String2AI
中文
AI Case Studies

From one line of business need
to an AI system that reports for duty.

These are Agentic application prototypes built for manufacturers and operating teams: production scheduling, business process re-engineering, inventory and sales forecasting, talent matching, executive decisions, agent platforms, and embodied intelligence on a robot arm. None of them is a slide deck. Each is a complete system you can open and operate, with real constraints and real data definitions.

9
case studies
9
live demos
7+
industries
7-day rolling
D+0D+1D+2D+3D+4D+5D+6
Stamping · 300T
Molding · PA66
Plating · Tin
Burn-in · 72h
Assembly · L2
Kiln · Firing
Agent proposes move　Awaiting planner On-time 94.1% · Kitted 11/14 orders

Cases

Five groups by business domain. Green means open it and try it; blue means a password is needed because the demo contains material from a client's requirements document; grey means not yet deployed publicly. All demo data is synthetic, constructed to match each industry's characteristics.

01 / Agentic APS

Production Planning & Scheduling

Agents discover, compute, compare, explain and push. Anything that changes a promised date or issues an instruction stops in front of the planner. A deterministic solver produces the plan; the language model understands and explains.

Terminal blocks & connectorsStamping → Molding → Plating → Assembly

Agentic APS for Terminal Blocks & Connectors

A scheduling prototype across the whole process chain. Multi-level BOM explosion feeds a global material allocation by customer tier and need date, yielding kitting rate, kitting date and bottleneck materials. A 7-day resource Gantt over 15 work centers shows agent proposals in copper for the planner to accept or reject. Four what-if scenarios, equipment failure, late material, added demand and slowed process, trigger incremental rescheduling.

  • 6 agent roles
  • Ontology: 19 classes / 25 relations / 15 constraints
  • CP-SAT scheduling solver
  • Natural language → SPARQL
  • Mold life & changeover matrix
Live demo degsonaps.string2ai.com
Automotive catalyst materialsSubstrate · Coating · Kiln firing

Agentic APS for Catalyst Substrates

A scheduling command center for multi-spec honeycomb ceramic substrates (DOC / SCR / GPF / DPF). The kiln is the hard bottleneck, so agents evaluate order acceptance and due dates against 14-day kiln load. From order breakdown to schedule release is one agent pipeline; a human clicks once, at "confirm breakdown", and the rest runs automatically and traceably.

  • Kiln load & due-date slack
  • Agent explanation + dialogue
  • Dashed markers for moved operations
  • Agent auto-tuning
Live demo sdgc.string2ai.com
PV inverters & energy storage · export-ledGlobal plan coordination

Agentic APS · Export Manufacturing Edition

An 18-agent global edition. The promise changes from ship date to arrival-at-port date. Kitting becomes a three-way test: common parts, country-specific parts, cell matching. The burn-in room is a fixed-duration physical bottleneck. BOMs are configured by platform × country × firmware. Changing a critical part means an 8 to 16 week certification gap in that country. Switching clients proved the value of layering: guardrail and orchestration layers barely moved; the agent and tool layers did.

  • 18 agents
  • L0–L3 autonomy levels
  • Five elements of a human checkpoint
  • Mandatory evidence fields
  • Certification & compliance guardrail
Not yet public
02 / Operations

Business Process & Executive Decisions

AI embedded in the business action itself, not a chat box beside it. Every suggestion carries a reason and a confidence; rule hits route to a human; every decision leaves a trail that can be replayed.

Semiconductor distribution & manufacturingLTC · P2P · WMS · Finance shared services

Business Process Re-engineering MVP

A runnable end-to-end prototype built from the client's requirements specification. Sales loop: AI quote suggestion with explainable reasons and confidence → rule-engine approval in seconds → credit check and ATP → order, goods issue, automatic invoicing. Expense OCR with confidence-based routing; AI sourcing top-3 and three-way match in purchasing; AI tiered collections, AI reconciliation and NLG monthly reports in finance; a Text-to-SQL data analyst. PC portal, six-theme executive dashboard and a six-role mobile app.

  • Zero-dependency Python backend + SQLite
  • 21-interface SAP catalogue
  • Rule engine configuration
  • AI digital-employee ledger & run logs
  • Text-to-SQL
Live demo huaqiang.string2ai.com
Cross-border e-commerce · metal & plastic furnitureAmazon multi-warehouse, multi-batch

MarginMind · Inventory Decisions & Sales Forecasting

The decision object is SKU × warehouse × batch, not SKU. A domain ontology serves as the executable contract between the data team and the decision layer: 32 attribute definitions, 17 source-system mappings, 13 constraint axioms, all validated live before any recommendation. A three-state decision-readiness gate; "blocked" refuses to recommend and generates dispatchable fix items. Attribution, expected-value estimation of options, forecast inflection points, and replenishment / promotion / advertising / competitor responses.

