Definition and lifecycle
Agent, goal, tools, autonomy, verify, start, stream, replay.
Getting started, reference, runbooks, the autonomy model, the CLI and the platform limits. The documentation describes the design; sections marked design preview will change before release.
Illustrative scenario: Block B-214 · double-bottom block · panel line PL-2 · dock D1, one block, one approval gate
From an empty yard-edge node to a shadow-mode run.
The Hullira SDK is typed Python. Tools are declared with schemas, limits and the qualified WPS envelope, and the policy engine enforces them at call time, not in a review meeting. This is a design preview; the API will change before release.
# Design preview. Bound the weld agent to robot cell R3 on panel line PL-2.
from hullira import Agent, Tool, Limit, Autonomy
params = Tool(
name="robot.set_parameters",
tags=["PL2.R3.WIRE_FEED", "PL2.R3.TRAVEL_SPEED"],
limits=[Limit(max_step="4%", per="60s", envelope="WPS-117")],
)
weld = Agent(
id="agent.panel_weld",
goal="porosity risk < 0.1 per joint, inside WPS-117, least rework",
tools=[params, Tool("weld.read_pool", read_only=True)],
# bounded writes on parameters; class-critical seam order stays with a human
autonomy=Autonomy.L3,
# simulate on the twin before every write
verify="twin",
)
run = weld.start(line="PL-2", block="B-214")
for step in run.stream():
print(step.name, step.status, step.duration)
What each section of the reference covers.
Agent, goal, tools, autonomy, verify, start, stream, replay.
Tag naming, units, read_only, Limit(max_step, per, floor, ceiling, envelope).
Levels, approval routing, shift, block-type and interlock conditions, WPS and PSPC checks, versioning and review.
Steps, status values (queued, running, waiting, approval, succeeded, failed, blocked), durations, outcomes.
Fields, export formats, retention, class and naval evidence packets.
twin.simulate, candidates, horizon, ranking objectives, as-built comparison.
Every Hullira run is an ordered, inspectable sequence. This is run_4b7e19 on Block B-214 · double-bottom block · panel line PL-2 · dock D1. Scenario completes in 10 steps with one approval gate. Predicted distortion at the erection joint fell from 4.2 mm to 1.8 mm after a back-step sequence, one porosity precursor on seam S-31 was cleared inside the WPS envelope, the ballast-tank coat ran inside a 3.4 °C dew-point margin, and the block landed at 1.6 mm with its conformance record written.
Yard Orchestrator · Pulled the work package for B-214 from the yard MES: 46 plates, 212 seams under WPS-117, erection tolerance 2 mm at the D1 joint, PSPC-grade ballast-tank coating, erection slot on D1 in the dock plan. Locked the qualified WPS envelope and the class scope.
Ship Twin · Simulated 18 candidate weld sequences on the as-built twin with the weld-thermal surrogate, ranked on predicted distortion at the erection joint, crane reach on D1 and shift takt. Standing sequence predicts 4.2 mm; candidate #6 predicts 1.8 mm.
Cut-and-Profile · Released the nest for 46 plates on CL-1 with the fit-up gaps candidate #6 assumes. As-cut geometry logged per plate and handed to the panel line; material yield 91.4% against the nest plan.
Panel-and-Weld · Loaded the back-step sequence for the non-critical seams on PL-2 and set heat input and travel speed inside WPS-117. Weld-pool and seam cameras streaming on 12 channels.
Yard Orchestrator · Candidate #6 also re-orders two class-critical butt seams, S-31 and S-32. Sequence changes on class-critical seams stay with a human, so the step was held for the shift welding engineer and approved at 06:12:44.
Quality-and-Anomaly · Twelve weld cameras and the UT station fused at the yard edge. One porosity precursor on S-31 at 06:41 flagged and cleared by a 4% wire-feed trim inside the WPS-117 envelope; the joint was routed to UT and passed.
Fair-and-Form · Post-weld scan registered against the twin: 2.1 mm at the erection joint. Two line-heating passes planned on the inner bottom and executed by the fairing crew; re-scan read 1.8 mm, inside the 2 mm tolerance.
