Exploring AI Agents in Process Industries
- 1. Exploring AI Agents in Process Industries Simple Use Cases and Practical Insights Andre Moreira, July 2025
- 2. Agenda 2 ● About lyfX.ai ● The Opportunity ● Agentic AI in Process Industries ○ Emerging Use Cases ○ Frameworks: From Prototype to Production ● Examples From Our Project Portfolio ● Lessons Learned and Conclusions
- 3. 3 ● Founded in January 2024 ● Data and AI for producing industries (chemicals, agriculture, engineering, etc.) ● Working with leading multinationals, SMEs, Startups Dr. Andre Moreira ● PhD in Theoretical Physics ● Executive experience in Chemicals BASF and AgTech Novihum) ● International expertise having worked in Germany, USA, Asia ● Technical development + Data & AI integration strategies into businesses Prof. Peter Lenz ● PhD in Theoretical Physics ● Expert in modeling complex biological and physical systems ● Professional background in Germany Max Planck), USA Harvard, Stanford), France Institut Curie) ● Advanced AI/ML and big data analytics Certifications:
- 4. Process Industries: The Opportunity 4 Thereof: DE Thereof: EU Chemicals 2023 global sales 5,195 bn EUR 655 bn EUR 225 bn EUR Figures from: CEFIC, GTAI, CESifo Status of AI in manufacturing 2024 Some adoption barriers ● Concerns about IP ● Inertia / lack of urgency ● Regulation / uncertainty about liability Example: Chemical Industry
- 5. Agentic AI: Emerging Use Cases 5 Some AI examples (not always agentic) ● Marketing ○ Monitor customers and competitors ● HR ○ Candidate selection ● Manufacturing ○ Process monitoring and optimization ○ Planning “co-pilotsˮ ● Logistics ○ Route optimization AI agents are software systems that can go off and accomplish tasks on their own, with minimal supervision “Routine & intelligenceˮ processes: mostly routine, but not mindless This can be a browser, a robotic arm, a factory control room, etc!
- 6. Frameworks - A Personal Journey 6 Frameworks handle state management, error recovery, and flow orchestration of multi agent systems ● BYO (“Build Your Ownˮ) ○ Must pay attention to API call management (e.g. timeouts); pydantic very useful ○ Pain to maintain, or extend beyond the favorite models, etc. ● Swarms / OpenAI Agents SDK: good starting point ● Hugging Face smolagents: better starting point (course) ● LangGraph: great, despite its relatively steep learning curve (course) ○ LangSmith is very useful to trace agentsʼ steps, number of tokens, latency, etc.
- 7. Use Cases 7 Selected Examples From Our Project Portfolio
- 8. Carbon Emissions Estimator 8 Tech stack: ● python, LangGraph, Django (backend + frontend), Open AI API ● Claude in pair programming 🤖 = LLM (with or w/o tools) “Calculate the CO2 emissions of sulfuryl fluoride” https://agents.lyfx.ai/agent/cccalc/
- 9. Carbon Emissions Estimator 9 Graph Calculator agent (ReAct) Triage (coordinator) ● States are TypedDict LangGraph optimized for it) ● Structured responses are BaseModel (pydantic; true to its name) https://github.com/andremoreira73/lyfx_agents_public (see langgraph_agents / cccalc folder)
- 10. 10 AI workflow for a systematic, steady stream of interim management postings Workflow ● Scrape web pages ● Analyze: is this a list of jobs or a specific job posting? ● Decide: ○ Job lists go to a list of pages to be followed up (scraped) ○ A job posting goes to a DB if relevant to our goal ● “Rinse and repeatˮ Interim Jobs Scanner
- 11. Interim Jobs Scanner 11 Single-tool (scraper) ReAct agent ReAct ☝is it “Lazyˮ to employ a single-tool ReAct agent?
- 12. Interim Jobs Scanner 12 👍Not lazy at all… Using a ReAct agent makes it very easy to have dozens of instances running in parallel (map-reduce pattern)
- 13. Interim Jobs Scanner 13 Graph node Agent with structured response ● LangGraphʼs Send function for parallelization ● Control over what the agent sees ○ OverallState vs. PageState https://github.com/andremoreira73/lyfx_agents_public (see IM_Scanner folder)
- 14. Interim Jobs Scanner - Metrics 14 Comparison (two runs) Version 2.0 with LangGraph Version 1.0 “vibe codedˮ with Claude Sonnet 3.7 Run date 18 June 2025 24 May 2025 Total run time 793 s 13 min) 2082 s 35 min) Target sites 25 25 Visited sites 25 10 (selected randomly) Visited pages 77 3 passes) 17 2 passes) Relevant postings found 44 9 Total tokens (input & output) 8,656,980 3,026,782 Model(s) used o3-mini gpt-4o-mini, o3-mini Run time / pages visited 10.3 s 122.5 s Effective tokens / second 10,917 1,454 Costs as of 18.06.2025 $10.26 ○ Input tokens: 8,424,501 $9.24 ○ Output tokens: 232,479 $1.02 1.0 🠊 2.0 11.9 x faster 1.0 🠊 2.0 7.7 x Version 2.0: https://github.com/andremoreira73/lyfx_agents_public see IM_Scanner folder Version 1.0: https://github.com/andremoreira73/curious_surfer_public
- 15. Other Routine & Intelligence 15 Transform process flow diagrams into accurate descriptions Research participants in a conference/trade show and prepare a targeted list + email
- 16. Lessons Learned, Conclusions 16 Market ● Start with a concrete project or highly specific use case, even if not really “super sexyˮ or “the next unicornˮ ○ Importantly: using AI agents actually helps the workflow ● Producing industries are in the very early stage of adoption ○ Open to conversations, but still reluctant to take it seriously; this will change over time Technical ● Use an existing framework to orchestrate workflows ○ After the learning curve, it saves time and simplifies scaling and maintenance ○ It is a fast-moving ecosystem, staying up to date is essential ● Prompting remains central to development ○ Tools like playgrounds or Jupyter Notebooks help iterate quickly before setting the agents “free in the wildˮ ● Add “human checkpointsˮ to boost reliability and build trust in the outputs ● Traceability is a must (e.g. via LangSmith)