Why We Built What Nobody Wanted to Build.

September 2024. Jio World Convention Centre, Mumbai.

We had no paying customers.

We had been selected from hundreds of applicants by the Government of Maharashtra to exhibit at the 27th National Conference on e-Governance. Two days. A free booth. Jointly organised by the Department of Administrative Reforms and Public Grievances, the Ministry of Electronics and Information Technology, and the Government of Maharashtra. The theme: Viksit Bharat. Secure and sustainable e-service delivery.

The room was not a startup expo. It was not a paid sponsorship. The people walking past our booth were Heads of Department from MMRDA, Defense, Railways, Finance, and the Head of the National Informatics Centre, the organisation that runs India's entire government digital infrastructure. Senior officials from central ministries and state governments across the country.

We were a two-founder company with no revenue, a rebuilt product, and a story about why the foundation had to come before anything else.

What surprised me about those two days was how quickly people understood the problem. Conversations did not start with "what does your product do." They started with "can this solve our workflow issue." The problem we had set out to solve in 2021 was apparently alive and well in the most consequential rooms in Indian governance.

We came close to something significant that week. We could not close it. Not because the product was wrong. Because the company was not yet the right size. That is a particular kind of frustration: the fit is real, the timing is not.

But what that room gave us was something more valuable than a contract. It made security non-negotiable.

Not a feature. Not a selling point. A prerequisite. Government-level data sensitivity meant that every architectural assumption about privacy and encryption had to be hardened. We came back and started fixing for it. The PBKDF2-derived encryption, the domain-level salts, the user-level salts, the zero raw data transmission to any model, the AES-encrypted preview samples that auto-delete on session exit. All of it shaped and accelerated by what two days in that room taught us.

We had been building the foundation. The NCeG conference told us exactly how strong it needed to be.

Late 2024: The Second Reckoning

The first AI reckoning had hit in late 2023. Three months of genuine doubt, the decision to carry forward with zero clarity, and the conclusion that an LLM on top of ungoverned data did not solve the problem we had set out to solve.

The second one hit differently.

By late 2024 AI was not a panic. It was a market reality. Every tool had an AI layer. Every deck led with AI. Every company that had spent years building fragmented stacks was now slapping a chat interface on top of them and calling it intelligent analytics. A Gartner survey of data management leaders conducted in Q3 2024 found that 63% of organizations either do not have or are unsure whether they have the right data management practices for AI (Gartner, Lack of AI-Ready Data Puts AI Projects at Risk, February 2025). And in 2026, research by Harvard Business Review Analytic Services with Cloudera found that only 7% of enterprises say their data is completely ready for AI adoption (Harvard Business Review Analytic Services and Cloudera, Taming the Complexity of AI Data Readiness, 2026).

And something became very clear watching this happen.

I had watched this pattern play out in every company I worked at before Edilitics. The fragmented stack, the five or six disconnected tools stitched together with engineering effort, had never been the solution. It had always been a workaround. Now the same thing was happening at the AI layer. Tools were moving fast, shipping AI features, winning press cycles, and leaving the foundational problems completely untouched. The data was still ungoverned. The pipelines were still fragile. The answers were still unverifiable. The only thing that had changed was that the errors now sounded more confident.

AI sitting on top of unvalidated, ungoverned, fragmented data does not produce intelligence. It produces confident-sounding errors. The domain expert asking the question has no way to verify whether the answer is correct. The CFO cannot defend the number in a board meeting. The analyst cannot trace the output back to the query that produced it. The problem I had set out to solve in 2021, the structural dependency that locked experienced people out of their own data, had not disappeared. It had been dressed up in a chat interface and shipped faster.

The insight that had been forming slowly finally landed with full clarity: the three years of slow building between 2021 and 2024 were not a delay. They were the prerequisite that everyone else had skipped and was now desperately trying to retrofit.

Clean data. Validated data. Governed data. A semantic layer that gives an LLM the context it needs to query correctly. Encryption architecture that ensures no raw data ever reaches a model. An audit trail that makes every answer verifiable. This is what has to exist before AI can be trusted to answer a business question.

Edilitics had built all of it. Not because we were visionary. Because we had no existing stack to bolt onto and no shortcut available.

The urgency was visible now in a way it had not been before. The product was not just relevant. It was necessary.

May 2025: Building AskEdi

Three to four months of discussion before a single line of code.

That discipline was intentional. We had watched the market move fast and skip the thinking. We were not going to do the same thing for our own AI layer. The internal question was never how do we add AI. It was always: can this genuinely reduce effort for the people using it? Can a non-technical business user ask a question, get a verified answer, see the query that produced it, and act on it without a data team standing between them and their own numbers?

