Say Hello to Ask Steve

 TLDR: 

Sometime back, rather than writing a book about Steve, I ended up building a custom GPT, what would Jobs do in ChatGPT. It was bugging me, it was behind chatGPT and not with a clean UI chat interface. No more, finally, I have Ask Steve on lovable powered by Gemini at the end. 

For all the common folks, it is a chat window to ask Steve Jobs what you want and hear from him based on 100 plus hours of video transcripts, mostly his interviews and talks 100+ articles written from the days of Apple launch and definitive 13 books written on him.

For all the AI nerd bros end-to-end RAG application with a hybrid search architecture, optimizing vector embeddings (Gemini/pgvector) and keyword fallback. 

Yes, yes, I  can finally say, You Know I am something of a AI Native builder myself. well sort of ;)

Recap/Evolution

I am a big fanatic fan of Steve Jobs, I ended up writing his biography in Tamil. I couldn't help but get mad whenever Steve Jobs and bad behavior of leaders and founders were referenced in the same breath. 

I would scream aloud in the head, "this is not what he intended, what he meant and this is not what you learn from him or you are totally wrong or why use Steve as an excuse for your A behaviour"

I wanted to write a book on learnings. It stuck me in these day and age, rather than a static book, it would be good to have a dynamic, customized, tailored, unique AI

I ended up building a custom GPT on chatGPT. more on that here

When I wanted to write on how I built it and how to build customGPTs, I learnt, it could be more powerful, If I can build an end to end RAG application. 

More than that, the thing bothered me was not having a better UI as chat interface. 

So I started wondering, how do i have a good front end and a proper back end.  and the result is https://asksteve.lovable.app

How is it has been built?

I had a collection of 13 books, 100+ articles and videos around 100 plus hours. I had all of them in text and used Notebook LM to create the core brain. Basically, more like a synthesis of all of them. Ideally all of them could be chunked as well. However it might not be efficient as it would mean AI is looking at the repetitive stuff. Rather it is critical to have a high density, single perfectly synthesized chunk. In other words, it is essential  to pre-distill the data.

I have fed the core files, and source of articles and video transcripts. Using the books as such or chunking them might also lead to copy right issues. The 13 files have been converted to 3140 chunks and is stored in supabase (postgreSQL). To make this library searchable by "meaning" rather than just "words," have enabled the pgvector extension. This allows the database to store embeddings—mathematical vectors generated by the Gemini API that represent the semantic essence of each text chunk. Then used a recursive chunking strategy to break down the massive books into bite-sized pieces of roughly 1,000 characters, ensuring the AI can pinpoint specific stories or principles without getting lost in the documents

The "intelligence" of the system comes from its Hybrid Search architecture. When a user types a question, the backend doesn't just look for exact word matches; it performs a dual-track search. First, it uses Vector Search to find chunks that are conceptually related to the query (e.g., finding "craftsmanship" when you ask about "quality"). Second, it runs a Keyword Search (Full-Text Search) to catch specific names or historical terms like "NeXT" or "Xerox PARC" that might be mathematically blurred in vector space. By merging these results, we achieve a 99% retrieval precision, ensuring the response is grounded in actual facts rather than AI hallucinations.

When you hit "Send," a Supabase Edge Function acts as the traffic controller. It immediately converts your question into a vector via Gemini, queries the database for the top 10–15 most relevant chunks, and bundles them together. This "Knowledge Package", consisting of the Steve Jobs Persona instructions, the user’s question, and the retrieved primary source text, is sent to Gemini 1.5 Flash. The LLM then synthesizes a response that is blunt, direct, and "Insanely Great," streaming it back to your screen in real-time. This "Closed Loop" ensures that Steve isn't just guessing; he is effectively "reading" your curated library to answer you

what remains the same: Both across the custom GPT and the Gemini powered RAG, the data, the source and system instructions remain more or less the same. However, the way the data is processed or treated is different. Here it is a proper end to end RAG powered by Gemini. Yet it should be fun to see how they behave differently

Key Takeaways

In the world of AI, there are lot of things for some one to know. To make things worse, things evolve and change at a super fast pace. Both the quantum and the speed of change is just impossible to keep up.

