Autonomous Industrial Control Agent
The Rigidity of Industry 4.0.
Our client operates one of the most advanced manufacturing facilities globally. Despite heavy investments in Industry 4.0, their production lines remained fundamentally rigid.
Control logic was hardcoded into Programmable Logic Controllers (PLCs). Whenever a new product variant was introduced, it required weeks of costly downtime for specialized automation engineers to rewrite and validate the logic. When unexpected sensor anomalies occurred outside the hardcoded definitions, the entire line executed an emergency halt — causing catastrophic production delays.
The Gap Between Cloud AI and Factory Floors.
The client envisioned using LLMs to create 'Software-Defined Manufacturing,' where production logic could be altered via natural language. But injecting cloud-based frontier models into a factory floor presented critical roadblocks.
- 1.Latency & Connectivity: Industrial control requires millisecond latency. Cloud API roundtrips over factory networks were too slow and unreliable for real-time robotic actuation.
- 2.Data Sovereignty: Manufacturing IP and telemetry data are highly classified and cannot leave the air-gapped facility networks. Cloud-based models were architecturally incompatible.
- 3.Domain Incompetence: Off-the-shelf LLMs possess vast general knowledge but fail at understanding proprietary industrial protocols, sensor telemetry, and strict Standard Operating Procedures.
Deploying 'Living SOPs' at the Edge.
Factory telemetry stays inside the facility while edge agents reason over SOP changes and anomalies.
I. Domain-Specific SFT (Supervised Fine-Tuning)
We bypassed massive generic models. We constructed a proprietary dataset mapping the client's complex sensor data, actuation commands, and safety protocols. Rigorous SFT on efficient 7B models created a domain-specific control layer whose private benchmark evidence can be reviewed under NDA.
II. Event-Driven Agentic Framework
We built a closed-loop perception-action engine. The LLM Agent continuously ingests structured telemetry from the factory floor, operating dynamically and reacting to asynchronous events in real-time — executing tasks precisely when preconditions are met.
III. Dynamic 'Living SOPs' & Anomaly Reasoning
We replaced thousands of lines of PLC code with semantic prompts. The client can now alter production workflows by updating a natural language document. When confronted with unexpected sensor events not covered in the SOP, the fine-tuned Agent uses zero-shot reasoning to deduce the safest mitigation strategy — rather than blindly shutting down the line.
Software-Defined Manufacturing Realized.
"We don't just chat with data; we command the physical world."
Related implementation briefs.
Claude + Token101 + Edge DLP
Auditable Financial Compliance Agent
A 4-agent auditable compliance system with document parsing, risk flagging with citations, regulatory checklist verification, and a complete audit trail — with Zero Data Retention and Token101 complexity routing.
Multi-Agent + MCP + Token101
E-Commerce Intelligent Operations Agent
A full-funnel e-commerce intelligence system covering pre-sale marketing, after-sales ticket management, and operations monitoring — connected to Shopee, Lazada, and TikTok Shop via MCP.
Claude Extended Thinking + Citations
Legal Contract Risk Review Agent
A contract risk review system built on Extended Thinking reasoning chains and clause-level citations, enforcing a unified review standard across Chinese, English, and Japanese contracts.