AI Growth Platform for
Three Distinct Businesses

This is the clearest near-term AI unlock in the manufacturing business — and your team has already validated the concept. Two FTEs spending the majority of their time on comparative validation of ~15-file drawing packages is a quantified cost with a direct AI substitution path.

The CAD Validation module builds a structured pipeline: drawing packages are ingested (STL, 2D, CAD, DXF), parsed into a structured representation, and run through an AI validation engine that checks for: dimensional interference across parts in an assembly, tolerance conflicts between drawings and specifications, missing views or incomplete callouts, and consistency between 3D models and 2D drawings. Results are returned as a structured validation report with findings ranked by severity — critical interference flagged immediately, minor discrepancies noted for engineer review.

The human engineer shifts from doing the comparison to reviewing the AI's output and approving or overriding findings. ISO9001 compliance is satisfied by documenting the validation process and statistically tracking accuracy — as Gus noted, the compliance framework requires documentation of process and accuracy, not a specific method. Altir builds the audit trail into the system.

The Metal Service Center sells to thousands of customers with wide SKU variety. The core competitive lever is quote speed and accuracy — customers who wait hours for a quote will find one faster elsewhere. The RFQ-to-Quote module builds an automated quoting engine that handles the majority of inbound requests with no human touch, reserving human attention for complex, high-value, or relationship-sensitive accounts.

Inbound RFQs — by email, web form, or portal — are parsed, matched against available inventory and pricing rules, and returned as a quote within seconds for standard materials. For spec-based requests, the system interprets material specifications, quantities, dimensions, and tolerances against available stock. Instant-buy is enabled for in-stock materials: the customer sees availability and current price and can commit to purchase immediately.

High-touch accounts (large industrials, relationship customers) are flagged for human review with the draft quote already assembled — the rep edits and sends rather than builds from scratch. This is the same architecture Altir built for global electronic components distribution, adapted for metals service center product types and pricing logic.

Manufacturing quoting at Schupan bifurcates cleanly into two paths that require different handling — and Gus articulated this perfectly. High-mix/low-volume (prototyping, fixtures, repairs, obsolete part reproduction via laser scanning) is price-insensitive and repetitive enough to automate substantially. High-volume standard production runs require relationship and human judgment at key decision points.

The Spec-to-Price engine ingests customer-provided CAD files, BOM spreadsheets, or spec sheets and extracts the key quoting parameters: material type, quantities, linear feet or dimensions, surface finish requirements, tolerances, and delivery needs. It then generates a draft quote based on current material costs, machine time estimates, and historical job costing for comparable work. For the high-mix/low-volume segment, this quote can be delivered to the customer in minutes rather than days.

The system also handles laser-scan workflows for obsolete part reproduction — one of Gus's named use cases — where a scan file is processed into a manufacturable specification and quoted automatically. The human engineer reviews before customer delivery for complex assemblies, approves for straightforward geometries.

The trading desk operates at the intersection of commodity market signals, customer demand, and supply availability — making pricing and procurement decisions under time pressure with incomplete information. Next Best Action intelligence doesn't replace the trader's judgment; it gives them the context to act faster and more confidently.

The module aggregates Schupan's historical trading data from Remas, external market signals, customer purchase patterns from Salesforce, and inbound volume signals to surface: recommended buy/sell prices by commodity grade and volume tier, customer accounts due for proactive outreach based on their purchase cadence, commodity positions approaching risk thresholds, and vendor/supplier performance signals.

Coordination across the trading organization is where the multiplier lives — Gus and Seth aligned that speedier results come from synthesizing behaviors across historical and third-party data. This module builds that synthesis layer as a persistent operational tool, not a periodic report.

Carlos named the exact right foundational problem: AI agents operating across four systems without a semantic layer will produce inconsistent, unreliable outputs. The same entity — a customer, a commodity grade, a load — has different representations in Remas, Visual Info, Salesforce, and MineHub. Without a unifying ontology, agents hallucinate joins, miss context, and produce results that operators can't trust.

The Semantic Data Layer is the infrastructure investment that makes every other module more reliable. It builds a domain ontology for Schupan's business — encoding grades, customers, loads, contracts, parts, and their relationships in a machine-readable, context-rich form. This layer sits between the AI applications and the source systems, resolving entity references and providing consistent context on every query.

This work aligns directly with the Databricks migration already in motion. Altir builds the semantic layer to sit on top of Databricks, using it as the lakehouse foundation while adding the ontological structure that makes AI outputs reliable and auditable. The result is deterministic behavior for enterprise decisions — exactly the concern Carlos raised about LLM non-determinism in production contexts.