Blast-and-Coat · Ballast-tank blast and first coat scheduled to a window with steel temperature 3.4 °C above dew point. Dry-film thickness read 312 µm against a 280 µm minimum on the stripe coats; PSPC evidence assembled per area.
Schedule-and-Yard · Re-solved the D1 dock and crane plan with B-214 ready 6 hours early and a neighbouring block late: erection slot pulled forward one shift, crane G2 reassigned, no berth conflict. Planner approved the revised plan.
Erect-and-Align · Both blocks scanned on D1 and registered against the twin; fit predicted at 1.6 mm. Shimming guidance issued to the erection crew, block landed and welded out, final scan 1.6 mm. Conformance record written: welds, scans, coating readings, approvals and class evidence for B-214.
Every tool invocation, argument and result is written to an append-only, human-readable log, and every reasoning step is expandable. Nothing about a run is hidden from the yard or the surveyor.
Plan, thought, action, observation, exactly as the orchestrator would record it.
Build and erect B-214 inside WPS-117, hold the erection joint within 2 mm, coat the ballast tank inside the PSPC window, and land it on D1 by the dock plan.
Simulate weld sequences on the twin first. Lead with the sequence because distortion is decided at the arc, then cut to the fit-up the sequence assumes, then weld and inspect continuously. Scan after welding and fair only what the scan shows. Hold any class-critical seam change for a human.
The standing sequence welds the inner bottom straight through and predicts 4.2 mm at the joint, which means fairing hours and a risk to the D1 slot. A back-step sequence with two class-critical seams re-ordered predicts 1.8 mm at a cost of 0.4 hours of takt.
twin.simulate(candidates=18) returns plan #6: back-step on the non-critical seams, S-31 and S-32 re-ordered, heat input held inside WPS-117.
Plan #6 predicts 1.8 mm at the joint and clear crane reach on D1. Three candidates scored lower on distortion but moved heat input outside the WPS envelope and were discarded.
Execute plan #6: eight bounded writes permitted on cutting, non-critical weld sequence, parameters inside WPS, coating windows and dock plan; the class-critical seam order routed to the shift welding engineer.
Porosity precursor on S-31 at 06:41:03. A 4% wire-feed trim inside WPS-117 cleared it in 15 s; the joint was routed to UT and passed. Post-weld scan 2.1 mm, two line-heating passes brought it to 1.8 mm.
Scenario complete. Block landed at 1.6 mm, ballast tank coated inside the window, conformance record written for class review.
What to do when.
| Situation | What happens | What you do |
|---|---|---|
| Edge node offline | Control returns to the controller's last known-good state; interlocks unaffected | Check the node, restore, resume at the current autonomy level |
| Model version fails canary | Automatic fallback to the previous validated version on that process area | Review the evaluation record before re-promoting |
| Approval not answered | The step waits; the run does not time out into a write | Named approver or delegate decides in the console |
| Deadlock between agents | Escalates to a human rather than resolving by timeout | Decide in the console; the decision is logged |
| Rollback condition met | The tag returns to its previous value | Review the trigger signal and the run record |
A yard does not go from paper travelers to unattended welding in one step. Hullira makes the level explicit, auditable and reversible at any time. The first release runs shadow mode, then assist mode, then supervised control on low-risk actions. Class-critical seam changes, lift plans and block release stay with named people.
| Level | What the agent does | What the human does | Where it sits in a pilot |
|---|---|---|---|
| L1 · Shadow | Reads weld, scan, NDT, coating, robot, crane and MES signals; predicts and recommends; writes nothing | Runs the process as today; compares predictions to NDT and surveyor outcomes | Shadow mode; the baseline is measured from the yard's own records |
| L2 · Assist | Proposes weld, fairing, coating, routing and schedule moves with evidence, confidence, envelope and rollback path | Approves each action in the review console | Assist mode, the remainder of the pilot |
| L3 · Supervised control | Executes pre-approved low-risk actions: inspection routing, coating pauses when a window drifts, weld parameters inside the WPS envelope, robot path variants tested in simulation | Approves class-critical seam changes, lift plans and block release | Graduated autonomy, after audited accuracy and ROI |
| L4 · Bounded | Supervised work sequencing and weld adaptation for qualified block types, across processes and yards | Sets the envelope; reviews the shift record and the evaluation results | Planned; after class, yard and insurer acceptance is earned |
The same run engine, the same policy checks, the same audit trail, from the terminal, the review console or the SDK. Design preview.
