yeet
43 posts
yeet
@DabbuMothsera
Learn,Build,Grow👾 Working @IISc
Bangalore Katılım Aralık 2024
71 Takip Edilen9 Takipçiler
2 years ago, we handled just 150 phone calls in an entire week at @bolna_dev.
Last week, we handled ~6.5 million. That's a 35,000x increase.
Looking at this, it's easy to think there was one breakthrough. There wasn't. It were thousands of iterations, compounding over time.
English
Why is Claude's STT so bad ?🤡 atp it don't make no sense
@claudeai @AnthropicAI
English
we’re in bangalore.
we’re four 4 with room for 6.
we philosophise software and build complex products.
hiring 2 engineers to live & work with us:
- epistemic humility
- curiosity
- rigour
- experience in shipping software
11LPA base, AI credits, tickets to bangalore all paid for.
English
Ran an independent WER benchmark on @Ringg_AI's own public dataset (10k Hindi samples) vs ElevenLabs, Deepgram, Sarvam.
Ringg wins on accuracy (0.16 WER vs 0.175/0.217/0.232). But every provider's error rate roughly doubles on Hindi-English code-mixed speech , an industry-wide hard problem, not Ringg-specific.
Real implication for anyone running Indian voice AI: expect meaningfully higher error rates on natural business calls (brand names, English phrases) vs clean benchmark numbers.
Github: github.com/lazerbeam47/Pa…
#VoiceAI #ASR #IndiaStartups
English
@DabbuMothsera @maitreya_wagh @bolna_dev Looks cool! Would love to understand this more. Can you check your DM
English
BolnaMonitor - real-time drift detection for Bolna's
ASR → LLM → TTS pipelines.
Across 20+ provider combos and 10+ languages, one
degrading combo affects thousands of calls before
anyone notices.
This detects it in ~30 min using Bolna's existing
webhook data. Zero code changes needed.
→ Z-score drift alerts (per combo, not global)
→ Error rate + absolute latency thresholds
→ Live sparkline dashboard
Live demo: bolna-monitor.onrender.com
GitHub: github.com/lazerbeam47/Bo…
@bolna_dev @maitreya_wagh @xan_ps - Excited by what Bolna is building for India's voice stack - happy to chat / contribute / integrate further!
#VoiceAI #Bolna #LLM #Observability
English
@tanishasharmax @maitreya_wagh @xan_ps @bolna_dev Hey Tanisha, really appreciate it! Blog sounds great - happy to discuss more in the DMs.
English
@DabbuMothsera @maitreya_wagh @xan_ps @bolna_dev ok this deserves more eyes. drift detection per combo is basically canary analysis for voice infra and nobody’s doing it. want to write this up for our blog / demo it on a video? dming you
English
@_Suresh2 @maitreya_wagh @xan_ps @bolna_dev Hmm makes sense - quick fix is a schema validator on ingest that alerts if the expected schema is modified. Will push it asap.
English
@DabbuMothsera @maitreya_wagh @xan_ps @bolna_dev zero code until the webhook schema shifts and the monitor silently ignores the new field
English
bookmark this.
hidden $10K/month service:
> find businesses invisible in ChatGPT
> run CrowdReply MCP
> it fixes everything automatically
> invoice $10K
10 clients = $100K/month
we're still early.…
Biemjo@Biemjo
Today we're introducing Claude Fable for Marketing. We built an MCP that ranks your brand in AI search. Just talk to it. It finds where you're invisible, and it'll fix it for you. this is absolutely wilddd
English
@hijunedkhatri Hey, I am in. I have been building voice AI systems, Samvaad is a real-time AI copilot for live calls with Deepgram STT, hybrid NLU with Gemini, speaker diarisation, all under 2 seconds end to end. What is the role?
English
I am hiring for one, if you are a dev currently working on VoiceAI, I have an awesome role for you.
Jyoti Bisht@joeyousss
Why is everyone joining VoiceAI companies?