If the answer was yes, we would build it. If the answer was not yet, we would wait until it was.

The answer was yes.

In May 2025 the first working version of AskEdi was built in a separate branch. AskEdi is the intelligence layer of Edilitics: a non-technical business user asks a question in plain language, AskEdi queries the database directly, returns a verified answer with the exact query that produced it, and recommends what to do next. No SQL. No data team. No raw data leaving your database. Front end and back end. The entire flow working end to end. Then it was handed over for production readiness, and what happened during that process was as important as the build itself.

Working extensively with LLMs at the engineering level exposed areas where the codebase could improve that would not have been visible otherwise. Instead of treating production readiness as a finishing exercise, it became a maturation exercise. Architecture decisions reviewed. Code quality improved. Execution flows refined. Fixes applied while monitoring real impact. The process of making AskEdi enterprise-ready made the entire platform stronger.

The result was not a chatbot on top of unvalidated data. It was a governed AI layer on top of a clean, encrypted, semantically structured foundation. Every answer shows the query that produced it. Every query runs against validated, governed data. Every result carries a confidence level. The domain expert asking the question can verify the answer before acting on it.

That is what three years of foundation building made possible. Not the AI. Everything underneath it.

November 5, 2025: the launch post went live on LinkedIn. Four modules. Integrate, AskEdi, Transform, Visualize.

November 2025 to May 2026: The Decision Nobody Understood

After the launch post I did not reach out to a single potential customer.

From the outside this looked like hesitation. Like a founder who had spent five years building something and then got cold feet at the moment it mattered. I understood how it looked. It was not what was happening.

The decision not to go to market was not about perfectionism. It was about a product philosophy that had been at the centre of every decision since February 2021. A partial stack does not solve the problem. It adds to the fragmentation.

Integrate alone is just another connector tool. Integrate and Visualize without Transform is pretty dashboards on dirty data. Any of that without AskEdi is just another BI layer with no decision intelligence. And AskEdi without the governance layer, without DQ scoring, without AIR scoring, without Decision Intelligence built in, is just another chatbot on top of unvalidated data.

Launching too early would not have been a soft launch. It would have been a betrayal of the entire reason Edilitics exists. The whole point was to collapse the fragmented stack, to replace five to seven disconnected tools with one governed platform. Ship two modules and you become the thing you set out to replace: just another tool people have to stitch together with everything else.

The threshold was never perfect. It was complete enough to actually solve the problem end to end.

So through late 2025 and into 2026, Visualize got another pass. Transform got another pass. The governance layer, DQ scoring, AIR scoring, Decision Intelligence, the layer that makes AskEdi fundamentally different from every other ask-your-data product, got built properly and integrated.

Not a chatbot on top of unvalidated data. A governed AI layer on top of a clean, encrypted, semantically structured foundation. That distinction is everything.

May 2026: Five Years

The product that started as a frustration at a consumer tech company in Mumbai in 2019. That crystallised on a highway to Goa in January 2021. That nearly died in the ChatGPT panic of late 2023. That was rebuilt twice from scratch. That came close to something significant in a government conference room in September 2024 and had to walk away. That launched quietly in November 2025 and then went back to building for six more months.

Is ready.

All modules. All governance. Production-ready.

Five years is a long time to build something. Long enough that the person who started it and the person finishing it are not quite the same. The Growth Manager who sat down with W3Schools in late 2019 because he had no other option did not know he was starting a five-year build. He thought he was learning a skill. What he was actually doing was beginning to understand a problem deeply enough to eventually solve it.

I think about that version of myself sometimes. The frustration of having genuine expertise and being unable to act on it without going through someone else. The indignity of it, if I am being honest. Knowing what a number meant, knowing what to do about it, and still having to wait. That feeling is what every decision in this company has been made against.

I did not set out to build the most technically complex version of this problem. I set out to build the version that actually worked for the person I had been in 2019. The Growth Manager who understood the business deeply, could look at a number and immediately tell you whether it was good or bad and why, but was still dependent on someone else just to pull the number in the first place.

That person can now use Edilitics. Ask a question in plain language. Get a verified answer. See the query that produced it. Check the confidence level. Get a recommended action with scenario impact ranges. Act on it without a data team standing between them and their own numbers.

That is what five years built. Not a product. A removal of a dependency that the industry had decided was acceptable and that I decided, on a highway to Goa in January 2021, was not.

Sources

Gartner. Lack of AI-Ready Data Puts AI Projects at Risk. February 2025. Retrieved May 2026. gartner.com

Harvard Business Review Analytic Services and Cloudera. Taming the Complexity of AI Data Readiness. 2026. Retrieved May 2026. cloudera.com