Bhavika

Pure Engineering | Data & AI

[Bhavika @stacktraceee]

A few first principles I’ve picked up while building with LLMs. https://t.co/kiEEXO8jfC

[Bhavika @stacktraceee]

Do YOU guys trust AI agents with your money?

I've been reading about agentic commerce for a while now, and honestly, I think it's one of the most fascinating things happening in AI.

Think about how we shop today.

We visit a few familiar websites, compare prices, read reviews, hunt for discounts, check delivery timelines, and somehow spend 40 minutes buying something worth ₹2,000.

And after all that effort, we've probably compared products across just 2–3 websites.

Now imagine an AI agent doing this across dozens of merchants. Finding offers we didn't know existed, comparing final prices after discounts and delivery, and actually making the purchase.

Not just recommending something. Buying it.

I genuinely believe agents could become better shoppers than us.

And what's exciting is how much infrastructure is already being built to make that possible.

Google and Shopify co-developed UCP to help agents interact with merchants and handle commerce workflows.

Google's AP2 addresses delegated payment authorization, intent constraints, and cryptographically verifiable checkout mandates.

Visa and Mastercard are building agent identity and payment trust infrastructure. Coinbase's x402 enables software agents to make programmable payments using stablecoins.

Even NPCI has been exploring AI/software-initiated payments through UPI Circle in limited pilots.

And then there are ideas that almost sound like science fiction.

Verifiable digital identities for AI agents. Agent-controlled wallets. Machine-to-machine payments. A future where agents might transact with other agents without humans participating in every step.

We're essentially building the infrastructure for software to participate in the economy.

And I find that incredibly exciting.

But while reading through all of this, I kept feeling there was one problem that deserved more attention.

A lot of what I was reading focused on making agents capable of transacting.

How do we identify an agent?

How do we know what it's authorized to spend?

How do we securely connect it to merchants and payment networks?

How do we verify that a transaction was authorized?

All extremely important questions.

But I kept coming back to something that happens BEFORE the payment.

Who independently verifies that what the agent is about to purchase is actually what the human asked for?

Imagine I tell an agent:

"Buy me a water-resistant backpack that fits my 16-inch laptop, contains no animal leather, costs under ₹5,000, and arrives by Friday."

The agent finds one for ₹4,500.

Good reviews. Right price. Delivery on time.

Everything looks perfect.

Except the particular variant it selected contains leather.

Now think about this.

The agent could be operating within its spending limit. Its identity could be verified. The transaction could have valid authorization.

And I could STILL receive something I explicitly asked it NOT to buy.

Because proving a transaction was authorized doesn't automatically prove that every product claim used to justify that purchase was true.

To be fair, AP2 already supports intent constraints and their verification. UCP helps standardize merchant interactions.

So the ecosystem isn't ignoring correctness.

But the part I became particularly interested in was the evidence behind those decisions.

What if a product claim belongs to a different variant?

What if the LLM silently drops an important requirement?

What if the cart changes after verification?

What if there simply isn't enough reliable evidence to establish whether a product meets the user's request?

And perhaps the most interesting question:

Should the same probabilistic system that proposes a purchase also be trusted to declare its own proposal correct?

It's almost like asking the agent to grade its own exam.

Except this time, the consequences involve your money.

And naturally, the engineer in me couldn't leave this alone.

So I built BuyRightAgents.

GitHub: https://t.co/OrFhRgts5n

It's an early Go-based prototype exploring an independent, evidence-backed verification boundary between an AI agent's proposed purchase and checkout.

The principle is simple.

An agent can propose what to buy. But it shouldn't be the final authority on whether that purchase is acceptable.

An LLM interprets the user's request into typed constraints, but its output is treated as untrusted data.

A separate deterministic verifier evaluates those constraints against evidence tied to the exact cart item, product, and variant.

Price and delivery are derived from the cart snapshot. Canonical hashing binds verification to that snapshot, while mock evidence attestations and a one-time checkout gate help demonstrate integrity and authorization boundaries.

Confirmed hard violation? BLOCK.

Important requirement can't be verified? REVIEW.

All hard requirements verified? ALLOW.

I've also been testing adversarial cases involving tampered evidence, mismatched products, changed carts, omitted constraints, and replay attempts.

