Forem: Dharma Teja

How Kiro Helped Me Build the Most Complex Product I Ever Attempted - In Less Than a Week

Dharma Teja — Fri, 05 Dec 2025 17:07:39 +0000

Kiro helped us build the most complex thing that we could ever build.
And honestly, I still can’t believe the speed at which everything happened.

We built something called Stitch API.
It solves a real problem we face every single day as both founders and developers.
I’m not exaggerating — this is literally the tool I always wished existed.

Try it here:
https://stitchapi-g87p.onrender.com

Why Stitch API Exists

Founders keep switching between too many dashboards.
Calendars here.
Analytics there.
News somewhere else.
Stock prices on another tab.
Meeting tools.
Notes tools.
Tools for tools.

It becomes a mess very quickly.

Developers have a completely different but equally painful problem.
Every product we build depends on so many APIs.
LLM model APIs.
YouTube Data API.
Payment APIs.
Notification APIs.
Everything scattered, everything separate.

Both sides suffer.
Both sides repeat the same work.
Both sides waste time.

So we decided to unify EVERYTHING into one simple, clean interface.

One place for founders.
One place for developers.
One place for every API, every service, every integration.

That became Stitch API.

The Original Plan Was Impossible

I’ll be honest.
If we had to build this the traditional way, it would easily take more than a month of senior developer effort.
And that’s not even counting testing, fixing all the integration issues, and wiring everything together.

This is the type of product that normally requires:
• backend systems
• API routing
• rate limiting
• unified dashboards
• multiple service connections
• authentication layers
• UI polishing
• and a LOT of glue code

It wasn’t supposed to be a “one week” project.

But then we used Kiro.

What Changed When We Used Kiro

This is the part that shocked us.
We started by generating specs in Kiro.
Kiro immediately produced three complete task files — and those task files basically structured the entire product.

Those three files helped us build almost 90% of the full application.

From backend logic.
To API structure.
To functionality breakdown.
To clean workflow.
Everything was just… there.

And the experience didn’t stop there.
We used wipe-code to complete the remaining parts.
The flow felt natural.
The process felt smooth.
The clarity was unbelievable.

This is honestly the fastest I’ve ever shipped something this complex.

The Most Mind-Blowing Part: API Integrations

One thing that usually eats most of the time is API integration.
Connecting everything.
Handling errors.
Fixing mismatched responses.
Dealing with edge cases.

But this time…
Kiro handled the ENTIRE API integration process.

No manual interruption.
No spending hours reading documentation.
No trial and error.
Kiro just did it.

And it felt like magic.

Stitch API Is Now Live

Today, Stitch API brings one place where both founders and developers can manage every service and every API they use.
And I genuinely mean it — Kiro played a major role in making this possible so quickly.

Live link:
https://stitchapi-g87p.onrender.com

GitHub:
https://github.com/surya4419/StitchAPI/

How Kiro Changed My Approach to Building

Before Kiro, I used to think in terms of weeks and months.
After using Kiro, I started thinking in terms of days.

Instead of planning how much time something will take, I started thinking:
“How fast can Kiro help me turn this into an executable product?”

This shift is huge.
It changes how fast we experiment.
It changes how fast we validate ideas.
It changes how fast we ship.
It changes how fast we grow.

This tool genuinely changed the way I approach development.

My Honest Conclusion

I’m not writing this because of a competition.
I’m writing this because the experience I had with Kiro was completely unexpected.
It removed friction.
It removed complexity.
It removed the bottlenecks that normally slow us down.

And it allowed us to build something we’re genuinely proud of.

Stitch API exists today because of that speed.

And that speed exists because of Kiro.

I Researched about the “Research Tool” that's rewriting the Rules: Perplexity AI

Dharma Teja — Thu, 07 Aug 2025 18:12:25 +0000

When I first started poking around Perplexity AI’s documentation and blog, what struck me was not just the polished UI or the witty brand voice - but the sheer precision of its engineering. As a developer, you already know the theoretical distinctions between search engines, chatbots, and answer engines. What few get to see is the code-level choreography that makes real-time, sourced answers possible.

In this blog, I’ll walk you through:

Core Architecture & Technologies
Citation & Reasoning Pipeline
Model Orchestration & Agents
Scalable Infrastructure & Data Flow
Complete Founder Story & Growth Metrics

Along the way, I’ll sprinkle in code snippets illustrating how Perplexity’s dev team likely tackled key challenges. Buckle up-it’s going to get technical.

