Shashi Prakash

Gypsum Board Manufacturing
 
A gypsum board manufacturing plant uses an AI-based production designed to monitor, controls and simultaneously optimize multiple key parameters like board thickness with a target of 12.5 mm, surface finish, and moisture content by analyzing data from sensors. This AI based system covers all the stages of gypsum board manufacturing i.e. slurry preparation, forming, drying, and cutting sections.
 
The key functional parameters which constantly monitors, controls, adjust to optimize are slurry feed rate, additive dosage, line speed, and dryer temperature. Since these parameters directly influence product quality, yield, and significant energy cost, the system must be able to self-audit its own accuracy, bias, and reliability while staying transparent and accountable to operators and management.
 
AI Self-Audit Framework
This self audit system operates within three interconnected layers:
1. Model Integrity  2.Data Integrity 3.Decision Integrity
These three models work in a perfect sync so that AI always remains precise, traceable, auditable, and aligned with plant goals.
 
1. Model Integrity Audit
This ensures that statistical health, accuracy, and fairness of AI prediction and control models are constantly monitored and controlled.
(A) Continuous Performance Tracking:
On real time basis AI compares its predicted control settings like slurry flow, line speed etc against actual outcomes like board thickness and density.
 
There are KPIs such as Mean Absolute Error (MAE) and Process Capability Index (Cpk) are tracked on real time basis and at the end of every shift to observe any drifts.
If Cpk drops below 1.33 or MAE rises by more than 10%, then the system triggers a Model Drift Alert.
 
(B) Adaptive Learning Validation:
Regularly at a predefined interval or post failure alarms AI before any model self-tuning or retraining. The AI runs simulation exercise using historical data to confirm that new parameters would have improved yield and quality. It gives more weightage to most recent data and reduces weightages for older data. 
 
It studies patterns in the data using EWMA and CUSUM charts. It also uses Z-MR charts for shorter runs along with pre-control charts for very simple interpretations. Retraining of AI only proceeds only after supervisor authorization.
 
(C) Bias Detection & Self Correction in the AI model:
It’s possible for an AI model to over-optimize like consistently prioritizing thickness control at the cost of sacrificing the expected surface smoothness or efficient energy consumption. Hence, a fairness score is computed based on internal metrics to ensure balanced performance across all CTQs (Critical to Quality parameters).
 
2. Data Integrity Audit
This ensures that all process data and sensor readings remain reliable, complete, and consistent.
 
(A) Sensor Health Monitoring:
The AI runs SPC (Statistical Process Control) and Z-score analysis on each key sensor like temperature probes, density meters, level transmitters.
Outlier or drifted sensors are automatically isolated, and operators receive a “Sensor Suspect” notification for recalibration.
 
(B) Data Completeness Checks:
Missing or delayed sensor data triggers predictive gap-filling using regression models — but such interpolations are flagged for review.
If missing data exceeds 5% of any critical variable per batch, the AI pauses autonomous control and defers to manual mode.
 
(C) Material Batch Correlation:
The AI validates whether deviations correlate with raw material changes like gypsum purity or starch quality.
This helps ensure process variance due to input parameters isn’t misinterpreted as functional or system drift.
 
3. Decision Integrity Audit
This validates that AI control actions remain explainable, traceable, and aligned with quality objectives.
 
(A) Human-in-the-Loop Control:
Any major automatic adjustment like >3% change in line speed or additive flow must be acknowledged by a production supervisor.
The system logs the decision, timestamp, and operator ID for traceability.
 
(B) Work Performance Reports:
Work performance data goes through expected performance level wherein variance analysis is carried out and any negative variances are highlighted via Work performance report with proper reasoning as to why a particular corrective measure was taken.
 
“Line speed reduced by 2.5% due to rising slurry density i.e. 1.12 g/cc to 1.18 g/cc and increasing moisture variance by 1%.”
 
This builds operator confidence and audit transparency.
 
(C) Self-Validation Sampling:
At a fixed interval of 4 hours i.e. twice in every shift of 8 hours, a random selection of 5 boards are picked up for manual review for the key KPIs like thickness, weight, visual defects, smoothness etc.

Post manual review the AI predictions are compared with manual readings to compute a Validation Accuracy Index.

If this index drops below 95%, the AI triggers a Performance Review Mode.
 
Summary
In nutshell, an AI model self-audits it’s performance by making use of this three layers sytem which is
•            Monitoring model drift and prediction accuracy,
•            Ensuring sensor reliability and data integrity, and
•            Verifying that autonomous decisions remain transparent and traceable.

