Kiran

PongFox is the world's first #tabletennis robot integrated with a pad! Now, not only can you simulate serving and follow-up shots as if facing a real opponent, but our pad also tracks your accuracy by counting how many balls land on it during drills. This direct feedback offers a tangible measure of consistency, pushing your practice to new heights. Elevate your game with intelligent analytics. #NextLevelTraining #SmartTableTennis #india #hardware #Robot

@nikhilv Most of the towerpro available here are clones, look at hitec from robu and see if it works for you

Does anyone in India building hardware have access to TowerPro MG996 motors - 180 Degree? It’s a critical component for robotic arms (360 degree don’t work, they’re meant for continuous motion like wheels - and stockists don’t know the difference) please help?

If you’re running OpenClaw on OpenRouter and memory search keeps returning “No matches”, check your embeddings. OpenRouter handles chat, not embeddings. Without a separate embedding provider configured per-agent, OpenClaw silently skips indexing. No chunks written → no memory retrieved.

@akshay_pachaar Any recommended specs like RAM for hosting alongside openclaw

@DanKulkov We use trello and I have set this up and started using it recently , have not seen any issues so far

has anyone done it? > openclaw > connect to trello > trello has free plan & free api > multiple agents work in organized tasks > you can observe everything / review work good idea or nah?

OpenClaw on my local system using GPT-5.2 is smarter than the one I’ve deployed on the server with OpenRouter auto. Routing across different models seems to perform really badly for agentic workflows. Underlying models matter a lot.

I guess it was multifold and hard to explain but here is a gist, first there was an update to the config file with wrong model names and streaming options, this got triggered by an agent and it got into some wierd memory saving issue. the model got fixated on the the agent issue and could'nt understand the main config file, I was also parallely trying out in claude (I use openai in openclaw), I guess the error was pointing to the symptom and it did not understand the root cause

@mkirank Curious to know what limitation you hit. Was it some domain knowledge, lack of following instructions, or something else.

OpenClaw has been borking since morning. Now I’m deep in a rabbit hole trying random fixes. I hate when I’m reduced to black-box testing. Earlier I’d read the codebase and dive in. Now I’m spending more time trying to “prompt” the LLM to fix it than actually looking at the code. Not sure that’s progress.

this is an interesting one , llm's are basically next word predictor and work well on something which there is enough history/context, openclaw is new and while some of the generic errors it might be able to resolve, I feel for things like in this case the intern has hit the limit. PS : I was finally able to fix it by manually undoing the changes it had done

@mkirank That said, sometimes you do have to intervene and find out why it's getting stuck somewhere and perhaps help it out.

@vybhavram @harshilmathur @levelsio @supermemory @mem0ai This is a interesting test and I am curious to know how it goes

I'm giving @mem0ai and @supermemory side by side in an A/B test now with my openclaw inspired deployment. Curious to see results. @harshilmathur - Were you able to quantify precision/recall or was it a more qualitative test? I have some custom stuff include memory router & feedback router, which is why I had to create my own version of an agentic framework (heavily inspired by OC)

How did you guys fix persistent memory with OpenClaw? My bot keeps forgetting stuff, I already have qmd installed

@onlyequities There is always pongfox when you don't have anyone to train with 😀 x.com/mkirank/status…

Who wants to play? Humble TT setup

A lot of things that take off aren’t technically groundbreaking. They just remove friction. OpenClaw is a good example. The core idea isn’t magic. But it works out of the box. Telegram, WhatsApp - done. That 5-minute start matters more than clever architecture along with a bit of luck ;-) .

I’ve started using a few internal agents (openclaw) in our workflow. An ops agent to automate follow-ups and update trello. A “chief of staff” agent to add context when planning weekly goals and other product decisions. Also tracking my table tennis and gym training with TTLog and GymBro. The only way to know what can be automated is to try. Even a small reduction in cognitive load matters. Let’s see how this plays out.

@SHIVAJIBOSS199 Through our website

@mkirank Through Twitter?