English
Just read Sarvam AI’s blog on evaluating Indian language ASR - it showed me the depth you have to deal with.
If you work on Voice AI or have any interest in Indic speech tech, you should definitely read this.
What makes Indic languages so different from plain English: heavy code-mixing, fluid colloquial vs formal styles, multiple valid spellings/scripts for the same words, and rich suffix/agglutinative structures. All of these completely break traditional WER/CER - a perfectly understandable transcription often gets terrible scores.
Sarvam’s LLM-WER + Intent Score + Entity Preservation approach is so much more practical and honest for real Indian use cases.They open-sourced the tools too. Solid work!
Must read: sarvam.ai/blogs/evaluati…
@SarvamAI
#IndicAI #ASR #VoiceAI
English
Built this using:
-LangGraph for the multi-agent architecture
- Mem0 + Chroma for long-term memory
-Gemini 2.5 Flash
- APScheduler for reminders and daily morning briefings
It has task tracking, goal tracking, a coach agent, smart reminders, and even gives you a personalised morning briefing.
Runs 100% locally on your laptop. Fully private.
Check it out: github.com/lazerbeam47/Gh…
English
Guys,do you remember the movie Ghajini? The one where Aamir forgets everything every 15 minutes? 😭
Well, I got tired of forgetting things myself, so I decided to fix the problem.
I built Ghajini - a multi-agent Telegram assistant that remembers everything for me.
You can tell it about your goals, tasks, plans, ideas, or commitments,basically anything you want to , and it builds long-term memory around them. It tracks what you've said, follows up when needed, and sends reminders so things don't slip through the cracks.
Think of it as a memory layer for your life, living right inside Telegram.
The name was too good to pass. 😂
GIF
English
@harshil3004 app.notion.com/p/Hi-I-m-Dabbu…
this is my notion portifolio do check out
" Risk "is one hell of a board game
English
we're hiring for cracked full-stack engineers (12LPA) and Junior ML Engineers (12LPA) to work on internal tooling and rapidly scaling 3DCV pipelines at contineu.
if you love shipping quickly, and want to work with some of the smartest people building frontier 3DCV pipelines in construction (led by my amazing CTO @kantineu ), we want to talk to you.
experience/education no bar - we only evaluate skills and thinking ability.
brownie points if you love playing board games or will be a part of the company modded minecraft server (maintained intensely by @kantineu :D )
comment down below or DM with any proof of your work that you are proud of, and we'll reach out to you.
English
Been building a multi-agent system and noticing most people just slap on a memory layer and call it done Memory is cool, but almost nobody talks about the learning layer , actually making the agents get better over time instead of repeating the same mistakes with fancier recall.What’s the best way you have found to add real learning to multi-agent setups? I am experimenting with reflection after every task + saving winning trajectories as reusable “skills.” But I would love real opinions ,what’s actually working for you?Drop your suggestions #AIAgents #MultiAgent
English
Built this today🧬 Prompt Darwinism I let prompts evolve by themselves. Bad ones die, good ones breed & mutate. After a few generations you get way better prompts. Simple loop: Solver → Judge → Selector → Breeder. Built with LangGraph + Gemini github.com/lazerbeam47/Pr…
English
Claude Opus 4.6 you beauty
turns out on reconnect I was creating the same remote player twice. One sprite got tracked, the other became an orphan that never received updates or got removed.
Server showed 2 players but Client had 2 players and a ghost
aditya🇦🇷@adxtyahq
been debugging a multiplayer bug since last night 😭 when a player refreshes, the server correctly keeps only 2 players alive. the client, however, decides to preserve the old character, spawn a new one, and create a universe where the same player exists twice. spent hours tracing player joins, leaves, render lists, IDs, eviction logic and event ordering even AI has suggested multiple fixes and we're still finding new ways to duplicate people. using Colyseus SDK btw. if you've run into something similar or have any ideas, hmu.
English