By ensuring that there is a manual review wherever required, explainable logic and process KPI monitoring and controlling on real time and periodic basis, the system stays responsible, auditable, traceable and aligned with both quality and cost targets.

This framework represents a practical path to Responsible AI in manufacturing — where intelligence enhances control precision without compromising accountability.
 

AI Powered Investment Portfolio Management
Wealth management firms across the globe use different AI-based portfolio advisory system that integrates and synchronizes with multiple data sources  like live market feeds, macroeconomic indicators, client risk profiles, social sentiment data, and geopolitical signals — to dynamically rebalance client portfolios.
 
100% correct data is a myth:
·      Market sentiment shift predictions swings either which ways after unexpected or expected RBI announcements.
·      Social sentiment analysis returns incomplete / biased result.
·      Inflation updates lags behind real-world trends.
The true challenge is that an AI model must decide whether to rebalance client portfolios immediately, estimate the impact of missing variables, or delay recommendations for more information to emerge — this could be resulting in risking either opportunity loss or investor exposure.
 
How does Uncertain / Missing Market Data is handled
1.    Using statistics estimate using market-correlated Data:
When a data feed like currency index is not available or missing, an AI substitutes estimates using other correlated variables like commodity prices or VIX - Volatility Index via Bayesian method.
Each substitution is has a confidence score which is reviewed manually and has a traceable metric, ensuring transparency for auditors and compliance teams.
2.    Relying upon confidence score based rules:
If model confidence >95%, AI proceeds is programmed to continue with automated minor portfolio rebalancing (e.g., shifting 5% from equities to government security bond or gold ETFs).
Between 75–95%, AI flags it for financial analyst review.
Below 75%, the system delays rebalancing and triggers a “market alert” escalation to the risk desk.
 
3.    Flag for Human Review:
In ambiguous cases like conflicting social vs. technical data, the AI model requests validation from the portfolio manager / financial analyst.
 
4.    Delay and Contextual Re-assessment:
The AI temporarily freezes automated actions while running scenario simulations using historical analogs like similar price behavior during past inflationary cycles.
This prevents knee-jerk decisions that could amplify volatility during flash crashes or liquidity shocks.
 
Safeguards for Responsible AI Decision-Making:
 
1.    Dynamic confidence scoring: All model outputs include uncertainty flags visible to traders.
 
2.    Human-in-the-loop compliance: No high-value trades occur without validation when data completeness <90%.
 
3.    Continuous learning loop: After data validation or market closure, AI adjusts weights to improve resilience to incomplete data next time.
 
Impact and Market Responsiveness
In a market data that is fragmented or volatile, the responsible AI system behaves prudently but adaptively considering:
It only acts decisively when data confidence is solid i.e. confidence >95%.
It requests human review when uncertainty is high i.e. confidence 75% - 95%.
It self-adjusts model weights to learn from new market patterns if confidence is <75%.
 
Conclusion
In the financial sector, an AI managing portfolios under market turbulence must balance speed with accountability. The right combination of probabilistic estimation, confidence score based escalation, and human verification prevents reckless trading saving investors from potential unprecedented loss while preserving agility — ensuring resilient, ethical, and market-aware AI decision-making in an unpredictable economy.
 
 

AI-Based Quality Inspection for Tea Bags – Ekaterra Lipton - UAE
I am sharing this personal experience from my training & consulting experience at Ekaterra Lipton – UAE. Like most of the modern manufacturing plants, Ekaterra Lipton UAE also uses AI-powered vision systems to inspect tea bags for defects such as empty tea bags, torn/damaged tea bags, missing tags, smears etc. These tea bags once filled roll over the conveyor belt before packaging wherein the screener constantly scans every tea bag. Their inhouse AI model assigns each item a defect probability score (0–1) based on image analysis.
 
Confidence Threshold Logic
AI Confidence Level
System Action
Rationale
> 0.90 (High confidence – clear defect or no defect)
AI acts automatically — accept or reject the item
Their model has historically achieved >95% accuracy in these cases, allowing fast throughput without delaying the process.
0.5–0.90 (Moderate confidence)
AI flags item for human re-inspection
Mixed visual screener indicators (e.g., slight smear, tags) which reduces certainty to accept or reject; a quality inspector validates the decision.
< 0.5 (Low confidence)
AI pauses and escalates for manual inspection and potential model retraining with new information
new defect patterns, Image noise, or faulty sensors make autonomous action risky.
 
