bhasinanant

I'm definitely biased here, but the OpenClaw hype is making people disregard the economics of it all. Building Auto-CRM.com, my primary concern was building a system that runs well while not costing 200$ per month to keep up, and of course, while also maintaining security. I assume the good guys at Folk, Pipedrive, etc also had similarly priorities. A lot of good work is being done within the OpenClaw ecosystem regarding RAG and memory, but specialised orchestration process to be a more reliable system.

Forgot to add one detail: It has an Agent sidebar like Cursor/Antigravity/etc do, which basically reveals the agentic "thinking" and what actions it took and why, which allows for iterative improvements to the steering by the user. If the agent does something you did not want it to, you can tell it not to do that and it'll remember that for future conversations.

Gemini 3.1 Pro isn't necessarily the next frontier Coding model, but it is still an amazing model. Everyone has taken note of it's crazy good svg skills, but I had an idea to test this further. I needed to create a launch video for a project, and had this idea of an ongoing story going on in the background of the content, with a one-shot video of a person going about the daily office grind. When Gemini 3.1 Pro came out, I decided to try to make an animated version of it first.

So, the stack for this process looks like: 1. AI Studio Build. Very quick, great scaffolding. Already aware of React Motion and other related helpful libraries. 2. Gemini 2.5 TTS 3. OpenShot to compile the video and the audio together.

Manual work needed: 1. Getting the script together. Actually also done using Gemini in the Gemini Web App. Recorded a walkthrough of my application, and just asked it to go from there. 2. Adjusting the time allotted to each of the scenes. AIStudio created the scenes pretty well, but they followed the time durations in the ideal script. This needed to be modified to be synced with the generated Audio. Needed just one prompt though. 3. Recording the Animation. Go Full Screen. OBS Studio Recording the window. Animation runs on autoplay, and hides the player controls. Move mouse to another screen. Record and Reload page to start from 0:00. 4. Compiling the shot and audio together. New Project in OpenShot. Bring resources in. Clip Video to start and end at the correct spots. Audio is already synchronized. Export.

Total money spent: 0$

Additional Insight: I initially forgot to switch the model, and got a Gemini 3.0 generation instead. Needless to say, that chat was a lot more cumbersome and the result a lot less appealing.

I am so happy with the result that it just seems obvious to build a wrapper that does all of the steps from recording to editing for you, but would love to hear what everybody else thinks.

Zed is my goto editor when I'm not vibe coding, but that is rare these days. Their integration with Claude Code, etc really helped, but Antigravity completely pulled me away. And really, since they're catering to the same basic audience, the defaults should be the same as VSCode for most stuff. VSCode but performant would be an excellent pitch for the upcoming consumer RAM deficient world.

Dunno how they plan to get wider extensibility and community support without an embedded JS backend to support the existing Code plugins. That's where the real blocker is.

I haven't used anything else since I've found PDFGear. Have it installed on all my devices. Still surprised it isn't more known.

https://NitroQR.com Building the one stop QR Generator with a crazy amount of aesthetic customizability. Already at a pretty functional level, focusing on marketing now. Got something new cooking for the holiday season. Hopefully launching this weekend.

So, no custom scanners are really needed. But there are of course issues with scannability that become worse the longer the payload is. Easily solvable with link shortening though. The robustness centers around how modern scanners really work. No scanners really matching data module by data module. They're sampling the centroids in the approximate locations. So long as we keep the relevant centroids the right color, we can use the rest of the space as we want. There is more that can potentially be customized too, but I like how it currently looks. Might add rounded borders though.

Ohh, the good old AI generated ones. I loved them when they came out. They look really good. Each has their own merits though, in my opinion. It's more expensive at scale to generate those. The results are also non-dererministic meaning the output can vary wildly while keeping the Qar information intact through Controlnet, which is not the case with our approach. The AI generated QRs also often don't really look like actual QR Codes, until you pull a scanner out. This is why I'm not really considering adding anymore customisation to this tool. Gives it a unique visual identity.

These are really creative!

One of my absolute favourites. Here's a thread of some other great ones: https://x.com/bhasinanant/status/1956313568665571431 High-res helps the quality too, though.

There's ones that look really bad, and others that look like they were just supposed to be. Here's one of my favourites: https://x.com/bhasinanant/status/1956313568665571431

Agree with a lot here. QR Codes have however become a cornerstone of digital convenience in a lot of the world. All the more true here in India, where basically everything is using QR Codes, from payments to restaurant menus. People here are still not used to using Lens + OCR. Then, there are issues like long links or confusing zeroes and O's. QR Codes solve these pretty well, I believe.

There is a toggle to preserve color. Give it a shot!

Thanks! :)

Yep, found those really interesting when they came out. I used to run an AI startup until a short while ago. In my opinion, the cost to utility ratio still isn't there for most use cases. So, that's why I wanted to find something deterministic, without all the compute consumption. NitroQRs generate the same as normal QRs with some good old computing, and can be generated in bulk for basically nothing.

Agreed that these might not work flawlessly on T-Shirts for example. Does work on flat surfaces in my experience, but yes, there's a slight tradeoff.

It's a slight tradeoff. Digital works flawlessly though Still experimenting with how it works with print.

It's a commercial project but I'm of course not planning to monetize the QR generation. That'll always remain free. Currently adding more features that would appeal more to enterprises or businesses. Those will be the only paid features

Would love to hear your thoughts! :)

The generation is all offline. There's a "private" checkbox, which if you check, nothing goes to the server. Hope that helps :)

What I do want to understand is, why would anyone use this instead of NextCloud, for example. I'm all for having local setups though, as a general policy. It's just that from what I know, the difference in feature sets is just too huge right now.

Particularly for India, for example, all algo's have to be pre-approved, and the core principle is that you cannot operate in isolation; your broker is your gateway and your primary point of compliance.

For the US, the SEC and FINRA place the onus on the broker-dealer to supervise all trading activity, including automated trading by retail clients. This includes having risk controls to prevent erroneous or manipulative orders. An individual is not typically required to register their specific algorithm with the SEC, but they are subject to all overarching rules against market manipulation and fraud.

So, end of day, you'll still have to do a ton of back and forth with your broker, to make sure their controls cover your implementation. Not really feasible for the retail trader from the looks of it, regardless of location. But some smaller firms can perhaps utilise it.

Interesting choice to be using CPython. Very impressed with the completeness, and in particular the Risk Engine. Not having delved in too deeply yet, it does have the basics, but the markets are very highly regulated, and automated trading is always in the microscope. At Goldman, my team's whole job was to make sure the automated trading desks have all the checks implemented. It's a little spoilsport, but anyone seriously looking to use this, it's not gonna be very plug and play, at least while comlying with the market regulations.

Is it just America though? People are just going out lesser now, ever since Covid.

Bolt is my platform of choice to quickly prototype. In my experience, it generates coherent, beautiful frontends. That said, they've recently halved the daily free tokens. A token in Bolt's case is the literal LLM input/output token. Used to be 300K per day/3M per month. Now, it is 150K per day/1.5M per month. Once you are out of tokens, Bolt won't generate anything for you until more tokens(free or paid) become available. It is great tbh, but I can understand if it seems a bit too measly, but tokens especially from the frontier closed source models are pretty expensive though.