Six days. Six sessions. Seven interviewers. Weibull analysis, ALT, HALT, HASS, SPC, and FIT rates all on the table. That was my prep window for a 6-session on-site interview at a power electronics manufacturer where reliability engineering was core to the role. I'd taught Weibull years ago but hadn't touched it in a while. Facing the prospect of re-reading textbooks and scrambling through old notes, I did something different: I used AI to build a structured prep system that turned six days of panic into six days of focused, fluent preparation. The result wasn't just walking in familiar with the material—it was walking in ready to talk shop with VPs and principal engineers about their actual challenges. This post shows the methodology, not the Weibull content itself.

The old way vs. the AI way

The traditional technical refresher looks like this: dig out textbooks, highlight sections, make flashcards, do practice problems, and hope it sticks. For a topic as deep as Weibull analysis—which spans probability theory, statistical inference, engineering applications, and software implementation—that process takes weeks. You're constantly context-switching between foundations and applications, trying to build mental models from fragmented sources.

My AI-assisted approach inverted that. Instead of starting with textbooks and hoping to extract what's relevant, I started with the specific use case: talking to seven different interviewers about their actual reliability challenges at the company. Then I used AI to structure knowledge retrieval, generate targeted practice, and create interviewer-specific talking points—all grounded in sources I provided. The AI became a query and synthesis engine over a knowledge base I curated, not a replacement for domain expertise.

This matters because AI hallucinations are real, especially with formulas and technical details. If you let the AI generate Weibull parameters or ALT models without verification, you'll walk into an interview confidently wrong. The key insight: AI compresses prep time only when you supply the expertise to guide it and validate its outputs. You're still the driver; AI is just the accelerator.

The prep system architecture

My system had two layers: a zettel cluster as the knowledge base, and AI as the query/synthesis engine. The zettel cluster contained three types of notes:

1. Foundational sources: Key sections from Meeker & Escobar's Statistical Methods for Reliability Data, JMP documentation on Life Distribution and Fit Life by Stress platforms, and my own past teaching notes on Weibull analysis.
2. Interview-specific prep: The interview schedule, interviewer bios, and universal talking points that applied to every session.
3. Structured templates: Live note capture formats, formula verification checklists, and prompting patterns for different types of queries (definition, application, verification).

The AI's role was to query this knowledge base, synthesize explanations, generate practice scenarios, and tailor content to specific audiences—but only after verifying its outputs against the sources I provided. This created a tight loop: retrieve, synthesize, verify, apply.

Step 1: Knowledge retrieval and structuring

I started by feeding the AI the right sources. For Weibull fundamentals, I gave it the probability density function, cumulative distribution function, reliability function, and hazard function from Meeker & Escobar, along with the physical interpretation of shape (β) and scale (η) parameters. For JMP capabilities, I provided documentation excerpts about the Life Distribution platform, probability plots, confidence interval methods, and competing risks analysis.

The AI didn't just regurgitate these—it used them to answer questions like: "Explain how the shape parameter β indicates different failure patterns" or "How would you set up an accelerated life test for electrolytic capacitors using the Arrhenius model?" Each response grounded in the provided sources reduced hallucination risk because the AI was reasoning over facts I handed it, not inventing them.

Crucially, I made the AI show its work. When it explained the bathtub curve as a mixture of Weibull distributions, I had it cite the specific sections where this concept appeared. When it described the Norris-Landzberg model for thermal cycling fatigue, I verified the formula against my sources before moving on. This retrieval-augmented generation approach turned the AI from a potential source of misinformation into a reliable study assistant.

Step 2: The formula verification loop

Formulas are where AI hallucinations hurt most. A misplaced exponent or missing parameter can turn a correct concept into dangerous nonsense. My solution was an explicit verification loop: whenever the AI generated a formula, I immediately checked it against my trusted sources before accepting it.

For example, when the AI produced the Weibull hazard function as h(t) = (β/η)(t/η)^(β−1), I checked it against Meeker & Escobar. When it gave the Norris-Landzberg acceleration factor, I verified each term: the 1.9 exponent on ΔT ratio, the 1/3 exponent on frequency ratio, and the exact constant 1414 in the exponential term. When it suggested the Arrhenius model for capacitor life as Life = A·exp(+Ea/kT), I confirmed the activation energy range for electrolytic capacitors (0.9–1.0 eV).

This loop served two purposes: it caught hallucinations before they took root, and it reinforced my own understanding through active verification. Each time I caught a discrepancy—like the AI initially using 8.314 instead of 8.617×10⁻⁵ eV/K for Boltzmann's constant in an Arrhenius calculation—I strengthened my grasp of why the correct value matters. The process wasn't just about getting right answers; it was about building immunity to plausible-sounding wrong ones.

Step 3: Per-interviewer talking points

With knowledge structured and formulas verified, I turned to audience-specific preparation. Rather than generic Weibull review, I created tailored talking points for each interviewer based on their role and likely focus areas.

For the Sr. Principal Reliability Engineer, I focused on statistical rigor: MLE vs. median rank regression, likelihood ratio tests for comparing Weibull and lognormal fits, and how to handle right-censored data in field returns. I had the AI generate practice problems like: "Given 15 failure times and 5 right-censored observations at 1000 hours, write the Weibull likelihood function."

For the engineer championing JMP Live implementation, I shifted to workflow: how to set up a JMP add-in for non-scripters, the training challenges in rolling out statistical tools across 30+ global plants, and how to measure adoption success. The AI helped me structure responses around my actual JMP Expert System—framing it not as an AI tool but as "I built tools that make JMP accessible."

