Prediction timing and executive scrutiny6 min read
The model had read what happened after the fault
A separate test population cannot rescue a forecast supplied with post-event repair knowledge. The decision moment belongs in the definition of the task.
Alert evaluation and service capacity · 2026-07-16
Wind-service enterprise
420 staff
₹184 crore FY26 company revenue
Evidence and analytics
47 data staff
Four heads; ₹9.3 crore FY27 function expense
Reliability and data leadership
27 years
July 1999 entry; CDO since January 2022
The model had read what happened after the fault
I am Anamika Veqzari, chief data officer at the fictional Qenvathi Wind Services in Coimbatore. In 2022 I approved a validation account that appeared to support a predictive-maintenance proposal. The reported score was 91%. The dataset included repair-closure fields completed after the event we were claiming to predict. Dividing records into development and test groups had not removed that future knowledge. I withdrew my endorsement, funded ₹67 lakh of reconstruction and supported holding the ₹2.8 crore rollout proposal. We needed to establish what the model could have known at the intended decision time before asking anyone to trust its forecast.
My work began in reliability studies in July 1999, after a mathematics degree. I joined Qenvathi in September 2011 and became CDO in January 2022. The company has 420 employees and ₹184 crore FY26 revenue; my remit comprises forty-seven data colleagues through four heads, ₹9.3 crore FY2026–27 expense and a separate ₹3.1 crore evidence programme. An employee management trust holds 58% and an industrial-services investor 42%. I own no shares. Imran Zelqari leads the business, and Ritu Nezvethi leads engineering. Our data office supports their decisions without assuming responsibility for turbine operation, repair approval or safety judgements.
The other decisions here follow the record beyond a model score. We retained the observation periods of components that had not failed, rather than calculate fleet life from failures alone. We reconstructed component identities when a turbine position had hosted more than one physical unit. We examined the workload behind a rare-fault alert system whose high overall accuracy concealed many false alerts and some missed events. Each case required a specific account of what the evidence supported. A reliable-looking total could still answer a different question from the one the service team needed to decide.
A test split did not remove future knowledge
The 1,200 labelled episodes were divided into 900 development and 300 test records, but post-event repair fields contaminated the claimed prediction task. The 91% score was withdrawn.
Unfinished observation was part of the evidence
Among 360 component histories, 54 had observed first failures, 246 reached the common cutoff without that event and 60 left observation earlier. Their different histories remained explicit.
An alert carried a workload as well as a score
In a separate 10,000-episode review, 600 alerts contained eighty confirmed events and 520 false alerts. Twenty events were missed. Those counts mattered more than an unqualified accuracy headline.
Current focus
Maintaining the connection between physical component histories, prediction timing and the practical questions asked by engineering and service colleagues.
Conversations I welcome
A first board contribution on analytical evidence and management judgement, with preparation, employer consent and realistic time commitments established first.
Beyond the page
The record behind the perspective.
Formation
Bachelor of Science in MathematicsDorzevi Mathematics College · 1999
Anamika brought the withdrawn prediction claim back to the executive discussion with its cost and unfinished questions intact. That made the next investment decision clearer.
Imran ZelqariChief Executive Officer, Qenvathi Wind ServicesThe full record ↗
Contact
A first board contribution on analytical evidence and management judgement, with preparation, employer consent and realistic time commitments established first.