What would really give the edge is, knowing what needs to be done/built, why it should be built, and how it should be built matters the most. (when I say how it means, the behavior, the output, rather than the tech stack)

If someone knows them well and if they have critical thinking and are good at prompting, they can build anything in the best possible way. I had to ask the LLMs what other approaches are there? why this? why not that? what are the pros and cons and it walks you through the trade offs and you can make a choice accordingly [For e.g the keyword vs vector search, why not to use books, even if you have copy right, at what point more data or training material doesn't move the needle much]

In fact lovable did the chunking by itself, i had to say use this method. It is not that I need to know this method, I need to know enough to ask the LLMs for it to tell me about possible methods and how to make it happen. In other words, if you are little tech savvy and if you can speak english and follow instructions, building most of things are child's play. [There are few applications, I am having a tough time to build them]

There is more than one way to skin the cat. I could have used Claude Code to build or google studio to do it end to end. I could have used ChatGPT API rather than Gemini API. To make it simpler, hey you could use AWS or GCP or Microsoft cloud but each comes with its own flavor. It is imperative to know which one would for you. Functionality, cost and so on.

More than that, the most important think is your source material, the data, how you organize it, how you train your LLM to deal with it. Again IMO, each LLM behaves differently. There is a variation in their flavors.

Building this also gives you the appreciation or make you wonder or appreciate, how difficult or challenging to would be to build something gemini or claude or chatGPT. or even a regular enterprise AI application with safeguards. (for e.g in the beginning, the feedback I got from my friend was it is mean and verbose.)

Who is Satoshi

 The Indian movie enthusiasts had to wait for 22 months to know the answer, "Why Bahubali killed Kattappa"

It took 16 months to unmask the Fake Steve Jobs blogger!

Crypto enthusiasts have been waiting for more than 17 years to know, " Who is Satoshi Nakamoto?" the bitcoin creator and author of the white paper on bitcoin

John Carreyrou of New York Times has written an article claiming to unmask, Satoshi and says, Adam Back, Co-Founder/CEO of blockstream.com is Satoshihttp://

However, Adam has gone on record denying it.

If I am Satoshi, (or for anyone who is Satoshi) it would take an immense will power and effort to not boast about such power and influence for so long.

Think about it, if you are Batman or Superman, though you want to keep your identity secret, wouldn't you at least wanna tell a select few. How come folks can keep quiet, if they know you are super man or bat man!

Neverthless, read this wonderful article https://www.nytimes.com/2026/04/08/business/bitcoin-satoshi-nakamoto-identity-adam-back.html and i also think and believe, certain secrets should stay secrets and this is one of them

Finfluencer vs AI tech Bro

 Last week, after building the REIT InvIT planner, I did share it few of my friends. Well actually, all of my friends. 

One of my friend responded saying, Nice, you have become a Finfluencer.

For a minute wasn't sure, should i be offended or feel proud. 

The whole point of it was to showcase how AI makes life easier, how it is easy to build something super fast and super easy and rather than be called as AI tech bro, me getting labeled as finfluencer was little upsetting.

Then it hit me! It was indeed a compliment

One of my mentor used to say something like this: True tech should be like magic, One should never know the intricacies and they should see and appreciate the output. (probably he was channeling his inner Steve Jobs ;))

Though the application was built using AI and leverages a lot of AI, the tech was invisible. In fact only one person asked me, "Oh you have built it, what stack you are using and how and where you are getting the data"

With the rest of the others, the conversation was about to what are these assets, how to invest, to invest or not, how much to invest and so on.

Looking back, in a way mission accomplished!

Embracing and Adopting AI

 Sometime back was talking to a senior executive from a traditional company. The talk veered towards AI and the impact/disruption due to that. During the conversation, he was wondering, how to convince stakeholders in firms that are not necessarily tech first but are one focused with accuracy, precision and reliability.  I did suggest, how they could think off an overall AI strategy, Identifying use cases with low impact low risk, clearly calling out boundaries in terms of where AI could be used versus not, rethinking their enterprise risk management with AI in focus, having sufficient guard rails, human in the loop design systems and so on.

In hindsight, I realized I should have told him the below as well

1. AI is not just GenAI or LLM. (where the repeatability and reliability could be challenges). There is lot more.

2. If the senior/exec leadership team is still worried about using AI and if they are not wondering how could they use AI and how could they leverage AI and are not proactive and taking initiatives to be AI literate, they are in for far more bigger challenges

3. The real nuanced challenges are more about data protection, privacy, security, and managing legal, risk and compliance.

Having said that, I was also reminded by what one of the tech person who is building/integrating AI features told me recently.

