Priya had been working with machine learning models in production for almost three years when she noticed something uncomfortable. She was busy, she was trusted with bigger systems, and yet her ability to work out why a model was wrong had barely moved since her first month. The hours were there. The improvement was not. If you are building a career in artificial intelligence and AIOps, her story points at a pattern worth borrowing: deliberate practice, meaning short drills aimed at a single weakness, with honest feedback and a small change afterwards.
The plateau nobody warns you about
Priya’s days were full of model reviews, pipeline fixes and dashboards. She could ship a retrained model on a Tuesday afternoon without drama. But when someone showed her a batch of wrong predictions and asked “what is going on here?”, she circled. She would try three theories, run a few queries, and eventually land on an answer by elimination.
That is the quiet trap in technical careers. Work gives you exposure, and exposure gives you familiarity, yet familiarity is not the same as skill. Most of Priya’s years had been spent repeating things she already did acceptably, which is pleasant and almost useless for growth. The skill she lacked, diagnosing error patterns quickly, was precisely the one her days never forced her to isolate.
Choosing one skill instead of “getting better at AI”
Her first move was to shrink the goal. “Get better at machine learning” is a wish. “Look at twenty wrong predictions and name the dominant failure pattern within fifteen minutes” is a drill. She wrote that sentence on a card and stuck it beside her monitor.
This matters because vague goals produce vague practice. A narrow target tells you what to do on Monday morning and, more importantly, tells you afterwards whether you succeeded. When you can say “I named a pattern in nineteen minutes and it was the wrong one,” you have something to correct. When you say “I worked on model quality,” you have nothing.
The drill, step by step
Priya’s exercise was simple enough to repeat several times a week.
- Pull a small sample of wrong outputs from a past incident or a staging run, no more than twenty rows.
- Set a timer for fifteen minutes.
- Write down one hypothesis about the main cause before touching any code.
- Test only that hypothesis, using the sample.
- Record whether it held, how long it took, and what clue she missed.
The discipline lay in step three. Her habit had been to explore until something looked interesting. Forcing herself to commit to a hypothesis first meant every failure told her something specific about her instincts, such as always blaming data freshness before checking label quality.
Where feedback came from
Practice without feedback is just repetition with a hopeful attitude. Priya needed an honest reading of her attempts, so she built three sources.
- The answer key. For past incidents, the postmortem already named the real cause. She hid it, ran the drill, then compared.
- A colleague’s two-minute review. Once a week, a teammate looked at her written hypothesis and said only whether it was plausible and what they would have checked first.
- Her own timing notes. A column in her notebook showing minutes to a correct pattern, which made slow weeks visible.
None of these cost money or required a formal programme. They did require that she look at her own misses without flinching, which is the genuinely hard part. Feedback that only confirms what you hoped is a rehearsal for complacency.
The small adjustment that changed everything
After about three weeks, a pattern appeared in her notes. Nearly every slow attempt began with a hypothesis about the model itself, when the wrong outputs were frequently traced back to how the input was prepared. So her adjustment was tiny: before theorising about the model, spend two minutes checking that the inputs in the sample looked the way the model expected.
That single change cut her typical time to a correct pattern noticeably, and, more usefully, it made her explanations to others crisper. “Here is what the input looked like, here is what the model did with it” is a sentence people can act on. The improvement came from noticing and fixing one habit, not from a month of extra reading.
Adding adjacent drills over time
Once the first drill felt comfortable, she rotated in others, one at a time, always with the same shape: narrow target, short timer, an honest check, one tweak.
- Writing a three-sentence summary of a model change for a non-technical manager, then asking whether it could be understood without follow-up questions.
- Sketching what she would check first if a service’s predictions suddenly changed overnight, then comparing her list with a seasoned colleague’s.
- Reading a monitoring chart and predicting the next few points before scrolling.
She resisted running all of them at once. One focus at a time kept sessions short and the lessons easy to apply. If you want ideas about the wider skills that sit around such drills, the site has a roundup of the data operations abilities that make pipelines dependable, and a walkthrough of first steps into a reliability-focused career for people who are moving across from other roles.
When a person can help you set the drills
There comes a point where a self-run routine has blind spots, because you cannot see what you do not know to look for. Priya eventually wanted someone who could watch her work and propose drills she would never have designed herself. The DevOpsTrainer directory lets you browse DevOps trainers, instructors and coaches by domain, and AIOps and MLOps are among the areas it covers. Treat it as a place to start a search, then judge any fit by how well the person gives specific, usable feedback.
What she would tell a newcomer
Asked later what she would pass on, Priya kept it short.
First, pick the skill that your current job never isolates. Production work teaches you whatever happens to occur, which is rarely the thing you most need to rehearse. Second, keep sessions to fifteen or twenty minutes; attention fades faster than people admit. Third, write your hypothesis down before you test it, because it is the only way to learn from being wrong. Fourth, find a source of honest feedback, even a modest one. Fifth, change one thing per round so you know what caused the improvement.
She also mentioned a quieter benefit. Once the drills became routine, she stopped feeling like an impostor who got by on elimination. She could point to a record: dates, minutes, patterns, and the habits she had corrected. Careers in AI move quickly and the vocabulary shifts every year, but the skill of diagnosing a wrong answer carefully does not go out of date.
If you have been busy for years and sense that your progress has flattened, try one drill this week. Choose a narrow target, set a timer, write down your guess before you look, check it against something honest, and make a single small adjustment. Then repeat it. The compounding comes from repetition with attention, not from repetition alone.