  • Ontology: 8 classes / 7 relations
  • Ready / Degraded / Blocked gate
  • Fault-injection drills
  • Inventory-age & net-price definition traps
  • TypeScript decision engine
Live demo haijia.string2ai.com
Listed software company · board governanceAnnual decision tool for the founding shareholder

Profit Allocation & Executive Scorecard Decision System

Allocates net profit among shareholders, management, employees and strategic reinvestment, and scores management with a balanced scorecard plus veto items to produce an incentive coefficient. Financial inputs flow into a waterfall allocation; 14 KPIs score on threshold / target / stretch tiers with deferral and clawback rules; a new-product-line board back-fills scorecard metrics; four preset scenarios plus custom scenario comparison; a six-step governance process with a decision record and printable report.

  • Four-axis radar & ten-year trend
  • Scenario simulation
  • Veto items & clawback
  • Decision audit trail
  • Mobile-ready
Live demo hrdecision.string2ai.com
03 / People

Organization & Talent

Turn "finding people" from something that depends on personal memory into something data-backed, explainable, auditable and reproducible. The system ranks and gives reasons; a human makes the call.

Large-group HR40,000 employees

Talent Profiling & Job Matching Platform

Ships with 40,000 employee profiles. One sentence, "find 100 people who can go to the Middle East", returns a ranked shortlist with reasons. Hard filters → recall → eight-dimension ranking → organizational balance constraints, each result carrying an evidence chain (language level, project months, relocation willingness). Set per-department draw limits up front; every match leaves a snapshot that can be replayed. An OWL ontology with SHACL compliance red lines: sensitive attributes are barred from the model and enforced by machine. A 300-term skills dictionary is the single source of truth, maintained directly by the business.

  • Eight-domain profile model
  • OWL 730 triples / SHACL 7 shapes
  • Skills dictionary: 14 domains / 300 terms
  • Skill substitution rules
  • LLM Q&A fallback service
Live demo talentmatching.string2ai.com
04 / Platform

Platform & Tooling

Pull out what recurs across single scenarios: knowledge elicitation, ontology modeling, agent runtime and autonomy, evaluation and error feedback. Make the second scenario far cheaper than the first.

Digital-employee factory for manufacturingThe eight-stage FDE methodology as software

MetaJet · AI Digital Employee Factory

Not an agent framework; a factory for digital employees. Four modules in a loop: the FDE Engineer module elicits veteran tacit knowledge through 29 standard question types; the Ontology module makes that knowledge white-box and verifiable with a "7+1" semantic specification and hard provenance binding; the Loop Agent runtime earns trust through L0–L4 graded autonomy, sensitive-action gates and end-to-end audit; the AI Coding console runs assets-as-code throughout. No eval, no deploy.

  • Ontology package library & reuse rate
  • Digital-employee roster
  • Eight-stage gate board
  • Run trace
  • S7 value review
Password required metajet.string2ai.com
05 / Embodied AI

Embodied Intelligence

Take AI off the screen and onto a robot arm. The capture rig, the training and validation rig and the product prototype are one and the same executor, so nothing migrates between training and deployment. Force targets are realized by the controller and constraints sit outside the model: safety comes from the controller and the hardware.

Medical ultrasound equipmentTeleop capture → VLA policy → force-controlled execution

Ultrasound Robot · Teleop Capture × Embodied Brain

The physician scans through the robot's hand, and the brain learns the technique. The physician holds a force-feedback leader arm; a 7-axis force-controlled follower arm holds the probe on phantoms and volunteers. Pose comes from the follower's encoders and force from a six-axis wrist sensor, replacing optical motion capture and IMUs. A three-layer brain runs at real cadence: task planning at 1 Hz, VLA policy at 10 Hz, impedance control at 1 kHz, with a safety filter clipping actions outside the model. Demonstration capture → autonomous policy with physician takeover (DAgger) → force-holding co-operation in hospitals; 16 weeks to a rig, 18 months to an L3 validation machine. The page is a live-simulated dynamic technical proposal.

  • Sync error < 5 ms · leader-follower latency < 8 ms
  • Contact force hard cap 15 N · passive limit 25 N
  • π0.5 VLA fine-tune + US-Encoder
  • DAgger takeover loop
  • LeRobot dataset 250 / 2,000 / 5,000+ h
Live demo ultrasonic-robot.string2ai.com
Methodology · Li Jiang · July 2026

The Real Barrier to AI Adoption,
and the Ontology Answer

Every CEO says "all in on AI". Look closely and there is a contradiction: AI users keep multiplying, yet companies whose way of operating has actually been changed by AI are vanishingly few. The biggest barrier is not technology and not data. It is a long-neglected asset: tacit knowledge. The five steps below are our complete answer to that problem, and the logic underneath all nine cases above.

