Vinod Kannan

Building AI products & gamification | Co-Founder @BooleanBeyond & @BricqsAI | AI engineering, product building & startup | Coimbatore

[Vinod Kannan @V1n0d]

IPL starts tomorrow. But something important is missing this year.

Fantasy cricket.

For the last few years, fantasy wasn't just a side feature.
It changed how people watched the game.

Fans didn't just support a team. They tracked individual players. Every ball had stakes.

Having spent two years building and managing the IPL fantasy league for BCCI, we saw this firsthand. Over 1 million users participated with minimal promotion, 300k+ daily average squads created just to play for free.

People didn't come for rewards. They came to play along.

Globally, this behavior is only growing. During the FIFA World Cup, most fans use a second screen while watching. Leagues like NBA and NFL have leaned into this with live predictions, real-time interactions, and data-driven experiences.

In India, that second screen still exists during IPL. But now there is no structured way to engage. Just scrolling, memes, and reactions.

The attention hasn't gone anywhere.
The format has.

#IPL2026 #FanEngagement

[Vinod Kannan @V1n0d]

@remembradev Memory is becoming the real differentiator for agents. The interesting question is when to use long-term memory vs just better context management. Sometimes a well-designed context window beats a full memory layer.

[Vinod Kannan @V1n0d]

@KaSentenza48360 This is so true. Context management is the most underrated skill in agent development. We started treating context like a budget, not a dumping ground. Decide what the next step actually needs, pass only that. Made a huge difference in output quality at step 15+.

[Vinod Kannan @V1n0d]

@_vmlops This is the missing piece. MCP solved agent to tool communication. A2A solves agent to agent. The interesting part is how fast this is moving. Six months ago none of these protocols existed in any usable form.

[Vinod Kannan @V1n0d]

@KrishXCodes This is so true. We spent weeks fine tuning prompts early on. The real improvement came when we fixed how data flows into the agent. Right context at the right time beats a perfect prompt every time.

[Vinod Kannan @V1n0d]

@Indianinfoguide Tamil Nadu has always had deep manufacturing roots, especially in textiles. Great to see that paying off at the national level. When one state builds real competitive strength like this, it lifts India's export story overall.

[Vinod Kannan @V1n0d]

@MikZavolko The gap between funding headlines and what most AI startups actually need is wild. You don't need $110B to build something useful. A small team, clear problem, and the right architecture can get you surprisingly far.

[Vinod Kannan @V1n0d]

If you are building an AI product, here is a simple architecture that works for most use cases:

1. User input comes in
2. A classifier decides what type of request it is (costs almost nothing, use a small model)
3. Route to the right handler. Not everything needs an LLM. Some requests just need a database query or an API call
4. Only send to the expensive model when reasoning is actually needed
5. Cache responses for repeated patterns

This alone can cut your LLM costs by 60-70% and make your app way faster.

Most AI apps are slow and expensive because they send everything to one big model.

[Vinod Kannan @V1n0d]

Good additions. The input hash for reproducibility is a smart approach. When a client reports a bug and you cant reproduce it, debugging becomes really hard. And the 3-tier routing makes sense, most people start with one model for everything and then get surprised by the bill when usage goes up.

[Vinod Kannan @V1n0d]

If you are building AI agents for clients, here are few things that will save you a lot of pain

1. Always test with real data, not clean sample data. Real data is messy and your agent needs to handle that from day one.

2. Add guardrails at every step. Input validation, output checking, and a fallback option if the agent is not confident. This builds trust with the client.

3. Track your costs early. LLM calls add up fast when you scale. Use smaller models for simple tasks and cache repeated calls.

4. Log everything. When a client asks why the agent gave a certain answer, you should be able to trace the full flow in minutes.

5. Start with one workflow, get it solid, then expand. Trying to automate 10 things at once is how projects fail.

Production AI agents are not that complex. But you need discipline and good systems.

[Vinod Kannan @V1n0d]

Tim Berners-Lee had this vision 25 years ago. He called it the Semantic Web. The idea was simple. Machines should understand data, not just display it.

