01
Sense
Collect signals, process data, context, operator observations, and missing-information cues.
Working thesis Public operating model Evidence-bound
Manish Sharma Lab
Industrial AI for Decisions You Can Verify
Sense -> Model -> Decide -> Verify
Industrial AI becomes useful when it does more than produce a confident output. It must stay connected to the signals it observed, the model assumptions it used, the decision it recommends, and the evidence that can verify the outcome.
Four-stage loop
Sense, Model, Decide, Verify
This is a practical operating model for public industrial-AI work, not a claim to own a new discipline.
Operating loop
Sense -> Model -> Decide -> Verify
Signals, assumptions, bounded decisions, and verification evidence.
Confidence is not approval. The loop is useful only when process signals, model assumptions, decisions, and verification evidence stay separate.
02
Model
Combine machine learning, engineering rules, uncertainty, constraints, and traceable assumptions.
03
Decide
Structure recommendations, trade-offs, risk priorities, next actions, and human-review boundaries.
04
Verify
Connect decisions to inspection, measured outcomes, feedback loops, and physical evidence.
Use Sense for observed inputs, Model for assumptions and uncertainty, Decide for bounded recommendations, and Verify for inspection, testing, documentation, measured outcomes, and expert review.
Stage links
Connect each stage to a source, framework, or tool
The thesis is useful only when it routes readers to concrete LMD/DED resources and evidence boundaries.
Sense
Process signals, images, measurements, logs, human observations, and RFQ inputs.
Model
AI, statistical, process, and rule-based models with visible assumptions and uncertainty.
Decide
Repairability, process route, inspection need, risk level, and next-step recommendations.
Verify
Quality evidence, testing, inspection, documentation, release criteria, and expert review.
Principles
Decision support has to keep uncertainty visible
The model is useful only when the decision path remains inspectable.
A signal is not proof.
A model output is not automatically an engineering decision.
Missing information should remain visible.
Confidence should not hide uncertainty.
Human review must be explicit where risk requires it.
Verification must match the consequence of failure.
Useful industrial AI creates traceability, not only predictions.
Interactive explainer
Signal is not proof.
Select a signal. The useful question is not whether the signal is interesting; it is what the signal can suggest, what it cannot prove, and what evidence closes the loop.
Selected signal
Melt-pool image feature
What it can suggest:
possible process drift, instability, spatter, geometry change, or review point
What it cannot prove:
final material quality, fatigue behavior, or service safety
Evidence needed next:
- dimensional inspection
- NDT where risk requires it
- material testing or metallography when acceptance depends on it
- expert review
Confidence is not approval. Monitoring can support review; it does not replace inspection, testing, expert review, or release evidence.
What changes the decision
Monitoring signals need context before they become useful
A signal becomes more useful when it can be tied to inspection, calibration, repeatability, and acceptance criteria.
Correlation with inspection
Drift history
Sensor calibration
Image quality
Process repeatability
Acceptance criteria
Applied proof
Grounded publicly in LMD / DED
The established public proof domain for this thesis is AI for Laser Metal Deposition and Directed Energy Deposition at Exafuse.
Process monitoring
Signals from LMD/DED systems can expose instability candidates, drift, and review points.
Melt-pool signals
A melt-pool signal can support awareness, but it does not prove final part quality by itself.
Bead geometry
Geometry indicators can guide process review and machining planning when tied to inspection evidence.
Repairability screening
Material, damage, access, tolerance, inspection, and criticality must remain visible before a recommendation hardens.
RFQ preparation
Known facts, missing information, assumptions, risks, and next steps should be separated.
Process route selection
LMD, DED, cladding, SLM/LPBF, machining, replacement, or no repair should be compared with explicit trade-offs.
Inspection evidence
Dimensional checks, NDT, material evidence, and expert review close the loop where risk requires it.
Use with
Public resources that make the thesis practical
These links connect the operating model to existing LMD/DED frameworks, evidence, and RFQ resources.
Operating model
Apply the thesis to the established proof domain
LMD / DED Domain HubWhere the operating model is grounded publicly.Open →LMD Quality Evidence LadderSeparate process signals from inspection proof.Open →LMD Repairability IndexScreen repairability before recommendations harden.Open →Full Site MapEvery page and public machine-readable asset in one linked map.Open →