There are two biggest challenges in AI

People in senior level who have no context or knowledge about the particular domain prompting AI to ask for tough questions and make a critique and parroting that to show they are smart while putting down people. Think of a finance leader commenting on the choice of a tech stack!

People in senior level playing around with AI, building a pet project or POC over the weekend and thinking it would apply to production grade, large scale systems as well and making insane requests and timelines. 

Building a AI Agent to analyze REITs and InvITs

 Ages ago, I came across the news that the minimum threshold of 100 units for REITs to buy have been done away with. I was curious what is REIT? Why the limit size and so on. There were only 3 REITs then. 

I was fascinated by them. I felt it might be interesting cause it helps to diversify, it helps to own a piece of real estate and regular income and yields were better than rental yields. (Think buying a flat and renting out) In addition, potential for capital appreciation as well in a clean and better manner. (Think about buying a plot and the loan and the hope it will appreciate and the challenge of liquidating and so on)

Though the lot size limit was done away with, I bought 100 units of all 3 REITs. I kept buying them regularly. I wasn’t sure of figuring out which is good and which might have potential and so on. So I bought all the three regularly and more or less in equal amounts. Now and then I would try to search for posts and analysis about them but they were never up to date.

Fast forward now, thanks to AI, generative AI and LLM, I have been building out few things. (How I ended up as a vibe coder is a story for another day) One of the things, I wanted to build was to figure out the retail bond with the best yield. I was struggling and end up facing structural challenges. It was then I realized, I could use that approach to analyze the REITS.

Basically, I wanted to know which REIT could be the best one to pick rather than blindly investing equally in all REITs. 

https://rankindiareitinvit.lovable.app/

So I have built it. Also the dividend yield is only one factor and mostly the past which may or may not help to look in to the future. So, I wanted it more of a strategy engine, what If i am keen on future potential, what If i am looking to reduce the risk, what If i have to Optimize for the dividend yield (again based on the past performance)

During the process of building it I learnt more than few things.

One, just because prompting and tools make it easier to build and gives a nice looking output doesn’t mean it does what it is intended to do. We should also know what we are building, why we are building and the nuances of it. Domain knowledge, Software engineering, product management matters more than ever. 

For e.g as simple as fetching the CMP was a challenge. In the first cut, it was fetching stale data from yahoo. Calculating Yield is one thing, calculating it the post tax yield is what makes it unique. There are nuances to tax calculation of REITs, in addition all REITs are not unique. In fact, though I was thinking it is better than rental yield, never thought about, hey what is the bench mark to see, if the yield is good (again to be fair, it is not just about yield as there could also be capital appreciation). 

Even when building it, I asked to fetch the parameters from the investor presentation and AI was taking it from the PDF. Considering, I am someone who knows about the industry, I should have asked the AI agent to fetch the data from the XBRL rather than the PDF as it is the optimal and ideal one.

Once I was done with building, the surprising thing was how the REITs were not necessarily doing well as the post tax yield was below the 10Y G-Sec Yield. 

Later on I also added the InvITs to the mix. (Again the taxation for InvIT is different) 

I could have also built this differently. For e.g have only one page with filter to have REIT only, InvIT only or both together. Rather than the current version of 5 pages, - Master ranking, REITs, InvIT, terminology and methodology, and what this is all about. 

I also learnt the following from Lovable: Feedback on how I prompt

• Skip explaining the problem — just state the desired outcome. I can see the code.
• One instruction per sentence. Your multi-option prompts ("do X, or alternatively Y, I prefer Y") cost extra tokens exploring options you don't want.
• Reference by component name when possible ("TerminologyCard: use inline badges") rather than describing the UI.
• Batch related changes in one message with numbered bullets — your audit prompt was a good example of this.

Roughly, your prompts averaged ~2-3x the minimum needed length. Tighter prompts = fewer tokens per message in build mode.

What next?

I could make it open source, I could build a admin panel or one with user input, where people give the inputs as country, taxation, local nuance, basically say where to look for the APIs to fetch the CMP, NAV and the returns, and it builds the engine for that market. 

I can try and build a plugin to calculate the taxation and make it easy to file the taxes in India.

Few other things that might be of interest

I had around 18-22 high intensity prompts. Probably would have spent 120K-150K tokens and have spent 8-10 hours. I could have achieved the same in 30% fewer prompts. If upfront, I had thought a lot about how to build, what to build, say i want both REIT and InvIT, I want the table to look this way, these are the common factors, this is how I want to visualize and so on.