interviews on S Anand

Kick-starting a PyConf Panelist Interview

Tue, 17 Mar 2026 08:32:35 +0530

I was a panelist at the PyConf Hyderabad AI in SDLC - Panel Discussion.

After that, one of the volunteers asked for a video interview.

“How was the panel discussion?” he asked.

Ever since I started using AI actively, my brain doesn’t work without it. So, instead of an eloquent answer, I said, “Good.”

He tried again. “Um… how did you feel about it?” he asked.

I searched for my feelings. Again, fairly empty in the absence of AI. “Good,” I said again.

He didn’t know what to do. I did.

“Look, I’m not going to give good answers without good questions. Let’s do this. Why don’t you think of a way of asking that question that will make me give interesting answers?”

To his credit, he tried. He thought for a minute before saying, “I can’t think of anything.”

So I pulled up ChatGPT and dictated this on his behalf:

I have been asked to record panelists at PyCon Hyderabad and give feedback.

I asked one panelist what his feedback was, and he said it was good, but that is a very simple question, and he said I should come up with better questions that will help him answer better.

Now, I have no idea what questions to ask that will come up with better answers. So what am I supposed to do?

Can you give me some interesting questions that will help this panelist give me good answers?

I pasted this into Claude - since ChatGPT has better dictation but Claude has flaire - and handed my phone to the volunteer.

He scanned through it, and his first question was a shocker.

“Was there a moment when you disagreed with a copanelist but stayed silent?” he asked. And then, the interview flowed beautifully.

Postscript: I read Claude’s actual response. The volunteer’s question was not in the list! Based on a quick scan of Claude’s ideas, he came up with that (very good) question on his own.

This validates my AI Advice. AI is great to get you started. It’s great for ideation. It also atrophies your brain - but those are parts many won’t use in the future.

IIM Bangalore PGP Interview Panel

Tue, 17 Mar 2026 06:42:32 +0530

Yesterday, I was part of an IIM Bangalore interview panel at Hyderabad, along with Professor Subhabrata Das and Debajyoti. Panels typically comprise of two faculty and an alumni, and handle 8 interviews in the morning and eight in the evening, though in our case, we had 9 each.

As we arrived, we were given a USB drive with the student’s resume, statement of purpose, and other documents that they had submitted, which included employment contracts, declarations, letters of recommendation, etc., depending on the student. Each interview was approximately 20 minutes. Luckily, Dr Das set a timer for 18, so we didn’t go too far beyond.

There was practically no time to read the documents before each interview, so I used Claude with the following prompt to suggest questions.

In the context of the below, suggest deep, probing questions that will reveal the suitability of the attached candidate for IIM PGP.

Feel free to search online for additional information about the candidate.

…followed by

The instructions email from IIMB
The attachment in the email
Claude’s earlier advice on how to evaluate candidates based on the above. This included suggestions to:
- Probe for internal consistency: “Your SoP says X but your job history suggests Y. Explain.”
- Probe for authenticity: stuff that coaching can’t teach. (It gave candidate-specific questions.)
- Probe for social sensitivity: “You’re from village X. Where does your aspiration help / hurt whihc segments of your village?”
The candidate’s documents (resume, statement of purpose, etc.)

It helped that there were three panelists. That gave me time to screen the questions and understand the good ones - while others asked their questions.

DISADVANTAGE: You have to spend time reading the questions. Not all questions are great. Hallucinations exist (It said one candidate had a claw-back in their employment contract - they didn’t.)
ADVANTAGE: You pick up little-known stuff. One wrote a monthly salary instead of yearly and didn’t know it. The gaps in the resume, the dips in transcripts, etc. surfaced instantly.

Initially, I just read out some questions.
Then I tried using my own judgement.
Then I went back to Claude’s questions.
It took a while to learn how to use it well.

The interview questions idea came from Claude. It also suggested:

Transcribe real time and pop-up questions. I wish I could and I’m sure it’ll happen soon.
Review scores & transcript with Claude. I did almost exactly that!

I told GitHub Copilot CLI (I had credits) running Claude 4.6 Sonnet (a sensible model):

You are evaluating candidates for the PGP Program at IIM Bangalore.
Read interview-guidelines.md fully to understand the process and criteria.

Then, go through each candidate’s folder. These contain:

notes.md: interview notes

documents.md: documents submitted (converted from the PDF in that folder).

Primarily based on the notes (using documents.md for reference - no need to read PDFs), evaluate all candidates on the criteria mentioned in interview-guidelines.md.

Plan like an expert interview evaluator first. In this context, first think about:

What patterns would an expert in this field check / recognize that beginners would miss?

What questions would an expert ask that a beginner would not know to?

What problems / failures would an expert anticipate that beginners may not be aware of?

How would an expert analyze this? At each step, explain what they are looking for and why.