1. Core Architecture & Technologies

At its heart, Perplexity combines three pillars:

Live Web Scraping & Retrieval
Large Language Model Synthesis
Citation Extraction & Formatting

1.1 Live Retrieval Module

Perplexity employs a two-stage retrieval:

First, a fast keyword filter (using Elasticsearch or Vespa) narrows documents;

Then a lightweight Transformer-based reranker (like a distilled BERT) selects top passages.

# Pseudocode for retrieval + rerank
docs = elasticsearch.search(query, size=100)
reranked = reranker.predict([q + d.text for d in docs])
top_docs = sorted(
    zip(docs, reranked),
    key=lambda x: x[1],
    reverse=True
)[:5]

1.2 LLM Synthesis Layer

Once top passages are identified, Perplexity feeds them into an LLM prompt engineered for concise answers with citations. Their prompts likely follow a pattern:

You are an expert assistant. Given these passages:
1. [URL1]: “...”  
2. [URL2]: “...”
Provide a summary, citing each fact like “[1]” or “[2]”.

Behind the scenes, they leverage OpenAI’s GPT-4 Omni for general questions, then fall back to Sonar (their optimized in-house model) and hell other languages for cost efficiency.

response = openai.ChatCompletion.create(
  model=selected_model,
  messages=[
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": user_query}
  ],
  temperature=0.2,
)

(see the generated image above)

1.3 Citation Extraction & Rendering

To ensure every claim is sourced, Perplexity parses the LLM’s output using a regular-expression post-processor.

import re

citations = re.findall(r'\[(\d+)\]', response.choices[0].message.content)
for idx in set(citations):
    source = top_docs[int(idx)-1].url
    print(f"[{idx}] {source}")

This lightweight approach seamlessly transforms inline markers into clickable footnotes.

2. Citation & Reasoning Pipeline

Beneath the user-facing simplicity lies a multi-agent pipeline:

Query Agent – normalizes input, extracts keywords
Retriever Agent – fetches candidate documents
Reranker Agent – scores relevance with neural network
Synthesis Agent – compiles answer with citations
Verifier Agent – sanity-checks facts via secondary searches

Each agent runs in Docker containers orchestrated by Kubernetes, enabling horizontal scaling as query volume spikes.

apiVersion: apps/v1
kind: Deployment
metadata: { name: synthesis-agent }
spec:
  replicas: 5
  template:
    spec:
      containers:
      - name: synthesis
        image: perplexity/synthesis:latest
        resources:
          requests: { cpu: "2", memory: "4Gi" }

3. Model Orchestration & Agents

Perplexity’s secret weapon is its model-agnostic orchestration:

Model Selector: routes queries based on complexity
Parallel Inference: runs multiple models concurrently
Cost Optimizer: shifts low-priority queries to cheaper models

# Simplified model selection logic
if query.complexity > threshold:
    model = "gpt-4-omni"
else:
    model = "sonar-small"

This micro-optimization yields massive cost savings without sacrificing answer quality.

4. Scalable Infrastructure & Data Flow

Handling 780 million queries monthly demands an iron-clad backend:

Edge Caching: Varnish caches frequent answers for <1 second responses
Stream Processing: Kafka pipelines ingest clickstreams for real-time analytics
Monitoring & A/B Testing: Grafana dashboards track latency, and Ray Tune runs prompt-template experiments

# Kafka consumer example
kafka-console-consumer --bootstrap-server kafka:9092 --topic query-events

Okay, so what happened next literally defies every startup playbook I've ever read.

Check out these numbers and try not to fall off your chair:

December 2022: Public launch
February 2023: 2 million users
March 2023: $26M Series A, 10 million users
January 2024: $520M valuation
April 2024: $1B valuation (unicorn status)
December 2024: $9B valuation
July 2025: $18B valuation, $100M ARR

That's a 120x valuation increase in 2.5 years. I've been tracking startups for years, and I've never seen anything scale this fast while maintaining quality and user satisfaction.

Their investor list reads like Silicon Valley royalty:

Jeff Bezos (Amazon founder)
Jensen Huang/NVIDIA
SoftBank Vision Fund
Yann LeCun (AI pioneer)

When your competitors' leaders are investing in you, you know you're onto something massive.

Let's talk about the man behind this!

The 0.01% That Changed Everything

Picture this: You're at IIT Madras, arguably India's MIT. You want Computer Science more than anything. The cutoff comes out, and you miss it by 0.01 points. Most people would be devastated. Aravind Srinivas? He called it destiny.