01 / The barrier

What a company is really worth lives in no system

Twenty years of digitization essentially moved explicit data online. But data is not knowledge, and knowledge is not wisdom.

System one · The database
Explicit data

Orders, customers, finance, approvals and production data in ERP, CRM, OA and MES. Machines read it directly.

What the CRM knows: how much a customer bought over the past three years. One fact, static and lagging.

System two · The human brain
Tacit knowledge

Experience, intuition, judgment, insight; unwritten industry rules, customer relationships, the rhythm of collaboration. Precisely the company's core competitive advantage.

What the sales lead knows: the customer's boss is considering a new supplier; what the decision-maker really cares about; which competitor is courting them; which project is most likely to close.

Datarecorded
Informationunderstood
Knowledgeaccumulated
Wisdomapplied
Decisioncreates value
Where AI excels: recording, understanding, accumulatingThe company's moat: applying wisdom, making decisions
02 / The experience gap

Frontier models have a PhD-level IQ and not one day on the job

Logical reasoning has hit the ceiling; business experience is zero. The gap between them is exactly where companies pay the premium.

WHERE COMPANIES PAY THE PREMIUM PhD-level IQ × 30 years of business experience PhD Senior Entry REASONING IQ 0 yrs · novice 10 yrs 30+ yrs · veteran INDUSTRY EXPERIENCE / BUSINESS CONTEXT Ontology supplies the years on the job Frontier foundation model PhD-level reasoning · 0 years on the job 30-year industry veteran Deep business intuition · outreasoned by the model
IQ is not the bottleneck

Frontier models already reason at PhD level and get stronger every quarter. Waiting one more model generation will not add a single year of experience.

Experience is the wall

Why this order cannot be scheduled, whether this lot can be released, why this machine vibrates. The answers live only in the veteran's head, in no system.

Ontology supplies the years

Write the veteran's experience as a machine-readable ontology, and a PhD-level intellect can start work on your line on day one.

03 / The answer

Turn tacit knowledge into an ontology machines can read

An ontology is a formal model of the business world: concepts, attributes, relations and rules defined in a machine-understandable way, forming a shared semantic vocabulary for the company.

Objects

Customers, orders, equipment, operations. The things that actually exist in the business.

Attributes

What describes an object: status, grade, priority.

Relations

How objects connect: who supplies whom, which operation depends on which machine.

Rules

Implicit judgment made explicit: when a quote should be marked up, when an order should raise an alert.

Before · Traditional data governance
  • Manages tables, fields, metrics, quality
  • For humans: meaning explained by documents and word of mouth
  • AI queries tables directly and cannot read the business context
Now · Ontology-driven data management
  • Manages business objects, relations, rules
  • For AI: a semantic layer lets machines read business language
  • AI understands the business through the ontology first, then joins the judgment

A data model defines the structure and format of data; an ontology defines its context and meaning. The ontology is the interpreter between AI and enterprise data: data supplies the facts, the ontology supplies the meaning.

04 / Engineering

Produce high-quality ontologies at scale, with a low barrier

Companies do not want a handful of one-off ontologies. They need semantic assets that move from "buildable" to "usable, manageable, operable". Domain experts need no technical skill; they answer questions.

Five linked rings of ontology engineering
  1. Pre-processTurn multi-source knowledge into standardized "semantic raw material"
  2. ModelAI and experts jointly generate the ontology
  3. RegisterSemantic assets stored centrally, verifiable
  4. PlatformVersions, permissions, reuse
  5. DeployPlug into business scenarios, serve AI agents
The "7+1" semantic specification × 29 sentence patterns: say the business clearly

Twenty-nine standardized natural-language templates are the "grammar" through which AI understands a business. AI reads forms, documents and manuals and extracts automatically; domain experts speak in the patterns and surface the implicit knowledge no document contains. Three kinds of expression map to three kinds of knowledge and three sets of open standards.

State "what objects exist"

Unify terminology, define attributes and relations, separate common, domain and instance layers.

SKOS · RDF · OWL
Conceptual knowledge
State "how tasks are done"

Decompose atomic actions and link them to objects; define how ontology data is operated on and queried.

OWL-S · SPARQL
Procedural knowledge
State "what rules apply"

Make business control points explicit as "if… then…"; set semantic boundaries and role permissions.

SWRL · OWL · ODRL
Rule knowledge
29 standard question types
  • Who approves this step?
  • When should a quote be marked up?
  • Which machines does this operation depend on?
  • Who decides when an exception occurs?
AI requirements engineer
FDE Agentinterview · probe · extract · structure
Generates the ontology directly

Domain experts only answer questions. The FDE Agent follows each answer to probe exceptions and boundaries, then distills the spoken material into conceptual, procedural and rule knowledge. FDE borrows from Palantir's Forward Deployed Engineer, the on-site role that translates client needs; we turned that role into an agent.

