Sumukh (@Sumukhmg) profile picture

NITK | Growth | Product

Mysuru
Joined October 2021
Photo by Sumukh (@Sumukhmg)Photo by Sumukh (@Sumukhmg)Photo by Sumukh (@Sumukhmg)Photo by Sumukh (@Sumukhmg)Photo by Sumukh (@Sumukhmg)Photo by Sumukh (@Sumukhmg)Photo by Sumukh (@Sumukhmg)Photo by Sumukh (@Sumukhmg)Photo by Sumukh (@Sumukhmg)Photo by Sumukh (@Sumukhmg)
Algorithms Every Programmer should Know 🧵👇
39
549
26
2,178
Some clarifications on various themes being discussed about @SarvamAI in the last few days: 1. Government Support: The IndiaAI Mission under MeitY has played a pivotal role in bootstrapping India's AI ecosystem. It gave ~20 startups subsidised access (40% discount) to GPUs for training models. We used that access to train Sarvam 105B from scratch, which we completed and open-sourced this February. The government did not pay Sarvam directly, instead like with every supported startup, the subsidy went to GPU cloud providers, helping seed AI infra in the country. 2. Sarvam Token Factory: Open-weight models saw massive adoption globally this year. They offer a cost advantage over closed models and let enterprises own their AI stack. But this choice is not available to all Indian enterprises, which often need data to remain within the country and low latency for interactive use cases. Sarvam Token Factory is one of the first large inference offerings in India. We serve multiple open-weight models on GPUs we have privately procured, without any government support. We have optimised datacenter scale inference to make our token cost comparable to or lower than established providers such as Fireworks, Baseten and Together. Today, India consumes a few trillion tokens a day and imports the majority of them. Scaling this consumption up and building sovereign capability requires investing in GPU infra, running efficient token factories, and guaranteeing advantages in cost, latency, and data residency. This is the key objective of Sarvam Token Factory. We will continue to reduce costs further and widen access for inferencing in India. 3. Our own models: Building inference capability does not conflict with our core mission: training our own models. Over the last few months, alongside putting together the token factory and scaling up our enterprise offerings, we optimised our training stack for Blackwell GPUs and achieved significantly higher performance. We used this optimised infra to train an improved checkpoint: Sarvam 105B-1711, which we are open-sourcing today. It improves on reasoning, tool-calling, and co-work tasks, while excelling in Indian languages. It is one of the best Flash-Lite class models, open or closed, achieving high accuracy at a low cost of ₹25 per million output tokens. We also optimised it to run on a single RTX Pro GPU with our agentic product Indus. Along with the weights, we release inference and training recipes so others can build upon it. In the coming two weeks we will release multiple other models across vision, speech, and edge. And crucially, we are working with partners to get the latest training grade compute into India. This will set us up for training our next generation of much larger models. 4. The Sarvam Journey. To put things in perspective, most AI labs invest 100s of millions of dollars into infra to train from scratch, scaling up to billions for frontier capabilities. We are determined to climb that ladder from India, while building a sustainable commercial engine across population scale and enterprise use cases. Irrespective of the different narratives, some of which appear manufactured, we firmly believe that India can and must win across the full-stack of AI - we have the talent, the demand, and positivity to adopt AI. At Sarvam we will stay focused and build towards this outcome.
82
253
52
1,454
65,111
2 reels I posted 1st one 200 views 2nd one 1.3k+ views . . . 1st without captions 2nd with captions Interessant🤔
1
27
Sumukh retweeted
Cardboard (desktop) is live on Product Hunt. Go check it out and leave us a review <3 cardboard.ai/ph
27
13
4
228
8,409
ಕನ್ನಡ ರಾಜ್ಯೋತ್ಸವದ ಹಾರ್ದಿಕ ಶುಭಾಶಯಗಳು! ❤️💛
1
33
ಕನ್ನಡ ಒಂದು ಭಾಷೆ ಮಾತ್ರ ಅಲ್ಲ, ಅದೊಂದು ಭಾವನೆ ❤️💛
96
1,711
73
8,014
97,244
Do you remember when you joined X? I do! #MyXAnniversary
41
For eight years, @Zetwerk, you’ve been turning the biggest of ideas into reality. And in that spirit, we thought we’d introduce you to prototypes of the future from your toughest customers yet. After all, most inventions start with a child asking, “What if?” Congratulations on your IPO, Zetwerk. We can’t wait to see what you make next. #AccelFamily • @prayanks • @acharyaamrit • @srinathramki
6
33
12
115
79,362
Claude is best for mobile apps design 😭😭
46
One guy built that pdf tool Now every third person is building the same lol 😭
1
1
56
Excited to submit AgentDoctor (AutoPatch AI) for #neatHack by @neatlogs & @TheVibeFounder! 🚀 With my friend @subhamm_dmg
1
1
87
Video made using @usecardboard ✨✨
22
Excited to submit AgentDoctor (AutoPatch AI) for #neatHack by @neatlogs & @TheVibeFounder! 🚀 With my friend @subhamm_dmg
building agents is easy getting them to production is hard today we’re launching @neatlogs to fix that
1
1
1
3
273
Autonomous bug triage, failure reproduction & self-healing agent platform: 🔍 Observability: @neatlogs comparative pre/post traces & span waterfall 🌿 Git Context: @entire_io AST code graphs & checkpoint lineage 💼 Economics: @cfo_ai unit economics ($0.017/patch, 99.6% margin)
1
2
19
Day 1 building AgentDoctor for #neatHack by @neatlogs & @TheVibeFounder ! 🚀 with my friend @Sumukhmg Tackling the AI reliability crisis: distributed agents & microservices fail silently, costing 2.5 hrs/incident ($230+).
1
1
1
22
Day 1 building AgentDoctor for #neatHack by @neatlogs & @TheVibeFounder! 🚀 with my friend @subhamm_dmg Tackling the AI reliability crisis: distributed agents & microservices fail silently, costing 2.5 hrs/incident ($230+).
1
1
1
66
We built a 5-agent autonomous swarm that repairs code without humans in the loop: 🩺 Triage Agent: Stack trace ingestion & error classification 🌿 Context Agent: @entire_io AST code graphs & commit checkpoints 🩹 Doctor Agent: Surgical patch synthesis & diff generation
1
24
🧪 QA Verifier Agent: Pytest execution & iterative self-healing 💼 CFO Agent: @cfo_ai dynamic unit economics From failure to verified fix in seconds. Demo dropping soon! #AIagents #SoftwareEngineering #DevTools #Neatlogs #Entire
23
So you are saying hindi version of Drishyam and malyalam version is different???????
51
ಪಿತೃ ಪಕ್ಷ 🙏🏻
Caw caw caw caw caw caw caw caw
1
39