Thought Journal
A digital garden for evolving ideas. Unpolished, honest, and focused on the craft.
Gemini Argon is here
I was hoping to get back to Antigravity now that Argon is out. Turns out Argon is on the Gemini web and mobile apps, but not in Antigravity for the pro plan and requires the ultra plan to experience it there. That feels a bit off. Argon is pitched for complex, creative work, and an agentic environment is where that capability should count most. Asking a model for a product idea is one thing. Letting it read the codebase, make changes, run the product, check the result and keep going is another.
My Flo2026 Talk - Hyperagile, building products when AI joins your team.
Spoke at Nagarro’s tech fest, FLO, about building products in the AI era, where AI isn't just a tool but part of your teams. We shared a practical framework which has already seen success in reimagining agile to suit the new tech shift.
I Built My Own Coding Agent with DeepSeek Harness
You can now build your own coding agent pretty easily. DeepSeek has open sourced its Harness, so you can take it, connect it to a model, give it access to a repo and create your own coding tool. Let's call it NeetCode. At first glance, it won't look very different from Codex, Claude Code or Antigravity. Give it a repo and a coding task and it can inspect the code, search files, understand the existing architecture and start making changes. With a good model, even the initial analysis can look surprisingly similar. Now the interesting experiment is to see how close you can get to the experience of the more mature coding agents by improving the harness. You can add better repository and code search, Git integration, proper test and build execution, browser automation, API and database access, documentation search and even another model to review the changes.
AI Beyond LLMs
Been playing around with Jev from TypeSafe AI recently. For the last couple of years, whenever we say AI, we almost automatically mean ChatGPT, Claude, Gemini and other LLMs. Give them some information, ask a question, and they give you an answer. Incredibly useful. But while experimenting with Jev, I noticed how often we're asking AI to do something that doesn't actually need an answer in words. Take an invoice coming into a company. You don't need AI to write out "based on the information available, this invoice appears to be legitimate..." You just need it to tell your system the invoice is 97% valid, 2% duplicate, 1% needs review, and let the software take it from there.
Beyond Chunks: Why I'm Rebuilding My Document AI Stack
Building AI products is humbling. Every few months, something arrives that makes you question the assumptions your last project was built on. Last year, I built
AI enablement shift
Been part of a few AI enablement discussions lately. Teams are excited about speed. Shipping faster, building faster, iterating faster. Speed is real. But I think it's the wrong thing to be excited about. When execution gets cheap, the bottleneck moves. It doesn't disappear. It moves to how clearly you can think about the problem. A vague idea used to get lost somewhere in the build cycle. Now it shows up immediately in the output and you have to confront it.
Debugging mental model for agents
Switching models to fix your agent is like tweaking your macros on 4 hours of sleep. It won't give the desired results. When something breaks, the first instinct is to switch models. GPT to Claude. Claude to Gemini It still gives similar results, often underwhelming.
Supabase and a relook at vendor risk
Vendor risk Is not just about outages or pricing A few weeks ago I ran into an unusual issue. A few of my projects stopped working overnight. The reason was unexpected. Supabase had been blocked in India. There was no warning and no migration window, something I had neither anticipated nor planned for. The incident was another reminder of how unpredictable the environment around software systems can be. The biggest learning for me was how vendor lock-in can introduce unexpected risk. Today it was Supabase, but the same thing could happen with an LLM provider, a cloud platform, or any other critical service in a project. A ban is just one scenario. Vendors shut down, get acquired, change pricing, or introduce breaking changes in their ecosystem.
Chaos is a Ladder, not a Pit
The markets are currently overreacting to every little update from major AI labs. Neither is the SaaS story over, nor are IT companies being destroyed already. Furthermore, the success of agentic coding in controlled environments does not mean software engineers are no longer needed. Recent developments have primarily created chaos. Enterprises are unsure how to navigate this shift. They need help and guidance through this, the companies best positioned to advise them are those that have already invested in understanding their specific niches, architectures, and current challenges, not because they have all the answers, but because they know which questions actually matter. Extending that expertise to help them integrate emerging AI technology is the most sensible path forward. While the shift is massive and sudden, the software component of SaaS may look fundamentally different in a few years. But the underlying business problems it solves will remain. Services provided by IT firms will adapt to new opportunities around that.
