ForgeLine Assist · Fictional demo company
The plant supervisor's workspace that answers from approved documents and acts only with your approval.
An agentic assistant for safety, maintenance and quality on the floor. Ask, find procedures, read the exact source, and turn findings into tracked work.
Problem
Supervisors spend the shift hunting for the current procedure, then act on memory.
Time lost finding the right page
Safety, maintenance and quality rules live in separate documents. A single incident, like a label mismatch on a packing station, touches all three.
Stale revisions in circulation
A sampling interval changes from 120 to 60 minutes and the old printout keeps being followed. Nobody is told which revision is active.
Actions go untracked
Follow-ups are agreed verbally, handovers are rewritten from memory, and tickets in external systems drift from what was actually decided.
Assumption: these pains are drawn from the planning interviews behind this demo, not from a measured customer study. Quantified time-loss figures are [TBD].
Why now
Agentic tooling finally makes document-grounded help practical, but only with guardrails.
Agent frameworks are production-ready
Platforms such as Dify expose tool-calling agents with bounded tools and published workflow versions. One agent definition can run many isolated investigations.
Retrieval over approved sources works
Answers can cite the exact document, revision and section. Provenance is checkable by the supervisor, not taken on faith.
Plants already keep docs in versioned repositories
Markdown in Git, review rules and webhooks give a clean signal for "this revision is now approved." Monitoring can be revision-aware from day one.
Trust is the gap
Generic chat assistants cannot be allowed to touch plant systems. The opportunity is an assistant whose every external effect is approved exactly as shown.
Product
Four connected views, one investigation object.
Investigate a question
Natural-language requests with inline sources and follow-ups. Case context persists.
Find approved sources
Search and filter by domain, equipment and revision. Opening a result keeps the investigation.
Read the exact evidence
Document, revision and section always visible, with the assistant beside the page.
Follow work and changes
Attention queue, tasks, document updates, handovers and the approval inbox.
The investigation is the shared object
Equipment and batch context, evidence, questions, tasks and approval requests travel together across all four views. Switching cases switches everything; a citation never starts a new case.
Three kinds of work, kept distinct
Answering, creating demo tasks, and executing external actions each show their own status, so the supervisor always knows what really happened.
Live demo · flagship scenario
PK-04 is printing the wrong labels on batch B-204.
Live captures from forgeline-assist-app.vercel.app, taken during this build — not mockups. The completed flagship run (evidence, two demo tasks, drafted handover, one issue awaiting approval) is not shown: a confirmed bug currently drops a freshly created investigation the moment the browser navigates to its detail page, because the in-memory store's process-local pin doesn’t survive Vercel’s serverless instances. Fix is the already-drafted Postgres migration (db/schema.sql), not yet wired in.
How it works
One agent, bounded tools, deterministic execution.
Dify Cloud is proposed, not confirmed. Database, worker and email providers are unselected. Dify workflow runs are stateless; the application supplies scoped case context.
Trust and safety
The model proposes. People approve. Code executes.
Exact-action approval
The review screen shows target system, account, resource, action type, exact fields, evidence revision and proposed actor. An edited payload, changed target or superseded source revision invalidates the approval. The model cannot set approved = true.
Provenance validation
Answers are validated against retrieved sources before display. Missing evidence is stated, not filled in. Failed tool calls stay visible so unsupported conclusions cannot slip through.
Emergency boundary
No emergency dispatch, no bypass instructions, no restart or product-release authorization. Closing an issue is never equipment clearance.
No machinery control
The assistant reads documents and writes tickets. It never connects to PLCs, telemetry or line controls. Credentials and the allowlisted repository are server configuration, not agent inputs.
Also: no success is reported without a persisted or external receipt; secrets stay server-side and out of logs; email review links never approve anything.
Keeping supervisors current
A document is not authority because it was uploaded.
Detect
Signed GitHub push on approved knowledge paths; 15-minute reconciliation catches missed events.
Verify
Only revisions approved by the designated document owner qualify. Drafts and unapproved branches are ignored.
Stage and check
Index the new revision, run a retrieval smoke check, then promote. A failed update keeps the last working corpus and shows a stale warning.
Compare
Old versus new, impacted equipment and open cases. Pending approvals tied to a superseded procedure are re-checked or invalidated.
Notify
One update event, in-app plus email to verified subscribers. Unchanged checks stay silent; routine changes group into a shift digest.
Seeded example
CHG-001: QLT-001 revision 2 → 3, inspection sampling 120 → 60 minutes. Expected result: one update investigation, one accurate alert, no alert on unchanged state.
Email done honestly
Durable outbox, stable notification IDs, delivery and bounce tracked separately. Provider acceptance is not proof of inbox delivery. Quiet hours with owner-defined urgent exceptions.
Go-to-market
Start with one plant, one connector, one measurable shift.
Who
Mid-size discrete manufacturers with written safety, maintenance and quality procedures and shift supervisors who own follow-ups. Assumption: they already keep documents in reviewable repositories or can migrate a subset.
How
Pilot with one plant. Seed their approved documents, connect one ticketing target, define one approver. Measure time-to-answer, citation accuracy and approved-action throughput.
Then
Expand connectors using the same approval and receipt contracts. Add plants, not agent varieties.
What we are not claiming
Market size (TAM/SAM/SOM), pricing, customer counts and revenue are [TBD]. No customers, pilots or revenue exist for this fictional demo company. Business model is an open decision: per-plant subscription is the working assumption.
Roadmap
18 to 30 focused hours to a full working demo, then pilots.
After the demo
Vercel production deploy completed & verified live (forgeline-assist-app.vercel.app →). Next milestone: first plant pilot. Timing depends on partner readiness: [TBD].
Deliberately out of scope for v1
Voice, machine control, a fleet of specialist agents, multiple connector platforms. Estimate assumes ready credentials and ordinary APIs; it is not a guaranteed date.
Team and ask
Placeholders until real people and numbers exist.
[TBD] · Founder / product
Background and manufacturing operations experience [TBD].
[TBD] · Engineering
Agent systems, durable workers and integration experience [TBD].
[TBD] · Plant advisor
Shift-supervision and safety-compliance perspective [TBD].
The ask
Raising [TBD amount] to fund: completing the full demo (18–30 h), the first plant pilot, and connector and email provider onboarding. Use of funds split, runway and milestones: [TBD]. No revenue, valuation or traction figures are asserted on this deck.
ForgeLine Assist
Ask. Verify the source. Approve the exact action. Stay current.
A supervisor workspace that keeps humans in charge of every effect on the plant.