He wanted ontologies, linked data, and structured formats so that systems could talk to each other without humans in the middle.

It didn't really take off on the public web the way he imagined. Too complex, too academic, and hard to implement at scale. Most of the internet stayed as messy HTML pages built for humans.

But ontologies didn't die. They quietly evolved into the foundation of knowledge graphs. Google built one of the largest knowledge graphs in the world. Every time you search and see that info panel, that's this idea in action. Healthcare, finance, and logistics all use knowledge graphs today to manage complex relationships in data.

Now the interesting part.

Anthropic introduced MCP (Model Context Protocol), which is an attempt to standardize how AI systems connect to tools like databases, APIs, and internal services.

At the same time, there is growing work around agent-to-agent communication, referred as A2A, where multiple AI agents coordinate tasks across systems.

Now this is where things start to come together. Knowledge graphs provide structured context. Protocols like MCP provide access to tools. And multi-agent systems enable coordination.

So in a way, Tim Berners-Lee's vision is starting to happen. Just not in the way it was originally imagined. Instead of forcing all data into rigid structures, we now combine structured systems like knowledge graphs with models that can handle messy data. Instead of one universal standard, we are seeing protocols and patterns emerge that help systems work together.

Sometimes the right idea just needs the technology to catch up.

[Vinod Kannan @V1n0d]

@jumbld This is the kind of infra that was missing. Building AI agents is one thing but running them at scale for enterprise clients needs proper observability, access control and cost tracking. Without this you are flying blind in production.

[Vinod Kannan @V1n0d]

@Krishnaik06 This is becoming more common now. For any serious business use case, data privacy is a real concern. Clients dont want their data going to third party servers. Self hosted agents with open source models is the way forward for enterprise.

[Vinod Kannan @V1n0d]

@eng_khairallah1 This is the real problem. Everyone gets the demo working and thinks the hard part is done. The hard part actually starts after that. Real data, edge cases, error handling, cost management. Building for a client is completely different from building for a tutorial.

[Vinod Kannan @V1n0d]

@TA_Purvesh @Sanju_Lakshya This is visible on the ground. States like TN are actually executing on this. TIDCO investing in EV and deep tech startups, GCCs growing fast in Chennai, electronics manufacturing expanding. The shift from just services to real economy is real.

[Vinod Kannan @V1n0d]

@epic_bengaluru @Maheshbr4U @DevanahalliBLR @Kuma50531Kumar @MBPatil @investkarnataka @CommIndustries @ITBTGoK @PriyankKharge BLR deserves the crown but Chennai is quietly catching up on the GCC front. Lot of new centers opening up, especially in OMR and Ambattur corridor. Not competing with Bangalore but the pie is big enough for both cities to grow.

[Vinod Kannan @V1n0d]

@HackingDave This is exactly right. Building AI agents for real use cases is mostly debugging, error handling and integration work. Not the shiny demo stuff you see on timeline. The gap between a cool demo and something that actually runs daily without breaking is huge.

[Vinod Kannan @V1n0d]

I use a simple trick. Keep 2-3 base passwords that you remember well, then add the site name or some pattern of keyword from the site to it. So your Netflix password and Amazon password are different but you can always figure it out because the base is same. Apple passkey is also a good option if you are in Apple ecosystem.

[Vinod Kannan @V1n0d]

@startupindia This is happening on the ground. Chennai and Coimbatore are seeing real growth in EV and electronics manufacturing. TIDCO backing deep tech startups is the right move. TN already has the talent and industrial base, just needed this kind of push.

[Vinod Kannan @V1n0d]

6 months ago AI agents were just demos and Twitter threads. Now Perplexity runs on your Mac in the background, Claude works inside your Excel and browses web for you. The speed at which "AI doing actual work on your computer" became normal is crazy. And we are still just getting started.

You can start with giving your simplest tasks like morning research, cleaning up spreadsheet data, drafting emails from rough notes to Claude Cowork or Perplexity Computer.

The jump from "asking AI a question" to "AI running a part of your day" is where things get interesting. It takes some effort to set up but once it clicks, you won't go back.