That microscopic "failure" in 2017 became the catalyst for what's now an $18 billion company processing 780 million queries monthly. I mean, come on - you can't make this stuff up!

Born June 7, 1994, in Chennai, Srinivas grew up in the same city that produced Google CEO Sundar Pichai. But unlike most success narratives, his story starts with what seemed like academic disappointment. Getting stuck in Electrical Engineering instead of CS at IIT Madras felt like the end of the world to him.

But here's where it gets interesting - and this is where I really started appreciating his mindset. That EE background gave him the mathematical foundation essential for machine learning. Plus, a forward-thinking professor's Python programming class equipped him for what would become a Python-centric ML world.

"Nobody was really into Python that much in IIT at the time, and he was very prescient, and that helped me a lot, because obviously, ML is largely being done in Python."

- Aravind Srinivas

The guy literally turned a setback into a setup. And honestly? That's the kind of resilience that separates world-changers from the rest of us.

The Strategic Academic Journey: Building The Foundation

What impressed me most about Srinivas's trajectory is how intentional every move was. After IIT Madras (2017), he didn't just randomly apply to grad schools - he strategically positioned himself at UC Berkeley for his PhD in Computer Science.

But check out his internship game - this is where I realized this guy was playing chess while everyone else was playing checkers:

2018: Research Intern at OpenAI
2019: Research Intern at DeepMind
2020-2021: Research Intern at Google
2021-2022: Research Scientist at OpenAI

I mean, seriously? This isn't just impressive - it's a masterclass in strategic career building. Each role built upon the previous one, creating comprehensive understanding of the AI landscape from multiple perspectives. More importantly, it built the network and credibility that would prove crucial for Perplexity's success.

His PhD research portfolio reads like a greatest hits of modern AI:

Contrastive Learning for Computer Vision (CPCv2)
Reinforcement Learning (CURL)
Transformers for Image Generation (Flow++)
Decision Transformer for RL

The guy was basically positioned at the epicenter of every major AI breakthrough of the last decade.

The Future Vision: Where This Goes Next

Srinivas has set an ambitious target: "a billion queries a week" by end of 2025. That's 30%+ growth from current levels and would position Perplexity as a true Google competitor.

The Roadmap:

Comet Browser: Full web browsing with AI integration
Advanced AI Agents: Autonomous task completion
Voice-First Interfaces: Compete with Alexa/Siri
Hardware Integration: Native device experiences

The vision extends to becoming the universal knowledge interface - the single point where humans interact with all information.

My Personal Takeaways

After diving deep into this story, here's what really stuck with me:

Embrace Your Setbacks: That 0.01% miss wasn't a failure - it was redirection toward something bigger
Build Strategically: Every role, every connection, every skill Srinivas developed had compound effects
Solve Real Problems: Perplexity succeeded because it solved genuine user pain points
Stay Humble, Think Big: From a $100 domain name to an $18B company - but never losing sight of the core mission
Network Intentionally: The relationships built during those internships became Perplexity's foundation

Why This Story Matters

Honestly, researching Perplexity and Aravind Srinivas's journey has been one of the most inspiring deep dives I've done. It's proof that in our rapidly evolving tech landscape, the ability to learn, adapt, and persist matters more than perfect GPAs or predetermined paths.

This isn't just another unicorn story - it's a demonstration that with the right mindset, strategic thinking, and relentless execution, you can challenge trillion-dollar incumbents and win.

For every developer reading this who's ever felt like they missed their shot, or took a "wrong" turn, or ended up in a different place than planned - Srinivas's story is proof that sometimes the detour becomes the destination.

The research revolutionary from Chennai has shown us that sometimes, missing the mark by 0.01% is exactly what it takes to hit the target by 1000%.

As Srinivas says:

"It's only over when you think it's over. Until then, you can always find a way."

And honestly? After researching this incredible journey, I'm more convinced than ever that the best stories are still being written.

Resources & References

1.Times of India: Aravind Srinivas educational qualification and career path
2.Wikipedia: Aravind Srinivas
3.Frederick AI: Founder Story: Aravind Srinivas of Perplexity AI
4.Financial Express: Meet Perplexity CEO Aravind Srinivas: Indian-origin tech visionary
5.Hindustan Times: Who is Aravind Srinivas, Indian-origin CEO who challenged Elon Musk
6.MWC Barcelona: Aravind Srinivas Speaker Bio
7.Perplexity AI SEC Filings (D forms, 2024-2025)
8.Technical Comparison & Analysis: SE Ranking - ChatGPT vs Perplexity vs Google vs Bing
9.YouTube Video: Perplexity CEO Aravind Srinivas: From Academic to $9B AI Pioneer
10.Arxiv Paper: Exploring the Limits of Language Modeling

From Rejection to Recognition: How a "Failed" Hackathon Led to Our Biggest "Win"

Dharma Teja — Wed, 30 Jul 2025 23:49:01 +0000

The story of building an AI shopping agent, facing disappointment, and finding unexpected success

Who are we?

