For the next couple of blogs, I'm going to talk about thoughts on entrepreneurship, especially entrepreneurship with AI, machine learning, and data science. In this one, I'll share my initial aspirations about entrepreneurship and machine learning, and in the next blog, I'll talk about what I learned when I was at Captain Fresh.
The Mu Sigma Days: Learning the Business Side
At Mu Sigma, I was learning all the nuances of how a company operates—different problem statements, client stories, and a lot of hands-on work. Around 2022, I happened to work with a lot of clients, especially on the business development side, which meant working on proposals and getting deals or engagements done.
I remember a couple of deals that were more than a million dollars—multi-year deals that I was able to work on. And those conversions showed me something: I'm not just good at building solutions, I'm also able to formulate solutions and sell them—sell the perspective, sell the capability of the company. That's what the role intended, because it's not just about doing hands-on work or building models and shipping them. It's also about sitting from the client's perspective, seeing what problems need to be solved, working on proposals, and ultimately winning business.
One thing I noticed after all these back-to-back engagement wins—I remember one engagement which was on network optimization of different product lines and their inventories. I spent 24 hours straight in Mu Sigma, didn't sleep, worked on the proposal. It might sound like a McKinsey kind of thing, but working on a proposal isn't just about coming up with case studies. It's about what methodology we're going to build, how we'll utilize the data, what hypotheses we'll consider, what assumptions we have, the plan, time and material—multiple things.
And if you think about it, that's exactly what startup venture capital pitches look like. In those 24 hours, I got the proposal ready, submitted it, and it won. We set up the team, and the process of building started.
There were also months of preparation for multi-million dollar proposals—presentations, prototyping to demonstrate capability.
The Dream Started Taking Shape
With all of that, I used to tell my friends that I have this dream, this vision of starting a company, running a company. At that time, I had heard stories of different companies started by ex-Mu Sigma folks, like MathCo, and other small ventures. Some people who started their career at Mu Sigma ended up as vice presidents of analytics or heads of machine learning and data science at major companies.
That gave me hope. We were on a similar path—dealing with large-scale machine learning and data science transformations, taking up leadership roles. I was leading teams, setting them up, getting the work done, presenting to different levels of leadership. We could build things, build solutions, and do the pitching and selling.
That's when I started sharing the idea of Sahara Analytics with people. It was a rough idea: why don't we start yet another analytics company? Not a ditto copy of Mu Sigma, but something with our own original thought process built into it.
The initial idea was about building a services company—data science, machine learning services. But with a different thought process in approaching solutions. One thing I noticed is that people don't explore much beyond typical solutions. In many services companies, they just share off-the-shelf approaches or whatever that particular individual knows. It's not unique to one company—it's across many services companies and even product companies. It's their own bias and limited knowledge.
So the vision was: what if we could make a setup that helps companies build truly robust machine learning solutions? More uncertainty-aware forecasting, more causal-centered solutions—not just the typical stuff.
The Demand Sensing Problem: This Is What Excited Me
There was one problem statement we were working on: demand sensing—understanding how short-term changes and real-time impacts affect demand. This is fundamentally a causal inference problem. Consider demand as a system with multiple variables impacting it. Some are long-term—raw material procurement that happened three months back will impact production and supply down the line.
Some are short-term—sudden seasonality drifts, a new competitor entering the market, controversies around the brand. These can quickly change perception. If you want to accommodate all of this, traditional approaches just use historical data. But with demand sensing, you may not have meticulously collected data for these scenarios. It's an open-ended problem, and we were exploring many directions.
One direction was bringing in causal-based signals—actually accounting for what's causing the ups and downs, not just predicting patterns. That's when I realized there was a lot of scope for building a pure applied machine learning company, one that doesn't get deviated by data engineering migrations or reporting work. I'm deeply inclined towards core machine learning and data science work, not support and infra deals.
Why Sahara? The Middle East and EMEA Focus
We called it Sahara Analytics because we were targeting the Middle Eastern market—EMEA: Europe, Middle East, and Africa. When I say "we," I mean the like-minded people I connected with to discuss this idea.
That region had high-potential customers with very low adoption of machine learning and data science. The focus was on companies facing operational deficiencies—manufacturing, production, raw materials, petrochemical industries, processing industries. In Europe, it would be telecommunications. Very rarely did I have banking or other sectors in mind.
Everything revolved around sales, supply, demand, production planning, inventory optimization—machine learning plus operations research. In the Gulf countries especially, there was high demand because many companies were still dealing with Excel sheets or very old platforms. That seemed like a real opportunity.
Why I Didn't Take the Next Step
This was the idea, but I didn't have further steps. I wasn't sure how to proceed. Typically, you have to leave your job and focus completely. At that time, I wasn't mature enough to leave and take that risk. Now I've left my job to do my masters, but back then it was different.
So the idea just kept running in my mind and in discussions. Meanwhile, I observed how others were building machine learning and data science focused companies.
Back then, product companies weren't completely built on machine learning and data science. OpenAI existed, AI-based companies existed, EV companies existed—but thinking about machine learning for operations and operations research, it was mostly about building end-to-end platforms. Companies like o9 Solutions, Kinaxis, Celonis, Palantir—they were all building operational systems for bigger companies.
It's not just about building a great machine learning algorithm or model. For these problems, there's no free lunch—no single model that can be sold to every client. Every client struggles first with fragmented data sources. So these platform companies provide infrastructure to combine data, modules to tune and run models, and an operational layer connecting to planning tools for decision-making.