So far, 109 synthetic adversarial scenarios have passed with zero unexpected ALLOW decisions or unauthorized mock checkouts.

Of course, this is still a prototype.

The carts and product evidence are synthetic, checkout is mocked, and no real money moves.

Getting trustworthy product evidence from actual merchants remains a much harder problem.

But building this reinforced something I've been thinking about.

If an agent saves me 40 minutes of shopping but makes me spend another 40 minutes verifying its work, have we actually solved the problem?

I don't think users should have to become the security and quality-control department for their own AI agents.

And I'm completely in support of agentic commerce.

I think we'll realize its true value when agents can operate autonomously within boundaries we've defined, without needing us to supervise every decision.

Human intervention should be reserved for genuine uncertainty, not used as a substitute for reliable engineering.

Because the real breakthrough won't be when an agent CAN spend my money.

It'll be when I can trust it to spend my money correctly without having to babysit it.

Anyway, I'm still learning and experimenting in this space, and I find the possibilities genuinely fascinating.

Curious how you guys are thinking about AI and money.

What's been your experience with AI for shopping or payments so far? Would you trust an agent to complete a purchase without reviewing the final cart?

[Bhavika @stacktraceee]

The infrastructure for AI agents to shop and pay on our behalf is starting to emerge.

But are people actually using it? And do they trust agents with their money?

If not, what's holding them back: security, control, wrong purchases, or accountability?

I'm exploring something interesting in this space. Would love to hear from people who've tried agentic commerce or have thoughts on it!

[Bhavika @stacktraceee]

A really interesting read on Natural General Intelligence and the idea of AI learning from the physical world, rather than just human-generated data.

I particularly liked how it makes you think about what it actually means for a model to understand the world.

https://t.co/q7LtIB27oP

[Bhavika @stacktraceee]

@ai_change_world Maybe the quiz could trace the relevant code path beyond the diff and check whether the surrounding code supports a definitive answer. If the evidence is incomplete, it could flag that instead of forcing a guess. I’ll explore this for future versions!

[Bhavika @stacktraceee]

Still a work in progress, but I've been experimenting with something called VibeCodeFolks. And it started with a question:

AI can write 1,000 lines of code before you finish your coffee.

But can you explain those 1,000 lines?

I've been seeing senior engineers and engineering managers discuss how vibe coding is making code reviews harder, not easier.

We've made writing code ridiculously cheap. But understanding, debugging, and maintaining it? Not so much.

An AI agent can modify your frontend, backend, APIs, and infrastructure in one session.

Git tells you WHAT changed. Understanding WHY, HOW the pieces connect, and WHERE things might break is still on you.

So I'm building VibeCodeFolks, an open-source, local-first CLI in Go.

Run vibecodefolks . after an AI coding session. It attempts to:

Explain changes with actual code references.
Visualize how affected components might connect.
Quiz you on the code AI just wrote.

And the quiz is the part I'm most interested in.

Getting a question wrong won't block you from shipping. It's not a certification or a correctness check.

But if I struggle to answer questions about a change, I'd probably want to understand that part better before shipping it.

Sometimes the most useful thing a tool can do isn't give you another answer. It's help you realize what you don't understand yet.

Now, the obvious counterargument: isn't this just spending MORE AI tokens to understand AI-generated code?

Fair point.

But anyone who's vibe coded long enough probably knows this loop:

"Fix this bug."
"Now something else broke."
"Try again."
More prompts. More tokens. More time.
I wonder if spending a few extra tokens understanding changes upfront could sometimes be cheaper than debugging something you never understood.

It's still an early experiment. The explanations, visualizations, and quizzes need improvement, and I haven't established whether this approach actually saves time or tokens.

But it raises an interesting question:

Are we getting better at building software, or just faster at generating code we don't fully understand?

Curious how other engineers and teams are handling vibe coding at their orgs

https://t.co/XqAGXYtNoA

[Bhavika @stacktraceee]

The infrastructure for AI agents to shop and pay on our behalf is starting to emerge.

But are people actually using it? And do they trust agents with their money?

If not, what's holding them back: security, control, wrong purchases, or accountability?

I'm exploring something interesting in this space. Would love to hear from people who've tried agentic commerce or have thoughts on it!

