Notes · how I think

Signals from the products around!

I write about AI, product and growth: what changed, why it matters, and what I'd build with it.

Follow on LinkedIn ↗
01Product management isn't safe from AI. It's being rewritten.Careers · Oct 202602Seven motion graphics, and not one video editorAI tools · Oct 202603Jev, a model that decides instead of writingAI models · Sep 202604Swiggy Money on MCP: agents can now payAgentic commerce · Sep 202605The ₹10 nudge that built trustPricing · Sep 2025
01 · Careers · AIOct 14, 2026

Product management isn't safe from AI. It's being rewritten.

"Product management is the safest job from AI" is the popular take. I think it's half the picture: the title is evolving, and the job underneath is being rewritten.

The PM's core job was translation: customer to spec, spec to design, design to code. AI makes that translation cheap and quick, so the role moves to wherever judgment is still scarce.

01
AI PMCloser to the models. Owns specs, evals and prototypes for AI features.
02
Forward deployed PMCloser to the customer. Embedded in enterprise deployments, deciding which asks change the core product.
03
Full-stack builderCloser to the code. Idea to launch, end to end. This is where I'm placing my bet.
Titles are a lagging indicator. Judgment is the leading one.
Read the full post on LinkedIn ↗
Illustration: three directions for the modern PM — AI PM, forward deployed PM and full-stack builder
02 · AI tools · MotionOct 7, 2026

Seven motion graphics, and not one video editor.

I made seven motion graphics in a weekend with Claude Opus 5.5. Instead of generating a video, it builds the film frame by frame in code, and it can work from your real product: reference videos, brand assets and design systems.

Give it an evaluation loop and it critiques its own output, fixes the weak moments and renders again.

Prompt→Critique→Improve→Render→Repeat
Read the full post on LinkedIn ↗
Made with Claude Opus 5.5
03 · AI models · AutomationSep 30, 2026

Jev, a model that decides instead of writing

40–200×faster
~445×cheaper
$0.042per M input tokens
TypeSafe AI's own figures, versus today's LLMs.

Diogo Almeida, co-inventor of RLHF, launched Jev: TypeSafe AI's first "System One" model, built to make fast, structured decisions that software can use directly.

It's trained to attach honest probabilities to its answers, and it answers every question in one pass instead of token by token. It can't generate text, so it won't replace LLMs. It's a decision engine, used on its own or wrapped around an LLM as a judge or router.

Best fit for now: conditional routing, like picking the right model per request or sending support tickets to the right team.

Read the full post on LinkedIn ↗
TypeSafe AI playground: Jev classifying a personal-finance question with probabilities
My experiment · Jev inside a financial decision workflow
04 · Agentic commerceSep 5, 2026

Swiggy Money on MCP: agents can now pay.

I ordered dessert from Swiggy Instamart without touching my phone. Until now, AI agents could do everything up to checkout, then stop at a payment QR code. With Swiggy Money on MCP, they finish the journey, payment and confirmation included.

01Add Swiggy MCP to your AI tool
02Top up Swiggy Money
03Chat and confirm the order
04Done, paid in the chat
Read the full post on LinkedIn ↗
05 · Pricing · TrustSep 26, 2025

The ₹10 nudge that built trust.

While booking a Rapido, the app nudged me to cross the road, and the price instantly dropped by about ₹10.

Instead of adjusting fares behind the scenes, Rapido showed me how to get a better deal: a small interaction that gave me transparency and control. Compare that with fares that seem to rise when your phone battery is low. Even as a perception, that erodes trust.

Done right, a small product decision turns a routine transaction into a moment of delight.
Read the full post on LinkedIn ↗
Rapido app: check your pickup pointRapido app: fare before moving the pickup pointRapido app: lower fare after crossing the road