We Built 6 AI Products in 6 Months. Here's What We Learned.

Some people spend six months planning their perfect product. They build intricate roadmaps, debate architecture, and wait for the "right" moment to launch. I took the opposite approach. I decided to build six AI products in six months, shipping them to production as fast as I could.

This is the origin story of Uptrail. It's a look back at the failures, the pivots, and the unexpected wins that came from simply moving faster than my doubts.

The Theory of Rapid Iteration

When the capabilities of foundational models (like Claude 3 and Gemini 1.5) started to explode, it became obvious that the bottleneck was no longer what AI could do, but how fast we could integrate it into practical workflows.

My thesis was simple: the only way to figure out what actually creates value is to put it in people's hands.

Here's what I shipped, what worked, and where I fell flat.

1. PropAutopilot: The Real Estate Engine

My first attempt was PropAutopilot. The idea was to automate property management workflows using AI. It seemed like a massive market with terrible legacy software. I figured if I could just automate tenant screening and maintenance requests, I'd have a winning SaaS.

• The Result: It worked technically, but the go-to-market was a nightmare. Property managers are slow to adopt new tech, and the sales cycles were brutal.
• The Lesson: A great technical solution cannot fix a fundamentally resistant market.

2. BoxCast: turning any webpage into a video

After the sluggish pace of enterprise real estate, I wanted something visual and immediate. BoxCast was born from a frustration with explaining UI bugs. What if I could just draw a box on a screenshot and have an AI describe it, then turn that into a mini presentation?

Using the Vision API, BoxCast took screenshot regions and generated robust descriptions.

// The core insight: bounding boxes mapped to localized vision prompts
const response = await ai.analyzeVision({
  image: screenshot,
  regions: [
    { x: 10, y: 50, w: 200, h: 200, prompt: "Describe UI element in this box" }
  ]
});

• The Result: Developers loved it. It gained traction fast, especially with early-stage teams doing QA. The freemium tier with a watermark created a natural viral loop.
• The Lesson: Visual, shareable utility tools grow themselves.

3. ChaiScout: An AI Concierge for India

ChaiScout was an exploration into WhatsApp-native AI. In India, WhatsApp isn't just a messaging app; it's the operating system for daily life. ChaiScout was designed to be a concierge that could handle local queries, book reservations, and answer questions.

• The Result: The latency was the biggest hurdle. WhatsApp users expect instant replies, and routing requests through heavy LLMs often took 5-10 seconds.
• The Lesson: Platform context is everything. What works on a web interface feels sluggish on a chat platform.

4. ModelMesh: Optimising LLM Routing

As I built more products, the api costs started to sting. I realised I was using heavy models (like Opus) for simple tasks, and fast models (like Haiku) for complex reasoning where they failed.

ModelMesh was built out of sheer necessity. It's a 14-dimension classifier that intercepts prompts and routes them to the cheapest model capable of handling them.

• The Result: This was the biggest surprise. It cut my API costs by nearly 75%.
• The Lesson: The most valuable products are often the internal tools you build to solve your own pain points.

5. PayIndia: The Tourist Payment Gap

This was less of a product and more of a deep-dive research initiative. I noticed that while India’s UPI is incredible for locals, tourists are completely locked out of the digital economy.

I built a prototype that bridged international cards with the UPI network.

• The Result: Regulatory hurdles made it impossible to launch as a solo founder in 30 days.
• The Lesson: Some problems are structural, not technical. You can't out-code compliance.

6. Marketing Agents: Automating the Hustle

By month five, I had these products running but no time to market them. So, I built a system of 27 interdependent AI agents. A Scout agent would analyze Reddit conversations, a Content agent would draft tweets, and a Distribution agent would schedule them.

• The Result: It worked surprisingly well. It felt like having a tireless, slightly erratic marketing intern.
• The Lesson: Agentic workflows aren't just a buzzword; they are the future of solo-entrepreneurship.

The Takeaway

Building six products in six months wasn't about finding six unicorns. It was about building at-bats. It was about training a muscle for rapid execution and learning how AI actually works in production, not theory.

The biggest lie in tech is that you need a perfect plan to start. You don't. You just need to ship.