Manufacturing Operations & Delivery Assurance

How do you make on-time delivery measurable in a CNC plant?

Digitize the inputs, schedule against real capacity, and classify every shipment. FACT covers Sales, Engineering, Production and QC: PO PDFs extracted under a validated schema with human confirmation, 10k+ drawing-revision rows moved out of Excel with a three-way check, shift-aware scheduling with cascade preview, and OTD computed with day deltas.

Alex Scharifker · Product architect and implementation lead

Designed and led by Alex Scharifker for FACT as a manufacturing operations and delivery-assurance system.

10k+

Revision Rows Digitized

Bounded Contexts

Departments Covered

Problem

FACT On-Time Delivery Hub

A precision CNC plant supplying oilfield equipment ran order-to-delivery on PDFs, WhatsApp, Dropbox folders and a whiteboard. Purchase orders were retyped by hand — sometimes with the wrong part number, because source PDFs list sub-components beside the customer part number. Drawing-revision control lived in one person's 10,000-row Excel file. Scheduling was a whiteboard, so committed dates had no capacity basis, and nobody could produce an on-time-delivery number or explain a late shipment with evidence.

System

How the system is put together

Sources in, deterministic logic in the middle, AI restricted to interpretation.

Data sources

Purchase-order PDFs, AI-extracted and human-confirmed
Bulk spreadsheet imports of order lines
Dropbox design plans and drawing revisions
Machine shift calendars and part-time standards
Shop-floor run start/stop and pieces completed

Layers

Role-scoped operator surfaces on the edge
Typed RPC server functions, no separate API tier
Postgres with RLS across three bounded contexts
Pure-function scheduling and OTD core
Append-only event and audit tables
Server-side integrations: AI gateway, Dropbox, storage

Deterministic logic

Duration resolution: override → standards catalog → heuristic, source labeled
Shift-aware capacity from per-machine productive hours and active shifts
Cascade projection of downstream schedule impact before commit
Three-way revision comparison blocking engineering sign-off
On-time-delivery classification with day deltas
Nine-role, track-scoped permissions with time-boxed delegation

Where AI is used

Extracts line items from unstructured PO PDFs under a validated schema
Flags low-confidence part numbers for human confirmation
Narrates already-computed snapshots through three allow-listed prompts
Never schedules, certifies a revision, or appears as an actor in an audit log

Decision loop

Ingest documents and shop-floor signals
Verify through a five-step engineering funnel
Schedule against real capacity, preview the cascade
Measure OTD and variance, correct the standards

My role

Product architecture and bounded-context design
Data model and append-only audit design
Business rules as pure, testable functions
Workflow discovery across four stakeholder groups
Applied-AI boundary design and implementation leadership

Architecture

The architecture case file

Layer-by-layer architecture, deterministic model, AI guardrails, and the decision loop.

A machining plant could not prove whether an order would ship on time. This system makes the answer computable: documents become data, approvals become state, and commitments are measured against real capacity.

Intake→Verify→Schedule→Run→Measure

Key takeaways

—A precision CNC plant ran order-to-delivery on PDFs, WhatsApp, Dropbox and a whiteboard — nobody could answer, with evidence, whether an order would ship on time.
—A ~10,000-row personal revision spreadsheet became an auditable matrix where a three-way revision mismatch blocks engineering sign-off.
—Whiteboard scheduling was replaced by a shift-aware capacity engine with a cascade preview before any planner commits.
—Hard boundary: AI extracts documents and writes summaries; deterministic pure functions own every date, status, approval and OTD score.

Architecture Principle

One record for an order's life — from PO document to shipped date.

po_line_items is the pivot entity: commercial terms, engineering state, body specs, quality linkage and drawing references on one row.
The scheduling core is pure TypeScript — shift-aware, side-effect-free, independently testable, no model in the loop.
Integrations (AI gateway, Dropbox OAuth, PDF storage) are server-side only, reached through signed URLs and stored refresh tokens.

Role-scoped operator surfaces

TanStack Start (React 19, edge SSR) with file-based routing. Navigation is computed from the user's roles — quality never sees scheduling, operators never see pricing. Bilingual by surface: production and quality are Spanish-first, the commercial layer English-first.

Typed RPC on the edge

Server functions run on the edge worker — PO extraction, workflow transitions, work-order numbering, quality-matrix operations, Dropbox calls, delegations. No separate API tier and no container; secrets live in server-only modules kept out of the client import graph.

Three bounded contexts, one seam

Seam

About 20 Postgres tables spanning orders, engineering/quality and production. They meet at exactly one join — an approved po_line_item becomes a scheduled job — so one team's process change cannot break another's.

Authorization in the database

Three-dimensional access: nine roles, customer review track (in-house vs. third-party), and time-boxed vacation delegation. Enforced through SECURITY DEFINER helpers behind RLS on every table, never through client-side checks.

Append-only history

Auditability

Step events, revision events, date changes and downtime events are write-only tables. Every consequential change carries an actor and a timestamp, so customer disputes are settled with records instead of memory.

Outcome

What changed

Made on-time delivery computable — overall, by customer and by machine — for the first time.
Blocked wrong-revision manufacturing with a three-way check instead of one person's memory.
Replaced whiteboard promises with capacity-backed commit dates and visible delay cost.
Removed single-person dependency on revision knowledge and vacation-stalled approvals.
Instrumented step-level dwell times, making review bottlenecks measurable.