$ hullira run "block B-214 on PL-2 to D1" --autonomy L2+
→ plan composed 10 steps · 1 approval gate
→ twin.simulate 18 candidates · best #6 · joint 1.8 mm
→ policy.evaluate 8 writes permitted · class-critical seams held for human
→ executing cut.nest .......... ok 3m40s
→ executing weld.sequence ..... ok 2m05s
! approval required seam_sequence S-31, S-32
→ approved shift welding engineer · 06:12:44
→ executing inspect.fuse ...... ok continuous
→ executing fair.lineheat ..... ok 4m30s
→ scenario complete landed 1.6 mm · conformance record written
$ hullira runs show run_4b7e19 --format genealogy
Design targets. None is a measured figure.
| Item | Design target | Notes |
|---|---|---|
| Edge inference, weld and seam alerts | under 50 ms | Jetson Orin at the process area, TensorRT |
| Edge inference, coating or scan anomaly prompts | under 150 ms | Includes policy evaluation |
| Cameras per process area | 6 to 24 | DeepStream and Holoscan, 15 to 60 FPS by task |
| Training per yard | 100 to 500 GPU-hours | During onboarding, then monthly incremental retraining |
| Synthetic variants | 50,000 to 250,000 per vessel class | Cosmos or Omniverse Replicator, validated against real distributions |
| Models served per yard | Routed by latency, tenant, sensitivity and cost | Triton and NIM, canary by process area |
Hullira writes to production equipment. Every capability is scoped, every write is policy-checked against the qualified WPS, class and PSPC constraints, and every action is written to an append-only audit log the yard owns.
| Item | Scope | Status |
|---|---|---|
| SOC 2 Type I | Cloud control plane | QUEUED Planned, first 6 months, with a defense and ITAR-aware review |
| 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 |
| Class-society and defense assurance packages | Disposition and genealogy evidence | QUEUED Planned, 6 to 12 months; not a class approval or certification claim |
| Cited reasoning | WPS and PQR, ABS, DNV, LR, ClassNK, IMO PSPC, naval specs | RUNNING In the design, every answer cites its source |
| Per-tenant and IP isolation | Drawings, WPS, naval data, models, memory | RUNNING In the design from day one |
| On-prem and yard-edge deployment | Naval and regulated yards | RUNNING Design option |
The questions production directors, welding engineers and class-compliance leads ask in the first meeting.
Yes, but only within an explicit tag allow-list with per-tag rate and magnitude limits, inside the qualified WPS, coating, dimensional and schedule constraints, and only at the autonomy level your yard has set. Design partners start in shadow mode, where Hullira writes nothing. Class-critical seam changes, lift plans and block release stay with named people.
Control returns to the robot controller, PLC or planner's last known-good state. Hullira is a supervisory layer on top of the controllers and interlocks you already run, never a replacement for them, so an outage degrades the process to its current way of running, not to a stop.
Advisory recommendations start from the physics-based twin, the WPS and the block drawings in the first weeks. Defect, distortion and coating prediction improve as block history and labelled NDT and surveyor outcomes accumulate. Onboarding is sized at 100 to 500 GPU-hours per yard and 50,000 to 250,000 synthetic variants per vessel class, both design targets.
Only if you choose cloud training. Drawings, WPS and PQR, naval data, process models and quality records are isolated per tenant and never used to train another customer's models. Security is defense and ITAR-aware, a fully on-prem deployment at the yard edge is available for naval and regulated yards, and federated learning across yards is planned and waits on data-rights agreements.
You are, the same as with any procedure change, which is why every write is policy-checked, bounded, logged and reversible, and why anything above your risk threshold waits for a named approver. The audit log records the request, the evidence, the WPS or class citation applied and the human decision.
A paid pilot on one process, scoped to a single wedge with a success metric written into the order form: weld rework, fairing hours, coating non-conformance, dock days or class-approval cycle time. Shadow mode comes first and the baseline is measured from your own records. We do not publish a pilot length because we have not run one.
Documentation beyond this page opens with the design-partner cohort.
↑↓ navigate↵ openesc close