Schupan has done something most companies haven't: given Claude to the whole organization. That's the right first move. The next move is structured — identifying where the remaining two-thirds of employees are leaving value on the table and building the workflow integrations that make Claude genuinely productive, not just exploratory.

Altir's Claude Deployment Acceleration program addresses this in three tracks: (1) Finance & Excel — building on the existing Excel add-in adoption by creating structured prompt templates and data workflows for the most common financial analyses; (2) IT & Development — shifting from line-by-line validation to AI-assisted code generation, testing, and documentation with appropriate guardrails; and (3) Operations agents — building the first generation of Schupan-specific AI agents that operate against your actual systems (with the semantic layer as the foundation) for load triage, quote drafting, and drawing validation initiation.

This track also addresses the governance concern Gus raised around citizen developers and shadow IT — building the "empowerment with structure" framework so employees can build their own workflows without creating ungoverned dependencies or data exposure risks.

Schupan's Metal Service Center serves thousands of customers, but the depth of that relationship varies enormously across the book. Some accounts are active and transacting regularly; others placed one order a year ago and haven't been contacted since. The Customer Demand Creation module turns this static relationship data into a live outreach engine — identifying which accounts are most likely to need material now, generating personalized outreach anchored to real inventory availability, and tracking the response lifecycle from sent to replied to actioned.

The AI layer runs continuously against the CRM, segmenting accounts by recency, grade preference, purchase cadence, and market signals. When conditions align — a customer who typically buys aluminum hasn't ordered in 6 weeks, and Schupan has strong Zorba availability at competitive pricing — the system surfaces a prioritized outreach task with a pre-drafted message tailored to that account's history. Campaigns can be structured around specific signals: new material availability, seasonal demand patterns, grade-specific shortage alerts, or reactivating dormant accounts.

Each outreach is tracked at the contact level through the full response lifecycle. Response rates, reply content, and conversion to order are captured and fed back into the account scoring model — so the system learns which signals and messages actually convert for Schupan's customer base, not a generic benchmark. This is the compounding advantage: every campaign makes the next one smarter.

Today, a dealer or industrial seller who wants to understand the value of their scrap before bringing it to a Schupan yard has limited options — they call, they guess, or they go to a competitor potentially for a quote. The AI-Powered Scrap Identification module changes that equation: a seller takes a photo of their material, uploads it through a mobile-optimized partner experience, and receives a preliminary grade assessment and value estimate — along with a short series of clarifying questions to sharpen the evaluation.

The experience works in two modes. For consumer and small dealer sellers, it's a mobile-first interface where a photo of a scrap pile, appliance, vehicle part, or industrial offcut is analyzed by a computer vision model trained on Schupan's material types and grades. The system asks clarifying questions — "Is this material painted or coated?", "Approximately how many pounds?", "Is this ferrous or mixed?" — and returns a preliminary value range with an invitation to bring the material to the nearest Schupan location for a final verified assessment and purchase. For auto salvage and industrial partners, the same technology operates as a B2B intake tool — processing bulk photos of salvage inventory, automatically classifying parts and materials, and generating draft purchase offers for partner review.

This module is positioned as a Phase 3/4 expansion dependent on the semantic data layer and trading intelligence modules being in place. The computer vision model requires training on Schupan-specific material types and pricing, which the ADAPT blueprint will scope. It is explicitly subject to evaluation and prioritization by the Schupan leadership team.

Every ton of material Schupan processes has a sustainability story — CO2 emissions avoided versus virgin material production, energy saved, landfill diversion achieved, water conserved. Today, that story exists in the data but has never been quantified and delivered back to the customers who generated it. The Customer ESG Intelligence module changes that: it transforms Schupan's operational transaction data into a customer-facing sustainability dashboard, giving each industrial customer a quantified, auditable record of the environmental impact of their recycling program.

The Customer ESG Score Experience — accessible through the MySchupan customer-facing experience — delivers each customer their personalized sustainability metrics: total tons recycled by material type in any period, estimated CO2 avoided per ton (based on established EPA and industry benchmarks by material), landfill diversion rates, and circularity contribution scores. For large OEM and manufacturer customers who face ESG reporting obligations to their own boards and regulators, this data is not just a nice-to-have — it is becoming a procurement requirement from their sustainability teams.

The module generates automated ESG reports in formats suitable for corporate sustainability disclosures — with Schupan branded, auditable, and citable in customer sustainability reports. This positions Schupan as a strategic partner in its customers' sustainability programs, not merely a commodity service provider — a relationship dynamic that supports both retention and premium pricing. Altir will work with the Schupan team during ADAPT to define the specific metrics, reporting formats, and customer portal experience that best serve Schupan's target customer segments.