[Bhavika @stacktraceee]

@ai_change_world Noted! This is something I was thinking about too. In future versions, I want the quiz to flag questions that the available evidence isn't sufficient to answer, rather than just scoring how well you can explain the changes. Still figuring out the best way to approach this.

[Bhavika @stacktraceee]

This won’t just affect the 60K people directly involved. It’ll affect everyone watching, trying to stay optimistic about their own future.

In today’s job market, staying optimistic is a skill in itself.

[Asmit @coolcoder56]

this is so sad to read

[Bhavika @stacktraceee]

Sharing the first article here. You can follow the rest of the series by article number in ascending order :)

https://t.co/hLzQOSIy9y

[Bhavika @stacktraceee]

If you work on applied AI, it’s worth understanding how these models work internally. It’s always better to learn things from first principles.

I wrote a series on GenAI internals a few months ago and intentionally kept the math light.

Sharing the first article link in comments

[Bhavika @stacktraceee]

Jev, Laya and Kev feel like AI going back to an old idea.

Before LLMs, we built specialized models for specific jobs.

Then we tried making one general model do everything.

Maybe the future is somewhere in between:

general intelligence underneath, specialized models on top.

[Bhavika @stacktraceee]

Maybe AI doesn’t need to feel an emotion to understand it.

It can study millions of examples of grief, love, fear and learn their patterns better than one human ever could.

Which raises a stranger question:

Can AI understand something at a human level without experiencing it?

[Bhavika @stacktraceee]

Some things I’ve learned about building evals for LLM apps and agents https://t.co/LIHIQNcc6h

[Bhavika @stacktraceee]

Had a good laugh watching this video

But this is also a great example of just how far AI video generation has come.

A few years ago, generating a realistic image itself felt impressive. Video is a much harder problem because you’re not just generating good-looking frames. You need temporal consistency across all of them.

The same face has to remain the same face. Objects can’t randomly disappear. Lighting, geometry and camera motion need to stay coherent. If something leaves the frame and comes back, the model still needs some sense of what existed before.

Modern video models are getting surprisingly good at this by learning space and time together. They compress videos into latent representations, use attention across spatial and temporal tokens, and use diffusion or flow-based generation to learn how scenes and motion evolve over time.

Pretty crazy how fast we got here

[DesiEsco🇮��� @DesiEsco7]

AI will take over the jobs in India

meanwhile AI in india-

[Bhavika @stacktraceee]

One thing I never expected while going deep into LLMs...

The more I study how they're trained and optimized, the more I end up learning how to optimize myself.

LLMs were inspired by ideas from neuroscience, even though they work very differently from the human brain. That's why it's fascinating that many of the techniques we use to improve them also seem useful for humans.

Curriculum learning is just learning the basics before the hard stuff.

Knowledge distillation is reducing a complex topic into a mental model you can actually remember.

Retrieval beats trying to memorize everything.

Attention reminds you that not every piece of information deserves equal focus.

Iterative refinement is why your second or third explanation is usually much better than your first.

Context often matters as much as raw intelligence.

Maybe that's why learning AI has slowly changed how I learn everything else.

[Bhavika @stacktraceee]

I used to be confused about when to use Skills vs code for AI agents

Now I think of it this way:
• Early stage → Skills. Optimize for iteration speed
• Production → Code. Optimize for cost, scale, and reliability
Let the LLM do the thinking. Let code do the orchestration

[Bhavika @stacktraceee]

"Agents are just scripts."

That's what people say.

Then you build one for production.

Writing the script is easy. Maintaining it across countless edge cases is the hard part.

Miss one tiny condition or an important piece of context, and the entire execution can fail.

[Bhavika @stacktraceee]

One thing that surprised me while evaluating LLMs:

The same step by step prompt that helped a smaller model hurt a stronger one.

The stronger model didn't need the extra scaffolding. I was in its way.

Prompt quality isn't universal.

That's why we need evals, not intuition.

[Bhavika @stacktraceee]

Some things I've been doing lately to save on AI tokens:

• Caveman mode
• Better model for planning → cheaper model for coding
• Run /insights to spot token waste
• Create a project index/map for the codebase
• Start a new chat before context gets too big

[Bhavika @stacktraceee]

@NTFabiano It’s been 4 months since I quit short-form videos on IG and YouTube. Less stress, more control, better life. Just stopping this one thing improved other areas of my life. Never want to go back to that again.