The QA Shift for probabilistic systems
With the rapid progress of coding agents, getting a working prototype or MVP out quickly has become significantly easier, even for small teams. At the same time, baseline expectations for software have risen sharply. What felt like a delighter in 2023 barely meets expectations today. And that is assuming the product avoids the usual blind spots. When building a product meant for real business users, these challenges cannot be treated as afterthoughts. In finance, legal, healthcare, or enterprise workflows, failures are expensive. Traditional software breaks and throws errors. AI breaks differently: confident nonsense, risky actions, silent data leaks. That is a fundamentally different failure mode. Classic QA assumed determinism, but binary pass/fail logic no longer holds. You are evaluating behavior quality, not just code correctness.
Game theory of model release
The game theory between the big AI labs has gotten impossible to ignore lately. It's interesting to watch for sure. But if you're trying to stay up to date with the latest, it's also quiet exhausting. February 5, 2026 was a perfect example of how intense this has become. Anthropic rolled out Claude 4.6 Opus with a massive 1M-token context window and a new Agent Teams feature for parallel coding workflows. Within minutes, OpenAI countered with GPT-5.3-Codex, highlighting a 25% speed boost and a top spot on the freshly released Terminal-Bench 2.0. This is textbook Nash equilibrium. If one company launches and the other stays quiet, the first mover owns the narrative. If both launch simultaneously, attention gets split, hype gets diluted, and everyone waits to see how things settle.
Open AI Codex App for Mac
I ignored openAI Codex for a long time for one simple reason, the $200 paywall. While everyone was talking about OpenAI’s “agentic” future, it was locked behind a Pro tier that cost 10x more than tools like Cursor or Windsurf. By the time it dropped to the $20 Plus plan, I had already built my workflow around Claude Code and Antigravity. Still, ChatGPT nudging me with to download the new app with free credits eventually made me give in to install the standalone app without ever trying the CLI or cloud version. First impression? It has very different vibe.
The Real Risk in AI Isn’t Adoption. It’s Ignoring It.
We talk a lot about ROI in AI. Returns. Efficiency. Cost savings. But lately, I’ve been thinking more about something else: RONI. Risk Of Not Investing. We’ve seen this pattern before. When mobile took off, BlackBerry didn’t move fast enough. When the internet reshaped photography, Kodak hesitated. When streaming arrived, Blockbuster stayed comfortable. None of them failed overnight. They just fell behind, slowly. AI feels similar today.
Critical Thinking in the age of AI
AI is everywhere now. Answers are cheap. Fast. Confident. What’s becoming rare is judgment. The risk isn’t that AI makes mistakes. The risk is that we stop thinking because the answer sounds convincing. In the age of AI, critical thinking matters more, not less.
Vibe Coding ≠ Software Engineering
Vibe coding != Software engineering for that matter even Coding != Software engineering I still spend a lot of time defining instructions and logic for agents. What I worry about less now is syntax missing semicolons or extra parentheses. Its especially tricky developing fluency in multiple programming languages to build full stack apps. Most programmers are predominantly backend or front end focussed, though there are rare exceptions.
Reliable AI Pitfalls
To achieve reliable AI outcomes, businesses must navigate a complex landscape of technical bottlenecks and strategic organisational shifts. The sources describe this current state as a mix of rapid progress and significant "choke points" that require targeted investment to overcome.
Will B2B SaaS survive ?
B2B SaaS thrived when the Buy vs build decision heavily tilted towards buying, as building was extremely expensive and simple subscriptions provided a way to get on board without shelling out a fortune, it also allowed companies portability and an option to exit if better tools were available. Contrasting it with enterprise products with long onboarding cycle and costs running into hundreds of thousands subscription based products were a no brainer, particularly for the small and medium enterprises. They were not catering to the localised nuances or custom requirements but most companies were fine modifying their own processes to suit that.
Prototyping in the age of AI
AI tools have transformed prototyping from a slow, technical process into something much more accessible. Now, you can describe what you want in plain English, and AI can help you build it—whether it's a website mockup, an app interface, or even working code.
Shipping speed with agents
The gap between idea and shippable code has reduced, and the progress an average person can make is vastly different from what it was in the early part of the year. With earlier versions of tools like Cursor, Gemini CLI, and Claude Code, builds still needed multiple rounds before they felt push-worthy. A lot of back and forth. A little more nudging. The updates in the later months like Antigravity, Gemini 3.0, and Opus 4.5, it’s different.
Deep Research with Agents
It feels like the first serious admission that chat is a poor interface for complex work as it needs constant supervision. It creates friction when the task itself needs focus, time, and synthesis. The newer agentic models (Gemini’s latest updates, GPT-5.2) change that dynamic. You give them a goal. They go away for minutes or hours. They return with a concrete artifact you can review, refine, or ship.