was fireside with @otooligan from @agoric and @KushagraCrypto at @EthereumDenver. talked about the future of chain abstraction and the role of orchestration in solving blockchain fragmentation. right now, web3 is a mess of different chains, liquidity silos, and execution

If you’re building, dreaming, or just want to jam - DMs are open. 🔗 📸 Onwards. ✨.

What’s next?. A pause. Some stillness. Some travel. Then… who knows. New ideas. New questions. New frontiers.

To everyone at CoinDCX I got to build, fight, learn, and grow with - you made this chapter unforgettable. And to the community, the customers, the believers - you gave all of this meaning. 🙏.

To the engineering team - you’ve been the heartbeat. Craft. Consistency. Chaos-resilience. You made it real. You made it work. You made it matter. Thank you. 🙏🏽.

Joined when CoinDCX was a scrappy startup. Helped across product, ops, CX, and security before focusing on engineering. We built core infra, scaled systems, managed risk, shipped fast - and learned even faster.

After 5 wild, intense, unforgettable years - I’m moving on from CoinDCX. From 90k to 18M+ users. From 15 engineers to 200+. Through bulls, bears, bans, and believers - we built. A thread on the ride, the people, and what’s next 🧵👇.

hyperliquid-native isn’t a feature, it’s the direction infra is compounding toward. after a year of building:.- 15+ hyperliquid projects integrated via okto sdk.- hyperevm sdk launched (10x faster build cycles).- eth denver hyperliquid track live ($10k builder bounties).-.

RT @PluralisHQ: We've reached a major milestone in fully decentralized training: for the first time, we've demonstrated that a large langua….

No wallet connection flows, chain switches or gas management. Just embed and ship. 17M wallets secured, 50+ teams building on top.

Platform PMF needs to be seeded. Owning both platform + app lets you control liquidity, user behavior, and infra evolution. Most chains outsource their most important feedback loop.

For devs: okto isn’t just embedding wallets. It abstracts the entire stack. -30+ chains behind one interface. -gasless flows (usdc/native sponsor). -intent-based cross-chain execution. Dev time drops, user conversion rises. Infra complexity stays where it belongs: Off your app.

RT @smtgpt: 🚀Launching the CoinDCX Crypto Trading Masterclass!. As India ka Crypto Coach, @CoinDCX has partnered with SEBI-registered resea….

RT @levie: Prompting doesn’t show any signs of letting up as a useful skill. The variability in the quality of what we can get Agents to do….

8/.underneath all of this:.- prompting is turning into kaizen. the people running the loop daily become the best at improving it. prompt engineers today aren't writers, they're system trainers. this is how vertical ai will be built.

7/.the real IP isn’t just the prompt, it's the eval set. evaluations = why the prompt is written the way it is. without evals you can't improve or debug reliably. startups that own high-quality evals own the quality bar.

6/.forward-deployed engineering (FDE) is back.many vertical ai founders are now basically acting like Palantir FDEs:. - sit with real users.- understand edge cases.- bake that into the prompts + evals.- close huge 6-7 figure contracts quickly. deep context = fast sales.

5/.as models scale, different models evolve different "personalities":. - Claude: more flexible, high-agency.- GPT-4o: more rigid, strict rule follower.- Gemini: better at nuance and exceptions. picking models is becoming like hiring employees.

4/.metaprompting is the unlock.instead of hand-tuning prompts, you use the LLM itself to improve the prompt:. - feed failures back into the model.- ask it to rewrite/refine based on examples.- let it generate better versions of itself. prompt folding loops.

3/.worked examples are critical. for hard tasks (like bug finding, N+1 query detection etc), you don't describe the pattern in prose. you just show many examples. the model learns better through examples than vague descriptions.

2/.the best prompts aren’t just text walls. they’re structured like code:.roles defined. - step-by-step plans.- strict output formatting.- escape hatches when unsure. the output format often uses XML or markdown-style tags to guide the model.