Document and plan your process in eval-copilot-claude/plan.md.
Then, analyze students (use sub-agents as required) and document your analysis and evaluation in eval-copilot-claude/evaluation.md.

It reviewed all 18 candidates summarizing their background, work-experience, and interview scores along with strengths, concerns, and notable moments, with an evaluation summary.

This serves as a good cross-check, I think. So I told it to:

Compare the ratings against my ratings (and notes) in README.md.
Sort the candidates based on the difference between my ratings vs your (normalized / scaled from 1-10) ratings.
For the biggest differences, analyze the reasons for the differences.
Specifically, what might I have missed in the candidate that I should take a closer look at?
Append this to eval-copilot-claude/evaluation.md.

I ran this on Copilot as well as Claude. We agreed on the best candidate. No question about that. Interesting!

This revealed some interesting observations. I seem to:

Ignore social insensitivity and artistic ability. Most likely because I lack both. Note to self: I’m blind to, and hence devalue, what I’m not good at.
Value genuineness far more than analytics. I rated one candidate with terrible analysis skills as only a marginal reject. Claude rejected them outright. This is true for another candidate, too. But I rejected fakers outright.
I calibrate for preparation. I expect more from those with more experience, with more interview preparation, etc. Not a bad thing - just an observation.

In short:

AI is great for preparation. I should make it easier to feed it context.
AI is great at post-mortems. I need to budget time for this.
AI can help during discussions. I need to figure out the technology for this.

Transcript AI-ded interviews

Fri, 20 Feb 2026 18:09:04 +0800

Priyanka was ghost-writing an interview request from PC Quest for Ankor. Two questions were a bit technical:

Straive combines data engineering, analytics, AI, and content services. At a technical level, how are enterprises stitching these capabilities together architecturally and operationally when addressing complex business problems at scale?
GenAI systems tend to behave unpredictably when exposed to real workloads. What engineering patterns, monitoring approaches, or runtime safeguards are becoming essential to maintain reliability, performance, and cost control in production settings?

… and she asked if I could review.

These days, “rewrite” is easier than “review”.

STEP 1: Copy all my meeting transcripts with Ankor. Below is my process, but copy-paste or a coding agent works even better.

1a: Download transcripts with rclone
1b: Convert Word to Markdown with pandoc
1c: Combine using uvx files-to-prompt and copy

STEP 2: Pass Gemini (which accepts 1MB!) this prompt:

Based on the documents below (transcripts of my calls with Ankor),
and the contents of my Google Drive where you'll find material about Straive,
how would Ankor answer this question to PCQuest -
in a way that's aligned with their audience?

Q: Straive combines data engineering, analytics, AI, and content services.
At a technical level, how are enterprises stitching these capabilities together
architecturally and operationally when addressing complex business problems at scale?

(Paste the transcripts)

The answer was a bit long and not very insightful. So I added 2 prompts:

Share this as a 150-word summary suitable for a PCQuest audience.

Then:

Frame this in a way that's insightful and evocative for a business audience.
Think about an approach that would make them think,
"Oh, yeah, that's insightful, and wasn't obvious until Ankor mentioned it."

The final result was quite a good one, reflecting some of our recent conversations:

Most enterprises are making a critical mistake with AI right now: they are treating foundational models like marriages, when they should be treating them like batteries.

If you hardwire your data and content pipelines directly to today’s best LLM, you are accumulating massive ‘AI Debt.’ When a better model drops next year, you’re stuck. The true enterprise differentiator isn’t the model—it’s building an agnostic orchestration layer. This middleware lets you swap out backend models instantly, without breaking your business.

Once you decouple the model, you must stop trying to make legacy, linear assembly lines run faster. Instead, we use AI to ‘squish’ the assembly line entirely. AI co-pilots extract metadata and trigger multiple agents to execute tasks in parallel, not sequentially.

This fundamentally shifts the workforce. We are no longer putting humans in the loop to do the heavy lifting; we are elevating them on the loop as auditors. The AI executes the work and provides the citation trail; the human manages the risk. That is how you bypass theoretical six-month roadmaps and move from ideation to a secure, live enterprise deployment in just eight weeks.

STEP 3: For the next question:

Think about how Ankor could answer this question -
applying the same process, style and approach.

Q: GenAI systems tend to behave unpredictably when exposed to real workloads.
What engineering patterns, monitoring approaches, or runtime safeguards are
becoming essential to maintain reliability, performance, and cost control in production settings?

It again needed some tweaking:

Rewrite: about 150 words, slightly different style
(to make it avoid looking like too close a clone of the earlier answer),
but just as evocative and insightful.

… and the answer, again, was quite close to some of our discussons:

“Most leaders try to fix GenAI unpredictability by endlessly tweaking prompts. That’s like trying to tame the weather instead of just building a stronger roof.

At Straive, we assume models will hallucinate. The true enterprise differentiator is the engineering harness built around them.