Factors Influencing Confidence Thresholds
Sensor Conditions & Calibration
The most common reasons are poor lighting conditions, dust accumulation on camera / screener lens, or camera calibration issues can distort images and trigger lower confidence score thereby triggering more human reviews. Past Model Accuracy & Drift
Thresholds are dynamic and it emphasizes by putting weightage on the most recent trend and reduces weightage on older results. If rolling 60 data points i.e. recent performance false rejection rate exceeds 2.5%, the system self corrects & tightens the auto-decision range to prevent wastage. Continuous Learning Loop
Every Human re-inspection results are fed back real time to retrain the model. Over time, this raises confidence reliability and reduces manual interventions.  
Balancing Speed, Accuracy, and Trust
Speed: AI handles routine inspections in real time, maintaining production flow. Accuracy: Ambiguous cases are reviewed by experts, reducing false positives. Trust: Operators see AI as a collaborator, not a replacement — decisions are explainable, auditable, and based on confidence logic. Example
In a tea bag production at Ekaterra Lipton, the AI system detects a torn tea bag.
AI Confidence Score = 0.97 : Clear Auto-reject case.
Another image shows a minor smear (confidence = 0.66) : Flagged for human review.
If confirmed defective, the feedback improves future accuracy.
 

Process chosen: I have personally worked with a couple of companies dealing with U.S. debt collection — credit cards, health bills, and loans. This is a highly regulated industry, as these companies must comply with Federal Laws, State Laws, the Fair Debt Collection Practices Act (FDCPA), TCPA and the Health Insurance Portability and Accountability Act (HIPAA). Violations can result in severe financial penalties of up to $500,000 for class-action lawsuits, plus actual damages and attorney fees. Attorney General offices and the Better Business Bureau (BBB) typically handle initial escalations from complainants before the matter moves to legal remedies.
Conflict: Maximizing collection efficiency (contacts per hour, reduced AHT, automated dialing) vs. preserving legal & ethical compliance and consumer dignity (FDCPA/TCPA requirements, state rules, empathy for vulnerable consumers).
Why this conflict is realistic
Offshore operations are measured on throughput and recovery. At the same time, U.S. federal law (FDCPA) and the TCPA impose strict rules on when, how and with what consent collectors may call; many states E.g., Illinois: Limits calls to seven per week per debt, Michigan: Also applies a “seven calls per week” rule, New York: Restricts collectors to seven calls in a seven-day period etc mandate an additional layer of requirements and disclosures. An overly aggressive AI dialer that optimizes contacts/AHT can therefore trigger statutory violations, damaging reputation and incurring statutory penalties.
How AI should decide — practical decision logic
Hard legal constraints (non-negotiable rules). Do not place autodialed or prerecorded calls without the appropriate TCPA consent for that number/type. If consent cannot be verified, do not autodial. (TCPA / FCC rules). Respect FDCPA safe-harbor times (assume convenient time i.e. 8:00 a.m.–9:00 p.m. consumer local time unless affirmative info otherwise) and always include required initial-contact disclosures (the “Mini-Miranda”) where applicable. Enforce consumer-specific flags: DNC, cease-and-desist, attorney representation, bankruptcy, time-zone mismatch — these must block calls immediately. Soft operational objectives (optimize within constraints). A constrained optimizer maximizes recovery KPIs (contact rate, promise-to-pay, AHT) subject to the legal constraints. The optimizer receives a dynamic risk score per call (combining consent certainty, debt age, consumer vulnerability flags, state of residence, prior complaints). Calls with low legal certainty get lower priority or are routed for manual contact. Risk-tiered runtime behavior. High legal risk or sensitive signals (no express consent for autodialing; detected emotional distress; regulated debt type): route to manual agent with a compliance-approved empathy script and supervisor on alert. Medium risk: attempt manual, human-initiated call (no previewed robocall), or send written validation notice first. Low risk + verified consent: allow optimized autodial + limited automated agent assist (script suggestions, real-time compliance prompts). Conservative default & “hard-stop.” Whenever the model’s confidence in a legally material fact (consent, timezone, identity) is below a threshold, the system defaults to the conservative behavior: no autodial, route to human, and log an audit event. This avoids costly TCPA/FDCPA exposure. Explainability & audit trail. Every decision (why a number was dialed, why AI suppressed a call, which script was shown) is logged with immutable timestamps, consent evidence, and model scores — essential for internal QA and external defense. Human oversight — who does what, and how often
Compliance Council (daily): reviews automated alerts (attempts blocked for legal reasons, flagged voicemails, consumer complaints), approves emergency rule changes, and signs off on changes to dialer aggressiveness. Legal team (weekly / on change): keeps rule engine synced with federal/state updates and court decisions that change TCPA/FDCPA interpretation. QA / Supervisors (real-time daily): sample calls, approve agent overrides, and convert representative overrides into retraining labels so the AI learns human judgment patterns. Model Governance Board (monthly): reviews performance vs. safety KPIs and decides whether to relax or tighten soft objectives. Measurable governance & feedback loops
Track Compliance Exception Rate (#calls blocked for legal risk per 1,000 attempts), Legal Incidents (regulatory complaints / suits), AHT, contact rate, and consumer complaint rate. Use a dual-threshold alerting system: immediate stop for spikes in complaints; slower policy review for drift in other KPIs. Simulate policy changes in a sandbox (play back historical call logs) and run a legal-scenario stress test before production rollout. Where legal precedent is shifting (e.g., TCPA interpretation), increase conservatism until legal clarity is restored. Final Conclusion
Treat compliance and consumer dignity as a non-negotiable constraint and efficiency as the objective to optimize within that safe set. AI should be the consistent enforcer of the rules and the suggestion engine for efficiency — humans remain the arbiters of judgment, upgrades, and legal accountability. When in doubt, do not dial; escalate.