We are two college students, had this crazy idea. What if we could build an AI agent that would revolutionize online shopping? Not just another chatbot, but a truly intelligent agent that could understand user emotions and budget, recommend, and help users make better purchasing decisions.

So we entered the Bolt Hackathon with high hopes and endless energy.

The Technical Nightmare

Building our AI agent was like trying to solve a puzzle with missing pieces. Every day brought new challenges:

The Supabase Struggle

Connecting our AI agent to Supabase became our biggest headache. Just when we thought everything was working smoothly, something would break. The agent would stop responding. Data wouldn't save properly. Edge functions felt impossible to debug.

The API Limitation Crisis

We were using a third-party API to fetch ecommerce products, but it had very limited free queries. During testing, we'd hit the limit constantly. Our solution? We kept creating new accounts with different emails and phone numbers just to get more API calls.

It wasn't elegant, but desperate times called for desperate measures.

The Breakthrough Moment

After countless sleepless nights and debugging sessions, we finally did it. We had a complete, end-to-end working AI agent:

Frontend fully functional with Bolt
Backend running smoothly with Supabase
AI agent performing exactly as promised

We were proud. We were ready.

The Crushing Mistake

Then came the moment that changed everything.

We couldn't submit our project video in time. In a panic, we added a random unlisted YouTubevlink just to complete the submission, thinking we'd replace it the next morning with our real video.

We were too late.

Our submission got rejected because of that one missing video. While other projects with broken frontends, non-working backends, and incomplete demos made it to the top, our fully functional product didn't even get a chance.

It hurts. A lot.

The Unexpected Turn

But sometimes life has different plans.

After the hackathon ended, something amazing happened. Mentors were a lot of interest on our idea during the build sessions and were genuinely impressed.

What we discovered:

Out of 50+ project showcases, we had stood out in every build session.
Many mentors were founders themselves, and got impressed about our product.
Some expressed interest in future investment opportunities.
We received internship opportunities from their companies.

The Validation We Needed

During Bolt's build sessions, we had worked extensively with mentors to:

Refine our core idea
Conduct competitive analysis
Strengthen our product strategy
Perfect our pitch

This mentorship proved invaluable. Even without winning, we gained something more important: validation from industry experts.

Rise Your Hack: Our Second Chance

Just when we thought our hackathon journey was over, we heard about Rise Your Hack - the world's largest AI hackathon.

We shared our Bolt Hackathon story with the organizers, and they believed in us enough to let us participate. This time, we reimagined our AI shopping agent specifically for their Llama x Groq track requirements.

The challenge: We were still college students dealing with:

Maintaining attendance requirements
Middle of exam season
Multiple assignments due
Limited time and resources

But we pushed forward anyway.

The Victory That Changed Everything

After weeks of intense work, balancing college with development, we submitted our project.

The result was surprising:

🏆 We won the Llama x Grok track at Rise Your Hack - World's Largest AI Hackathon!

Our prizes:

Acceptance into the Grok startup program
$5,000 USD in Groq credits
Co-marketing support from the Grok team

The Bigger Picture

Looking back, our "failure" at the Bolt Hackathon wasn't really a failure at all. It was a setup for something bigger:

What we gained from Bolt:

Technical expertise in building AI agents
Mentor relationships and industry connections
Internship opportunities
Product validation from experienced founders

What we achieved at Rise Your Hack:

Recognition on a global stage
Substantial resources to grow our product
Access to startup acceleration programs
Proof that our idea has real potential

The Journey Continues

Today, we're not just college students with a cool project. We're entrepreneurs with:

A validated AI shopping agent
Industry mentor support
Internship opportunities
Startup program acceptance
Resources to scale our product

Sometimes the path to success isn't linear. Sometimes getting rejected opens doors you never knew existed.

Our AI shopping agent journey is just beginning, and we're more excited than ever about what comes next.

Building something meaningful while juggling college life isn't easy, but with the right idea, persistence, and a bit of luck, even a "failed" hackathon can become the foundation for something extraordinary.

I document my journey, visit here - https://www.instagram.com/teja.techh/