For the joint session with the Sr. Dir CIM and VP/CIO, I prepared to discuss data architecture: how shop-floor test data flows to analytics, the pain points of MES/SCADA systems, and where JMP Live fits as the bridge between statistical analysis and data layers. Here, the AI helped me articulate how my experience building production monitoring systems and data platforms translated to the company's IT/OT integration challenges.

This wasn't about memorizing scripts—it was about mapping my knowledge to each interviewer's worldview so I could speak their language fluently.

Step 4: Live note templates

The system didn't stop at the interview room door. I created structured templates for capturing live notes during each session, designed to turn conversations into immediate action items and follow-up questions.

Each template had four sections:

1. Key points made (what the interviewer emphasized)
2. Knowledge gaps revealed (what I needed to brush up on after)
3. Follow-up actions (specific things to research or prepare)
4. Personal assessment (how well I connected, what to repeat or adjust next time)

After the session on reliability statistics, my notes revealed I needed to double-check the exact formula for likelihood ratio test degrees of freedom when comparing nested Weibull models. After talking with the JMP Live champion about rollout, I added "prepare concrete adoption metrics" to my follow-up list. These templates turned the interview from a one-way interrogation into a diagnostic conversation where I could adapt in real time.

What AI couldn't do

Despite its usefulness, the AI had clear boundaries it couldn't cross—and recognizing these was as important as leveraging its strengths.

First, judgment calls. The AI could explain competing risks models or when lognormal might fit better than Weibull, but it couldn't tell me whether, for a specific capacitor failure mode in the company's actual field data, a mixture model was warranted or if a simpler approach would suffice. That judgment required domain experience—knowing when to pursue statistical elegance versus practical sufficiency.

Second, contextual nuance. While the AI could generate JMP scripting examples for Weibull analysis, it couldn't know which specific platforms (Life Distribution vs. Fit Life by Stress) the company actually had licensed, or whether their IT department preferred point-and-click interfaces over scripting for broad user adoption. Those details live in organizational knowledge, not textbooks.

Third, verification responsibility. The AI could suggest formulas and methods, but the burden of checking them against primary sources always remained with me. Letting the AI "handle" verification would have defeated the purpose—it's like asking a study buddy to check their own work.

Most importantly, the AI couldn't care. It didn't feel the stakes of walking into that interview room, didn't experience the adrenaline of a tough technical question, and didn't have the professional pride at stake in representing my skills accurately. That human element—the motivation to prepare thoroughly, the discomfort of uncovered gaps, the satisfaction of a connection made—is irreplaceable.

The result: From prep to performance

The six days of structured prep translated directly into interview fluency. When the reliability statistics interviewer asked about Weibull parameter estimation with censored data, I didn't just recite MLE—I walked through how I'd handle field return data where some modules were still operating at 18 months, and why profile likelihood confidence intervals beat Wald intervals for the scale parameter η. When the JMP Live champion asked about adoption challenges, I spoke from my actual experience building the Expert System and RAG pipeline, not from generic talking points.

Most tellingly, in the joint session with the CIM director and the VP/CIO, I was able to bridge engineering and IT: discussing how my EDA Explorer platform turned raw manufacturing data into decision-ready dashboards (addressing the CIM director's concern about making test data actionable) while also describing the modern data stack I've worked with (Python/SQL/Snowflake/Azure ML) that spoke directly to the VP/CIO's analytics roadmap questions.

The prep didn't just make me knowledgeable—it made me conversant. I could follow when the Sr. Principal Reliability Engineer referenced storage-industry reliability practices, connect when the HR Director mentioned professional development budgets, and engage when the Sr. Dir Test Eng & Sustaining Eng asked about real-time ML monitoring on test floors. Each conversation built on the last because my prep wasn't siloed by topic—it was structured around the actual human interactions I'd have.

Your template for technical refresher prep

This methodology isn't specific to Weibull or reliability engineering. You can adapt it for any deep technical refresher—whether you're brushing up on Bayesian statistics before a data science interview, revisiting Kubernetes internals for a platform engineering role, or preparing to talk compiler design principles after years in management.

Here's how to structure it:

1. Start with the audience, not the topic. Who will you be talking to? What are their roles, priorities, and likely questions? One sentence per person: "VP cares about quality culture, Reliability Principal cares about statistical rigor, HR cares about cultural fit."

2. Curate a knowledge base of trusted sources: textbooks, documentation, past work, and industry standards. This is your anti-hallucination foundation.

3. Use AI as a query engine over this knowledge base: "Explain X using only these sources," "Generate a practice problem about Y," "How would Z apply to this scenario?" Always make it show its work and cite sources.
4. Implement a verification loop for anything technical: formulas, models, specific numbers. Check every AI-generated detail against your sources before accepting it.
5. Create audience-specific talking points that map your knowledge to their worldview—not scripts, but fluency pathways.
6. Build live capture templates to turn conversations into immediate learning and follow-up.
7. Honor the boundaries: know where AI helps (retrieval, synthesis, practice) and where it doesn't (judgment, contextual nuance, verification, motivation).

The CTA isn't to adopt my specific Weibull prep—it's to adopt the mindset: AI compresses technical refresher prep from weeks to days when you treat it as a structured knowledge tool over sources you trust, not as an oracle. You bring the domain expertise to guide it and verify its outputs. You're still the driver; AI is just the accelerator that lets you cover more ground in the time you have.

And sometimes, that's all you need to walk in not just prepared, but fluent.