AB KineticsContact

●Service AI & agent governance for European industry

Agents that act.Governance that holds.

I build AI agents that diagnose machines, and the governance that lets you trust them to act. Engineer from Aachen, working on service AI for semiconductor equipment.

Drawn byAbhimanyu Kanwarhi — that's me

PM1PM2PM3PM4PM5PM3 · Δ +4.2 K1 85012345

1 transfer robot · 2 process module · 3 load lock · 4 EFEM · 5 FOUP

Service agent · logIllustrative
  1. 1Monitoring5 modules · 142 signals
  2. 2AnomalyPM3 heater zone 2 · Δ +4.2 K vs. recipe
  3. 3DiagnosisThermocouple drift · SOP 114 §3.2 · 3 similar tickets
  4. 4Proposed actionApply TC offset −4.0 K, run one dummy wafer
  5. 5ApprovalAwaiting sign-off · you are the service lead
  6. 6Executed & verifiedΔ 0.3 K · ticket closed · audit #20931

Autonomy L2 · acts with approval

Title
Service agent · governed action sequence
Drawn
Abhimanyu Kanwar
Scale
1 : 1

General notes

Unless otherwise specified:

Human in the loop.

  1. 1.Agents act only within their approved autonomy level (L0–L4).
  2. 2.Every action is logged, attributable and reversible.
  3. 3.Sensitive data stays on EU-hosted models by default.
  4. 4.Designed against the AI Act, Machinery Regulation (EU) 2023/1230 and the Data Act.
  5. 5.Written by an engineer who has serviced the machines.

Tolerance± 0 unlogged actions

§02 — The situation

Adopt agents. Regulate them. At the same time.

European industry is being asked to do both at once. Most AI tooling only helps with the first half. My work sits in the overlap.
  • 57%

    of German companies now use AI — up from 20% two years ago.

    Src: Bitkom, Sept 2026 ↗

  • 56%

    of those users name legal ambiguity as a major challenge.

    Src: Bitkom, Sept 2026 ↗

  • 4

    EU regulations touching AI on machines take effect between 2026 and 2028.

    Src: AI Act · MR · Data Act · CRA

Revision table — AI on machines, 2026 → 2028

Full timeline →
  • A02.08.2026AI Act · Transparency duties (Art. 50): disclose AI interaction, label generated contentAnyone deploying chatbots & agents
  • B11.09.2026Cyber Resilience Act · Reporting of actively exploited vulnerabilities and severe incidentsMakers of products with digital elements
  • C12.09.2026Data Act · New connected products must give users direct access to their machine dataMachine builders, OEM service
  • D20.01.2027Machinery Regulation · Applies in full; ML-based safety components need notified-body assessmentMachine builders, integrators
  • E02.12.2027AI Act · High-risk obligations for Annex III systems (deferred by the Digital Omnibus)HR, critical infrastructure, access
  • F11.12.2027Cyber Resilience Act · All essential cybersecurity requirements applyMakers of products with digital elements
  • G02.08.2028AI Act · High-risk obligations for AI in regulated products — incl. machinery (Annex I)Machine builders

Not legal advice — I write about turning regulation into engineering controls.

§03 — Parts list

Four areas I work on.

Each area stands on its own. Together they take an agent from first idea to something you can trust on a real machine.
§04 — Section A–A · Autonomy

An agent earns the right to act.

Autonomy is a staircase, not a switch. My framework defines five levels, what the agent and the human do at each, and the controls every step up requires.
  1. L0Informs
  2. L1Recommends
  3. L2Acts with approval
  4. L3Acts within policy
  5. L4Acts & reports
Human does→ Autonomy →Agent does

ABK-AL v1.0 · Autonomy Levels for Industrial Service Agents

Read the autonomy framework
§05 — Track record

Built by someone who has held the wrench.

Before building agents, I trained engineers on the machines and serviced them on the shop floor. That's why I know what an agent is touching.
  • 100+

    field engineers using a production AI service assistant I led

  • 30+

    service engineers in a field test of my troubleshooting agent

  • 3

    continents where I have trained engineering and service teams

N-01Production

A global AI service assistant

In my role at a leading semiconductor equipment manufacturer, I led end-to-end delivery of a proprietary AI service assistant.

Prototype to production on Azure, with external development teams, LangGraph orchestration and Model Context Protocol for scalable retrieval.

→ In daily use by 100+ field engineers.

N-02Research

Troubleshooting agent for semiconductor robotics

Master thesis at the University of Cambridge Institute for Manufacturing on wafer-handling robot failures.

A troubleshooting agent combining retrieval-augmented generation with knowledge graphs, tested in industry with service engineers.

→ Field-tested with 30+ engineers; published at AIRSI 2025.

Release for production

Let's talk.

auf Deutsch geht's auch.

Conference talks, research collaborations, or a question about the framework — email is the fastest way to reach me.

Drawn
A. Kanwar
Location
Aachen, DE
Languages
English · Deutsch