Process Chosen: Underwriting Process for Life Insurance Policy
 
Globally there has been a debate over right AI model applicable for the underwriting process for Life Insurance Policy.
An Actuarial AI model used by Life Insurers is required to perform risk assessment and flag high risk profiles based on various factors like:
1.     Pre-requisite medical check-ups conducted at Insurance Company appointed labs which includes blood samples, X Rays etc.
2.     Evaluation of previous medical history of the applicant which includes being admitted to a hospital for a list of ailments.
3.     In addition, the declaration done by a person in the application like lifestyle preferences like if he / she consumes alcohol, tobacco etc. If he or she exercises or not etc.
Human underwriters are then required to review and analyze the indicators flagged by the Actuarial AI model. The final decision is made by human underwriters specifically in the borderline or complex cases, by applying qualitative judgments before approving, modifying, or rejecting policies.
 
What AI Should Learn from Humans:
Compensating & Changing Circumstances: There are cases wherein an underwriter approves an applicant with high-risk factors where they see a controlled chronic illness for years (Eg. An applicant has heart related issue but has been stable & controlled for 5 or more years), radical change in adoption of good lifestyle preferences for recent years (Eg. An applicant used to consume alcohol and smoke but has left it for more than 5 years or more). These compensating & changing circumstantial factors is what AI can learn instead of automatically flagging them as high risk. Handwriting & Documentation: Humans can interpret ambiguous medical reports and seek clarity if not fully understood. Handwriting of doctors, paramedics, nurses notes can be interpreted by humans easily and they can seek clarity if not fully understood. An Actuarial AI model will run its full course of action and end up misclassifying an ambiguous lab report or poor handwriting. Training an Actuarial AI model on these insights can improve reliability of AI model. Ethical Factors: An Actuarial AI model can learn incorporating ethical factors like fairness, anti-discrimination against a certain gender, race, religion based on past learnings of the model, or approving a policy for a high-risk applicant with additional conditions. What Humans can learn from AI:
Hidden Risk Patterns: AI can identify subtle correlations between lifestyle, lab results, and long-term mortality risk that may not be obvious to human underwriters. Consistency and Benchmarking: Reducing subjective variability in policy decisions and to standardize risk scoring across underwriters can be achieved by Humans using AI recommendations Predictive Efficiency: AI can suggest optimal premium adjustments or policy terms based on large-scale actuarial analysis, helping humans make faster, data-driven decisions. Continuous Learning Loop Implementation:
Every underwriting decision (AI recommendation vs. human override) is logged for model retraining. Monthly calibration meetings review high-impact discrepancies, annotating reasons for approval, modification, or rejection. Underwriters receive dashboards comparing their decisions with AI risk scores and rationales, enabling reflection and skill growth. Outcome:
Over time, AI develops a more nuanced understanding of applicant risk beyond pure data patterns, while human underwriters enhance their data-driven decision-making and consistency. This co-learning loop reduces underwriting errors, improves risk-adjusted profitability, and accelerates policy processing with greater fairness and transparency.