Why now: what agents consume has changed
  1. 2022Hard-coded flowsIn-house scripts · early LangChainHumans write thousands of lines; any business change means changing code
  2. 2023Drag-and-drop flowchartsDify · n8n · CozeHumans drag dozens of nodes; the flow is still fixed in advance
  3. 2024Graphs that loop and branchLangGraphBranching and retries, but every path still designed by a human up front
  4. 2026Just a goal and one documentClaude Agent SDK · Claude CodeA human writes a one-page process doc; the AI breaks it into steps and checks as it goes
  5. SoonNo document at allA trend taking shapeThe AI interviews business staff itself, writes the document and assembles the agent

Earlier agents consumed flows, and translating a business into a flowchart is exactly where tacit knowledge gets lost: a veteran's "depends on the situation" judgments cannot be drawn as nodes and edges. Only an agent that reads documents can catch them. Our approach: the newest agent engines, running on top of an ontology. A human gives the goal and the ontology; the agent reads the ontology to understand, plans, executes by querying data through the ontology, observes and corrects until done.

05 / Landing

Start with an ontology for one small process

The real starting point is not deploying a large model. It is "one small process won", plus a reusable ontology asset.

  1. Pick one painful processContract review, sales quoting, service tickets: high repetition, clear rules, concentrated data.
  2. Build a small ontology for itLay out objects, attributes and relations; make the veteran's implicit judgment rules explicit.
  3. Attach the data to the ontology to form a semantic layerLet AI understand the business through the ontology instead of guessing at tables and fields.
  4. Augment with AI, do not replaceAI suggests, assists judgment and handles part of the work automatically; a human makes the final decision.
  5. Get one working, then replicate the nextProcesses replicate, ontology assets are reused. Every process landed makes the semantic layer one layer thicker.
Knowledge sovereignty: don't let your moat become someone else's training data

The way most companies use AI today is to write tacit knowledge into prompts and upload it as documents to external models. Used once, it grows into someone else's capability, and competitors get it too. Landing AI means not only letting AI read the business, but keeping what it has learned in your own hands.

Data stays in-domain, knowledge stays in-house

The ontology and semantic layer run in the company's own private cloud, server room or plant edge; the model only infers and takes away neither data nor judgment rules.

Swap the model, keep the knowledge

The ontology is the stable knowledge base; the large model is a pluggable reasoning engine. Changing models never means redoing the knowledge engineering, so capability upgrades cost no sovereignty.

Open standards, portable assets

Ontologies are expressed in open standards such as SKOS, RDF and OWL. Semantic assets can be exported, audited and migrated, with no lock-in to a platform or model vendor.

Tacit knowledge→Ontology→AI that reads the business

AI is not a technology revolution; it is a change in how companies operate. Sovereignty does not mean keeping AI outside the door. It means the part of AI that gets stronger with use grows inside your own company.

Shared Practice

Above is the why; this is the how. Nine cases from different industries rest on the same engineering judgment. Each of these five has a corresponding page in every system. They are not slogans in a proposal.

Ontology first

Separate knowledge from execution

Business rules, field definitions and cross-source mappings go into the ontology, the executable contract between data layer and decision layer. Changing a rule means changing a sentence in the ontology, not the code.

White box first

No citation, no output

Every conclusion must attach to an ontology node and a data source. If no citation can be found, the system declines and escalates to a human instead of producing a plausible-looking answer.

Human in the loop

Graded autonomy, adjustable

Start in suggestion mode by default. Small routine adjustments run autonomously; rush orders, outsourcing and due-date changes stop in front of a person. Autonomy is driven by data and signed off by humans.

Solver + LLM

A deterministic core

Scheduling, allocation and matching, problems with hard constraints, go to CP-SAT and rule engines. The language model understands requests, explains plans and talks with people. The boundary is clear.

Evaluation driven

No eval, no deploy

The metric frozen in the scenario card is the first item in the evaluation set. Classified errors feed back into knowledge and ontology, closing the loop rather than ending at a one-time delivery.

About String2AI

Turn a description of the business into an AI system that runs.

String2AI works on the last mile of AI in manufacturing and enterprise operations: how requirements are elicited, how knowledge is modeled, how agents are given autonomy, and how people come to accept the recommendations. Every case here began with a requirements document or an interview and became an operable prototype within a short cycle, used to align requirements, review the approach and prove value before deciding on a full build.

All prototypes are zero- or light-dependency, run offline, and deploy directly inside a client's network.

Get in touch
Emailtomlijiang@qq.com
WeChatcvcreator

To discuss your scenario or see the full version of a demo, reach out directly.

Scan to open on your phone
About data and client names. Product lines, orders, suppliers, employee profiles and financial figures in every demo are synthetic data constructed to match each industry, and represent no company's actual operations. Some demos are password-protected because they contain material from client requirements documents.