That's how platforms in data mining, process mining, supply chain planning, demand forecasting, customer analytics, and marketing analytics (Adobe Marketing Cloud, Nielsen, Kantar) built their moat—on data and the decision-making layer. The machine learning in between is still in the hands of engineers; it can't be completely off-the-shelf. That's why these companies rely heavily on support engineers and forward-deployed engineers who do customization.
The Two Types of ML Companies I Saw
Then there were companies using machine learning as part of consumer products—Grammarly using NLP for translation, summarization, paraphrasing.
So I saw two aspects. First, B2B enterprise machine learning—not purely ML, but data science and machine learning platform companies. Every problem requires different algorithms, but platform-wise and decision-making layer-wise, a common stack can work.
Second, consumer products like Grammarly, or cybersecurity companies with machine learning for anomaly detection.
This was my understanding back then. When I was interviewing, companies were mostly in these categories—not like today where many startups are purely agentic or build models for multiple use cases, with downstream ML applications where the entire solution is machine learning focused. Back then, my understanding and confidence was around building a services company.
The Reality of What It Would Take
But building a services company demands a lot. I wasn't mature or courageous enough to quit, grind for three or four months on proposals, win deals, build and deliver, and keep hiring.
If I reflect honestly, it would mean competing against Mu Sigma, MathCo, Tredence. That's the real market. I observed many connections and friends who started their own ML/DS solutions companies, but they all operated at very small scale and didn't grow over time. They dealt with local clients or very small teams.
That wasn't my choice. I wanted to deal with Walmart, Mercedes-Benz, Mitsubishi—the customer base that frontier data science and machine learning service companies target.
That demands a lot. Maybe an MBA for credentials. Maybe six continuous months of struggle. When I joined Mu Sigma around 2018, MathCo was just building and growing. I remember a LinkedIn post showing MathCo reached 300 employees. Now it's 10x or 15x that. MathCo is the most recently built and successfully running data science and machine learning services company. Before that, Tiger Analytics, Latentview, Fractal—these are older. Impact Analytics, C5i came after but weren't as big.
After MathCo, I didn't see many companies venture into this space. Some confined themselves to small-scale businesses or scattered into training institutions and marketing. Some limited themselves to 30-40 people, operating more like agencies than building large client services teams of 1,000 or 2,000.
It's a people game. A billion-dollar services company might have 3,000 employees. Meanwhile, AI product companies can have billion-dollar valuations with 10-20 people. That's how the market and VCs value things. Services companies are people-dependent, talent-dependent, and the talent is scattered because you can't be picky about projects. You might need to take data engineering work, dashboard refreshes, and sometimes maintain engagements with one or two people on low billing, hoping for something bigger in the future.
This was my observation. At that time, I had never worked with any product companies.
The Chennai vs Bangalore Startup Ecosystems
I also observed product companies like Zoho, Freshworks, Chargebee. People used to say Chennai has a high SaaS ecosystem—lots of SaaS companies, not asset-heavy or people-heavy.
In Bangalore, you had Swiggy, Zomato, Uber, Ola—companies getting started and rolling out aggressively around early 2019. Lots of new unicorns. But Chennai rarely produced unicorns because the mindset is bootstrapping. Companies bootstrapped and were comfortable building SaaS products—they learned that from Zoho and Freshworks.
They focused on highly niche problems. Zoho on office suites and mail. Freshworks on support desk automation. Each company picked small niches with subscription-based customers worldwide.
From my perspective, building such products—I never had exposure, and I wasn't interested in building support systems, ticketing software, subscription management, or wallets. That's not where my mind goes. People say you develop taste and curiosity. My curiosity was budding around data-driven problem solving and machine learning, not building utility software products.
Having worked in services companies, my understanding was that building a services company is a feasible way to keep momentum in data science and machine learning. But I couldn't step into the next phase because it's a big game. As an individual, it felt very difficult. I had less risk appetite, and starting would mean going against financial comfort.
What I Learned From All the Reading
That's how I started developing the idea and reading books about how people built companies. But one thing I noticed: most books are about building products—build what customers want, whether it's e-commerce, testing software, documentation tools, or observability platforms.
Two things I learned. First, you can't build a "free lunch" company that solves all problems—that's common sense. Second, I'm not interested in building utility software. Either I should build a platform that helps companies do end-to-end problem solving in a specific area, or build a services company that can do that.
That's how the curiosity and exploration started back in 2022 with one rough idea: Sahara Analytics. And I kept exploring.
And Then Came Captain Fresh
Then I joined Captain Fresh. It was a good platform for me to explore the product side—how products actually get built, what happens in that process. Me and my manager Parag were working like entrepreneurs within the startup. The company supported that way of operating. VP Venu and founder-CEO Uttam Gowda were aligned with that approach—they supported and motivated different product lines to operate that way.
In the next blog, I'll share what I learned at Captain Fresh while working with Parag. That experience was valuable because I was moving towards what I had been wandering about during my Mu Sigma phase—building a product where machine learning and data science is the core component, solving real problems with scope for growth.
And it wasn't boring. The problem statement was so interesting that I joined Captain Fresh over offers from Jio, PayPal, Kmart, Paytm, and other startups.
Let's see what I learned in Part 2.
To be continued.