Rajesh Gheware@ghewaredevopsai
4
Events attended
2
Submissions made
India
20+ years of experience
About me
Founder of Gheware DevOps AI. Corporate trainer and builder in agentic AI (LangChain, LangGraph, RAG, MCP), Kubernetes and DevOps. I run hands-on agentic-workflow and multi-agent courses for enterprise teams and build AI agents on AMD/ROCm and k3s.
🤓 Latest Submissions
Compressor Maintenance Copilot
Compressor Maintenance Copilot watches compressor sensor data, warns before a fault stops the machine, and turns each warning into something a maintenance lead can act on. A statistical detector learns what normal looks like from two months of real data (UCI MetroPT-3, a metro train's air compressor) and is never tuned on failures. On the four air leaks recorded later that year it warned before three of them, 14 hours to more than 2 days ahead, with about one false-alert episode per week. Each alert goes to an Evolus workflow. Its Evolus agent thinks with DeepSeek-V4-Flash running on AMD GPUs through AMD Token Factory, connected as our own model, and searches an Evolus knowledge library of the compressor manual and past work orders. It explains what is happening in plain language, names the likely cause, lists the checks to run, cites its sources, and drafts the work order with parts to have ready. The technician can then ask follow-up questions in chat, such as what was done last time. Safety rules are enforced in code, not only in the prompt: every citation must be a real document, any "safe to keep running" verdict is removed because that decision belongs to the maintenance lead, and low confidence or an unreadable answer escalates to a person. A second machine (AI4I 2020 milling data, with per-failure-mode classifiers) shows that the same flow generalises. Built with Python, scikit-learn, Streamlit, the Evolus REST API (agent, knowledge library, workflow) and AMD Token Factory. Open source under MIT; the manual and work orders are synthetic and labelled so.
18 Oct 2026
pocketful: a four-seat dark factory
In one line: one human message; five real defects caught before acceptance (three by the modeler's second reading of the spec, two by reviewer rejections); 36 of 36 planted faults caught; every stage folder claims its own stage on the event harness. The reviewer's GAPS.md per stage lists what the shipped checks never asked. We built a software factory, not just an app. Four coding-agent seats work in one BAND Desktop room: a coordinator that plans and accepts, an implementer that writes a requirements ledger before any code, a modeler that writes an independent executable model of the specification without reading product code, and a reviewer that rebuilds every revision from a fresh checkout with no network and plants faults before it accepts anything. The whole run took one human message. The coordinator accepted each stage and opened the next one itself. Every seat commits under its own Git identity, and every commit traces to a message in the room log. The modeler's random operation sequences found three defects before review even started, including a 500 on every historical read for new accounts. The reviewer rejected a race where a capture could land after its hold had expired, then planted that exact fault to prove its probe catches it. Across four stages it planted 36 faults and caught all 36. On the event harness in isolated mode, from a fresh clone, every stage folder claims its own stage, one to four. The run took 8 h 46 min and $74.88 at list price. FACTORY.md documents setup, costs, two disclosed BAND delivery stalls, and known limits.
6 Oct 2026