Everyone in my team joined as interns and grew with us. They haven’t worked elsewhere. Sometimes I realize we’re discussing things that require context from larger teams, distributed orgs, or past failures. If you’ve never seen those environments, it’s hard to imagine why certain processes matter. Sometimes I forget that what feels obvious to me was learned over time working in a lot of different environments

AI is great at writing code. But we spend more time debugging than writing. And debugging is where AI still struggles most. Sure, it can generate unit tests. But the real bugs live in integration, where system A’s output subtly breaks system B. In massive codebases, the hardest part isn’t “fix this bug.” It’s “which of these 500 files is actually causing this behavior?” Maybe I’m out of touch, I haven’t worked in large codebases at bigger companies in a while. There aren’t enough voices here sharing that reality to get a balanced view.

It is satisfying to know that a few clubs are completely dependent on the robot for their training. But this is adding extra pressure on our small team. customers are hard to deal with when support expectations are real-time. At the same time, we are handling supply chain, production, export/import documentation and adding new SKUs , and each new SKU adds to the support load. The more complex and moving parts in the hardware, it just becomes harder. Sometimes I question my life choices doing this while there are so many exciting things happening in AI. Maybe the money is easier there. But here, the learning compounds differently. Whether it’s “worth it” , who knows.

Here’s a short clip from the Peakst8 festival, people trying out the PongFox robot, many of them playing table tennis for the very first time. I’ve pitched this idea to many event organizers. Most of the time the focus is on the event itself, but what often gets missed is how to keep the audience engaged instead of just watching. Thanks to @kmr_dilip for giving us the space and helping out multiple times without any expectations. I’ve never even pitched to them. Rare to find teams that genuinely cheer for founders, even outside their portfolio.

Wife out of town and daughter skips school, my mind thinks it’s Saturday and I’m pinging people to meet and play and they are replying with let’s do tomorrow

We also need to change how this happens and how fast. Time is the biggest constraint for students without financial runway. We should look at how AI can compress this learning cycle. One idea is to identify the subset that is actually open to change and combine psychology, counselling, and market reality. Some of them need tough, honest feedback. A brutally honest agent could show them where they really stand in the market today, what the gap is, and what it will realistically take delivered in a way they can understand and act on quickly.

@mkirank Totally agree. What can be done to get this new narrative to colleges?

For almost 2-3 decades, this was the default escalator for tier 2 and tier 3 Karnataka talent. Finish a degree -> join a testing or full stack course at places like JSpiders or QSpiders -> crack a small service firm, -> then get body shopped into a larger IT company. It was a bit slow, but it worked. It created some sort of predictability. AI is breaking this pipeline quietly and turning it irrelevant. Testing scripts, CRUD based full stack work, basic automation, repetitive backend tasks. These were never high leverage skills, but they were employable skills. Now AI tools can do 60 to 80 percent of this work faster, cheaper, and without fatigue. Service firms will not say it openly, but their hiring behaviour already shows it. Fewer freshers. Fewer benches. More expectation from day one. What worries me more is not AI itself. It is the time lag in understanding. In Bengaluru tech circles, people are already talking about copilots, agents, prompt driven workflows, and outcome based roles. In tier 2 and tier 3 towns, the conversation is still about which course has better placement, which institute promises 100 percent jobs, which syllabus matches interview questions. That mismatch is definitely going to hurt! I see resumes flooding WhatsApp groups with the same keywords. Manual testing. Selenium. Java full stack. React basics. It seems like we are still seeing a system that is training people for yesterday’s jobs. And these youngsters are not lazy or incapable. They are rational actors operating on outdated information. The tragedy is that AI actually lowers entry barriers if taught right. A sharp youngster from Davanagere or Kalburagi with good thinking skills can now do things that earlier needed years of experience. But nobody is teaching them how to think with AI, how to use it as leverage, how to move from using AI for writing covering letter to problem solving. If this gap is not addressed soon, we will see a painful phase. More trained people. Fewer real entry level roles. More blaming of the system, or the market, or fate. The conversation has to shift from learning tools to learning judgment. If not, AI will not just disrupt jobs. It will break a long standing social contract that tier 2 and tier 3 Karnataka (applies to tier2/3 India too) trusted for upward mobility.