Public build announcement

MADD — Maximo Assistant for Deep Diagnosis.

Not another chatbot. MADD is being built as an evidence-first diagnostic assistant for IBM Maximo: skill packs first, runtime evidence next, deep diagnostics where needed, and LLMs only when they actually earn their keep.

Read the announcement Follow the build story

Announcement

We are building MADD in public.

MADD stands for Maximo Assistant for Deep Diagnosis. It is a new MASReady capability designed to help Maximo teams move from vague troubleshooting to evidence-backed diagnosis.

The ambition is simple: when Maximo breaks, the assistant should not merely produce a confident paragraph. It should inspect the right evidence, explain the likely root cause, generate safe validation checks, and prepare a solution pack that humans can review and approve.

This begins with Automation Script Root Cause Mode and will expand into escalation, cron, integration, security, environment drift, patch impact, migration readiness, and deep stack trace diagnostics.

Why it matters

Enterprise AI should not be “LLM on every click.”

Many enterprise AI tools send every question to a large language model, even when the answer could come from approved product knowledge, local skill packs, or live system evidence.

MADD takes a different path. It is designed to use curated Maximo skill packs, customer-approved runtime evidence, environment fingerprints, BMX error parsing, psdi stack trace analysis, and human approval workflows before reaching for optional LLM synthesis.

That makes MADD both a MASReady product capability and a practical OverpayingForAI use case: use expensive AI only where it adds value, not as the default answer to every problem.

How MADD works

Evidence-first diagnosis, human-approved action.

The first version is deliberately safe: read-only evidence, no automatic Maximo mutation, and no external LLM call by default.

1. Understand the question

MADD classifies the issue: automation script error, escalation troubleshooting, cron task behaviour, integration failure, security access issue, patch impact, or migration risk.

2. Plan the evidence

The assistant decides which Maximo evidence is needed: AUTOSCRIPT, SCRIPTLAUNCHPOINT, MAXOBJECT, MAXATTRIBUTE, CRONTASKINSTANCE, ESCALATION, object structures, logs, or scan data.

3. Fetch read-only data

Where authorised, MADD uses customer-configured read-only APIs to fetch the minimum necessary evidence into memory. Demo tenants can fall back to seeded evidence.

4. Diagnose deeply

It parses BMX errors, psdi stack traces, launch point timing, MBO lifecycle clues, environment differences, and known Maximo patterns to identify likely root causes.

5. Generate a solution pack

The output includes technical analysis, read-only validation SQL, recommended steps, test plan, risks, assumptions, rollback notes, and an approval checklist.

6. Require human approval

MADD does not change production. It prepares reviewable guidance. Humans approve, reject, or promote learnings into customer knowledge packs.

The build story

AI building AI, with humans keeping the keys.

The MADD journey is also becoming a public build diary. The control centre will record key events: ideas, decisions, tool choices, coding-agent prompts, review outcomes, lessons learned, and cost-aware AI notes.

The story is intentionally nerdy. It is about humans and AI building a Maximo diagnostic assistant together, while leaving a trail future humans — and probably future bots — can understand.

The point is not to pretend AI magically built everything. The point is to show the actual operating model: consensus, routing, prompts, guardrails, human approvals, and evidence.

MADD Control Centre

The orchestration layer.

The planned MADD Control Centre will act as the command centre for the build. A task is described once, converted into a project-memory-backed brief, evaluated by OpenAI and Claude, then routed to the most suitable coding tool.

If the models agree, the system recommends the tool. If they disagree, a human decides. Tools with proper API, CLI, MCP, GitHub, or headless support can be dispatched directly. Tools without reliable automation use honest manual handoff.

Every major decision can become part of the redacted public story, without exposing customer data, credentials, or raw private logs.

Maturity roadmap

How MADD grows from useful to difficult to ignore.

Phase 1 — Automation Script Root Cause

Analyse BMXAA7837E, psdi.mbo stack traces, launch point configuration, MBO lifecycle issues, and DEV/TEST differences. Produce safe solution packs.

Phase 2 — Escalation & Cron Diagnosis

Inspect escalation state, reference points, cron task instances, schedules, where clauses, and environment differences to explain why scheduled logic is not firing.

Phase 3 — License & Access Optimisation

Use MAXUSER, GROUPUSER, MAXGROUP, APPLICATIONAUTH, LOGINTRACKING, and usage signals to support planning and optimisation, not legal compliance claims.

Phase 4 — Integration & Object Structure Diagnosis

Trace MIF, REST, OSLC, endpoints, external systems, publish channels, enterprise services, and object structure issues with evidence-backed recommendations.

Phase 5 — Patch/iFix Impact

Compare IBM release notes against customer fingerprints and customisations to identify likely impact areas, retest scope, and implementation risks.

Phase 6 — Deep Class Intelligence

Where authorised, inspect class metadata and relevant runtime behaviour in memory. Summarise diagnostic implications without storing or redistributing proprietary source.

For Maximo teams

What MADD should eventually feel like.

Ask: “Why is this work order status automation script failing in TEST but not DEV?”

MADD should inspect relevant evidence, compare configuration, parse the stack trace, identify likely MBO lifecycle issues, generate validation checks, and give you a testable plan — not a hand-wavy answer.

For AI builders

The OverpayingForAI lesson.

The future is not one massive model answering every click. The better pattern is an orchestrated bench: deterministic skills, runtime evidence, model consensus, tool-specific execution, human approval, and public learning.

Use AI where it adds leverage. Do not pay it to remember things your system already knows.