We secure workloads through Verifiable Autonomy. Instead of testing in live production, we run agents through simulated environments governed by strict policy-as-code. If an agent can’t deterministically pass the simulation, it doesn’t touch real data.

To scale without exploding headcount, we use AI to govern AI. We deploy a ‘Maker/Checker’ architecture: a creative model generates the work, while a cheaper, constrained model instantly audits it for compliance and accurate citations. Humans only handle the exceptions.

Finally, we stop treating every task like a real-time chatbot. By shifting latency-insensitive workloads to asynchronous batch processing, we bypass rate limits and slash compute costs by half. Reliability isn’t a model feature; it’s an architectural choice.”

The trick to re-purposing content quickly is:

Keep data / docs handy
Learn to quickly search & copy-paste
Learn to quickly judge and give feedback to AI

Sun, 24 Mar 2024 14:39:01 +0000

Oh, wonderful! They’re keen to get in. Wise enough to take help. Honest enough not to be able to cover it up. Sounds like a good hire!

Solving multiple choice questions

Fri, 19 Jan 2007 12:00:00 +0000

How would you solve this multiple-choice problem: What is 12345 x 45678?

201932843
563894910
402394820
384718349
938491834

It always amazes me when people try and multiply the two numbers. In any objective-type test (multiple choice question), the aim is not to solve the problem – it is to pick the correct answer!

Most people don’t seem to realise the difference.

If I had to solve the problem, I’d look for shortcuts. For example,

All answers have 9 digits. So that’s not going to help.
12345 ends with a 5. All multiples of 5 end with 5. So the answer is either (2) or (3).
1. ~~201932843~~
2. 563894910
3. 402394820
4. ~~384718349~~
5. ~~938491834~~
Multiplying the first digits, 1 X 4 = 4. So the answer’s got to start with a bit more than 4. (3) starts with 40. That’s too low. C’mon, the second number STARTS with 45. So… 2. 563894910 3. ~~402394820~~

This takes me 10 seconds. Multiplying the two takes me ten times as long.

Let me repeat the point: You DON’T have to get to the answer. You’ve already been given the answer. You just don’t know which one it is. And you don’t have to solve the entire problem to pick the right one.

There are many ways of picking the right answer. One is what I just used: the answer must satisfy some property. In this case, the answer must be a multiple of 5.

In other cases, the answer must be close to something. For example, what’s 37463 x 28438?

532686397
1065372794
2130745588
4261491176
8522982352

Well, let’s just multiply the first two digits: 37 x 28. That’s 1036. So the answer’s got to be close. OK, (2) is the answer.

These are not one-off techniques, nor are these applicable only to numbers. You can always pick the right choice without solving the problem.

A weaker, but more general version of the above rule is: you can always eliminate choices without solving the problem.

I’ll blog more on this shortly.

Comments

anand 19 Jan 2007 9:47 am: hats off. U made me to think…….
Vinodh 19 Jan 2007 11:16 am: First rule should be to eliminate the options based on the product of the last digits. So, for the second example, it’s easier to eliminate all the other four options based on this rule. Of course, there are easier techniques for generic multiplication/division as well…I have an somewhat interesting technique to calculate squares/products based on simple algebraic formulae…of course they are better used with some practice, but very powerful.
jake 19 Jan 2007 1:44 pm: its easier and faster to get the answer for the second one by looking at the last digit of the answers. it has to be the same as the last digit of the product of the unit place digits of the numbers being multiplied.
Shreyas 20 Jan 2007 8:14 am: “All multiples of 5 end with 5” - I guess it should be “All multiples of 5 end with 5 or 0” :)
Anonymous 20 Jan 2007 7:55 pm: CAN YOU PLEASE TELL ME WHERE i CAN FIND MORE TIPS ON SOLVING MCQ? I LIKED GOING THROUH THIS BUT I NEED MORE PRACTICE. THANK YOU
Anonymous 22 Jan 2007 5:58 pm: There is a much better way to solve Q2. the last digit of N1 is 3 and the last digit of N2 is 8. 3*8 is 24 so N1*N2 has to end with 4. so the answer should be
GV Rao 7 Mar 2007 7:42 am: I would like to multiply first 2 digits: 45 x 78, and in second case: 3 x 8, which is easy, last digits gives more guess on product.
Sri Lankan Friends 10 Mar 2007 10:23 am: hmmm very interesting
ali 24 May 2007 8:16 pm: good but should be more efficent via understanding
vasudha 19 Jan 2007 12:00 pm: HATS OFF TO YOU… THE INFORMATION IS PRETTY GOOD.. YOU MADE ME THINK
deepak 25 Feb 2011 6:20 am: phaltu trick sab hai… issss se kahi behtar tricks mere paas hai…. sirf tukka maar ke IIT me aa gaya hu YES I AM IN IIT…