are planned for the 6 to 12 month release.
You are, as with any process change, which is why every write is policy-checked, bounded, logged and reversible, and why anything that affects qualification waits for a named approver. The audit log records the request, the reasoning, the limits applied and the human decision.
Start with one printer or one inspection cell and one measurable baseline. A paid pilot on melt-pool and build-failure prediction, CT porosity triage or pre-build recipe optimisation shows the number before you commit further.
↑↓ navigate↵ openesc close
