I save things I never find again.
I build a product about that with friends.
In between, I get stuck, learn things the hard way, and post the unglamorous version.
I'm Harshit. Student in Delhi × Chennai.
This account is my build log - Tagzzs, the daily mess, and whatever I figure out along the way.
Expect:
- learnings from shipping
- problems that don't make it into founder posts
- the occasional jokes
- the journey without the polish
If that sounds useful, stay.
Say hi below.
1
4
679
Anthropic locking claude fable 5.2 thinking blocks so no other model or unlinked account can read them shows how real model lockin is getting keep your saved context and project notes in plain markdown you control instead of relying on vendor specific memory formats.
2
2
64
Saturday check-in for builders: what are you working on this week?
Drop your link and one line on what it does, and I'll check it out and leave real feedback.
7
4
106
I built a new audio feature in Tagzzs so you can save any type of audio alongside your links and files.
When you listen to something that teaches you a lesson or carries a real emotion, you can save it in Tagzzs and have it resurface based on your context whenever you need to revisit that exact feeling or learning from the first time you heard it.
2
3
29
When you ask Kai-AI a question in Tagzzs, it answers using what you have saved and it attaches numbered references right inside the reply.
We built it that way because getting an answer without seeing where it came from doesn't solve the real problem of forgetting what you saved in the first place.
Showing the exact reference lets you see which saved link or file Kai pulled from so you can click straight back and revisit the original page whenever you need it.
1
2
37
One small technical thing I worked on this week was making sure Tagzzs catches duplicated saves cleanly.
People often save the same page or file twice without realizing they already grabbed it week ago, so I fixed how our backend checks and deduplicates incoming saves before they clutter up your space.
Follow along for the daily build log as we keep improving #tagzzs.
1
4
51
try it free → tagzzs.com ($10 worth of credits, no card needed)
1
21
You set reasoning effort to max on every api call.
Your response time drags.
Your accuracy can actually drop.
Anthropic's new claude fable 5.2 evals showed the max effort setting costing way more while scoring lower than xhigh on coding tasks, proving that letting a model overthink a simple problem just burns tokens without making the output better.
4
170
Harshit Mathur | Building Tagzzs retweeted
4/5 — Tagzzs 🧠
Save any link, note or snippet in one click, then recall it in plain language — and carry that memory into ChatGPT, Claude and Gemini.
Useful for people who save a lot but can never find it again.
Built by @Col_ASY
🌐 tagzzs.com
Quick question for people who build things: where do you keep a link, code snippet or doc you know you'll need again next week?
I used to text links to my own whatsapp, leave tabs open so I wouldn't lose them, and save instagram reels that disappeared into a black hole. Now I just dump everything into Tagzzs, from technical docs right down to cooking recipes so I don't have to scroll for good amount of time when i want to make something.
Curious what your current setup looks like.
1
2
81
The very first thing I built on Tagzzs was the initial frontend (which was not the best version lmao :_( ).
Our designer at the time handed over the layouts, and my first job was turning those static screens into a working interface. Right after that was up, I moved over to backend with Ayush to wire everything together.
Sharing the real engineering and marketing log here everyday, so follow along if you want to see how we are building this.
2
3
152
Harshit Mathur | Building Tagzzs retweeted
applied to yc winter 2027 with tagzzs, building it with @Aditya_tagzzs
this is the video we sent with our application
whatever happens, posting it here so you can see where it started
#tagzzs
Harshit Mathur | Building Tagzzs retweeted
i reverse engineered wispr flow, then used deepseek to fine-tune LFM3-1.7B. the result is on par with the wispr flow models you pay for, and it's free, open source and runs entirely on your mac.
key-up latency is 65 ms on my m4 max vs ~520 ms for wispr flow on the same clips. a 6 minute yap session pacing around my office still lands in about a quarter second.
> rust backend with a native metal engine: each decode step runs as one fused megakernel
> a dflash2 speculative drafter is built into the engine, trained on my local 4 x rtx pro 6000 server
> parakeet transcribes, LFM3-1.7B refines. both stream while you talk, so key-up only waits on the last chunk
> never leaves your macbook, so slow internet or bad ping is no longer an issue
> import your own dictionary of terms for your workflow
> i picked every default myself. it feels amazing out of the box
> adapts to the app you're in: discord is lowercase casual, slack and email are formal, claude code gets clear prompts
> training data: 77k transcripts of public audio and synthetic dictations, each rewritten by deepseek v4.1 flash 1017 following a written spec of what "what you meant" means, in 4 writing modes. none of my own dictations went into it
> blind judge (claude opus) on an hour of lectures: phonon's text preferred 90 times, wispr's 23
> built with claude opus 5.5 [xhigh]
brew install --cask infatoshi/tap/phonon
Most AI memory tools trap your knowledge inside one chatbot. what happens when you're googling an error at 2 am or replying to a client in gmail or whatsapp?
We built tagzzs to follow you everywhere : save any pdf, yt video, or doc in 1 click, see it resurface automatically right next to your google search results and type /tagzzs to inject it into Claude, ChatGPT or Gemini in 2 seconds. Whatsapp and Email memory are up next.
2
34
Harshit Mathur | Building Tagzzs retweeted
you build every feature a loud user asks for.
your ui gets crowded.
your core workflow gets buried.
power users always ask for more buttons and custom views, but if a new user can't figure out the main action in five seconds, all those extra options just push people away.
you build every feature a loud user asks for.
your ui gets crowded.
your core workflow gets buried.
power users always ask for more buttons and custom views, but if a new user can't figure out the main action in five seconds, all those extra options just push people away.
injecting text into chatgpt or claude from a chrome extension is never simple as setting innertext on the input box.
modern chat inputs run on prosemirror or react state, so if you don't dispatch a real input event after inserting the text the app thinks the box is still empty when you hit enter.
3
54
ai coding tools don't fix bad engineering habits, they just speed them up.
if someone writes messy code without understanding the system, giving them cursor or claude just helps them ship a month of tech debt in a single afternoon.
Both drafts are ready to copy and drop straight into the composer.
3
47



