AI Assists - PlanQuill

AI Shadow Logging

See what AI would recommend — safely, transparently, and without changing your schedule.

📦

Clear change history

Every AI output is recorded as a traceable event: what it suggested, when, and why.

Reviewable history

🛡️

Production-safe by default

Built to run in real factories: predictable performance, sensible limits, and safe data handling.

Designed to stay lightweight
Captures only what’s needed for monitoring
Helps validate AI before rollout

⚙️

AI suggestions won’t auto-change plans

You stay in control. AI assists generate recommendations and explanations — planners decide what to apply.

No behavioral change

Live monitoring + export

Watch AI events in real time (helpful during pilots and incident response)
Export recent events for troubleshooting and internal review

Replayable history

Keep a consistent, ordered stream of AI events
Resume where you left off (useful for dashboards and monitoring)
Compare results across versions during controlled rollouts

Security + tenant correctness

Admin-only access to shadow logs and monitoring
Strict tenant isolation (each customer only sees their own events)
Designed to avoid accidental cross-tenant leakage

What this enables

Safer AI pilots (measure value before turning anything on)
Faster debugging when results look “off”
Clear governance for audits and internal approvals

Implemented AI assists (deterministic, no external LLM)

Production-safe helpers with predictable behavior, clear failure modes, and exportable artifacts.

🧪

ScheduleSpec data-quality gates

Validate inbound specs before optimizing (horizon sanity, resources/capacity, unique op IDs, precedence cycles). Fail fast with actionable errors.

API: POST /schedule/optimize → 422 with score + errors[]

🩹

Repair mode (conservative fixes)

Optional repair=true applies deterministic, safe fixes (clamps nonpositive times/capacity, fills missing setup_family, rewrites unknown pools). Cycles get a break-edge suggestion.

If still failing: 422 includes repairs_applied + suggestions

🧩

Spec enrichment (durations + changeovers)

Deterministically enrich processing times and merge changeover matrix entries from observations. Supports tenant-scoped DB-backed sampling when enabled.

API: POST /schedule/optimize with { spec, ai_enrich: { … } }

🧑‍🏭

Planner edits capture

Ingest overrides as immutable events (idempotent), store before/after + metadata, and fetch recent edits ordered by occurred_at.

APIs: POST /ai/learning/planner-edits/ingest · GET /ai/learning/planner-edits/recent

⚖️

Tie-break preference profiles

Infer deterministic tie-break weights from aggregated planner edit deltas (e.g., tardiness vs setups) and persist a latest profile per tenant/plant.

APIs: POST /ai/learning/tie-break/profiles/rebuild · GET /ai/learning/tie-break/profiles/latest

🧾

Explainability trace artifacts

Export a deterministic “why scheduled this way” trace per operation (stable keys + reason codes). Store artifacts and return a presigned download URL.

APIs: POST /ai/explainability/trace/export · GET /ai/explainability/trace/presign

🧭

Scenario sweep runner

Run deterministic multi-scenario what-if sweeps, rank scenarios with stable ordering, and persist results + artifacts for later review.

API: Scenario sweep endpoints available (deterministic ranking)

📉

Due-date probability what-if

Estimate deterministic due-date hit probabilities per order (stable outputs for the same inputs). No ML dependency.

API: Due-date probability endpoints available

Deterministic by default: same input → same output, with explicit artifacts and safe error responses.

No external LLM

Safe failure modes (422)

Traceable outputs

Tenant-scoped data

Developer-facing APIs (ready to publish)

Concrete endpoints, predictable status codes, and copy/paste payloads — designed for production integration.

🧰

/schedule/optimize (quality + repair)

Blocks invalid specs with HTTP 422 and a structured score + errors[] payload.
Optional repair=true applies deterministic fixes first; failures include repairs_applied + suggestions.

🧠

AI enrichment envelope

/schedule/optimize accepts { spec, ai_enrich: { percentile, merge_strategy, … } }.
Supports optional DB-backed sampling (tenant-scoped) when enabled.

🧑‍🏭

Planner learning endpoints

Planner edits: POST /ai/learning/planner-edits/ingest (idempotent) + GET /ai/learning/planner-edits/recent.
Tie-break profiles: POST /ai/learning/tie-break/profiles/rebuild + GET /ai/learning/tie-break/profiles/latest.

Note: profiles are built/stored now; wiring them into the solver's runtime tie-break rule is a separate step.

📦

Explainability exports

POST /ai/explainability/trace/export builds trace artifacts under artifacts/{job_id}/…
GET /ai/explainability/trace/presign returns a presigned download URL.

📊

Scenario sweeps + due-date probability

Scenario sweep runner executes deterministic what-if sweeps, ranks scenarios, and persists artifacts with stable ordering.
Due-date probability endpoints output stable hit probabilities for the same inputs (no ML dependency).

Developer docs include an overview + curl examples and copy/paste payloads (ai_api_examples.md + payloads).

Data & ERP Integration

AI assists learn from your plan-vs-actual signals. PlanQuill connects to your ERP/MES and keeps schedules synchronized.

🗄

SAP

Native integration with SAP ECC and S/4HANA

Connected

🌐

Oracle

Oracle ERP Cloud and E-Business Suite

Active

💻

Microsoft

Dynamics 365 Business Central

Synced

🔗

Custom

REST API for custom integrations

API Ready

Real-time sync

Data integrity

Zero downtime

AI Assists that make schedules more reliable.