Soma

An engineer’s last stand: Building agentic systems before it all turns to dust. Follow me!!

[Soma @SomanForever]

@Anjaliydv1706 yes

[Soma @SomanForever]

Excited that our AI voice ordering project for restaurants has been accepted into the Deepgram Startup Program.

Looking forward to building with Deepgram models to improve speech recognition for real-world ordering scenarios.

#Deepgram Deepgram

[Soma @SomanForever]

@SkyTabPOS Thanks. I just DM’ed you.

[Soma @SomanForever]

@SkyTabPOS How long will take to review integration request?

[Soma @SomanForever]

@elonmusk What’s your take on this?
As job security declines, even startup paths feel uncertain if platforms like Anthropic are also moving into vertical solutions.

https://t.co/XsvZrwdWra

[Soma @SomanForever]

Anthropic is rugging the entire ‘AI wrapper economy. 🧵
The "Platform vs. Builder" contract is officially broken.

Here’s why your favorite AI lab is now your biggest competitor:
1/ We paid $200/mo subscriptions and thousands in API credits to "beta test" the market for them. We come up the use cases in BFSI or Health. Now, they’re just building vertical solutions and "Sherlocking" their own devs.

2/ Anthropic's "Claude for Customer Support, Education, Financial Services, Government, Healthcare, Life Sciences, and Nonprofits" isn't an infrastructure update - it's a hostile takeover of the application layer. Did they use our prompt data as a cheat sheet to see what actually works, then built it "natively."?

3/ Why build a vertical startup if the model provider can just flip a switch and launch a competing product with a 100x distribution advantage? The ‘infrastructure-only’ positioning didn’t hold.

4/ If you’re building "AI for [X Industry]" on top of Claude you aren't a founder, you’re an unpaid R&D department for a $100B giant. How can a small builder compete when the platform owns the "brain," the data, and the distribution? While you’re dreaming of building on their platform, their roadmap may already be moving into your space.

6/ They’ve shifted from being the "horizontal foundation" to being the "vertical monopoly." If they keep moving up the stack, there is nothing left for us to build.

Curious to hear your perspective on this,@DarioAmodei how do you see the balance between platform and builders evolving?

What do you think? Is it "survival of the fittest" or a "monopoly move"? 👇 #AI #Anthropic #BuildInPublic #SaaS #Vertical

[Soma @SomanForever]

@andruyeung That only works if the platform doesn’t build what you’re building.

Increasingly, they do.
https://t.co/yTbiJwbETA

[Soma @SomanForever]

Anthropic is rugging the entire ‘AI wrapper economy. 🧵
The "Platform vs. Builder" contract is officially broken.

Here’s why your favorite AI lab is now your biggest competitor:
1/ We paid $200/mo subscriptions and thousands in API credits to "beta test" the market for them. We come up the use cases in BFSI or Health. Now, they’re just building vertical solutions and "Sherlocking" their own devs.

2/ Anthropic's "Claude for Customer Support, Education, Financial Services, Government, Healthcare, Life Sciences, and Nonprofits" isn't an infrastructure update - it's a hostile takeover of the application layer. Did they use our prompt data as a cheat sheet to see what actually works, then built it "natively."?

3/ Why build a vertical startup if the model provider can just flip a switch and launch a competing product with a 100x distribution advantage? The ‘infrastructure-only’ positioning didn’t hold.

4/ If you’re building "AI for [X Industry]" on top of Claude you aren't a founder, you’re an unpaid R&D department for a $100B giant. How can a small builder compete when the platform owns the "brain," the data, and the distribution? While you’re dreaming of building on their platform, their roadmap may already be moving into your space.

6/ They’ve shifted from being the "horizontal foundation" to being the "vertical monopoly." If they keep moving up the stack, there is nothing left for us to build.

Curious to hear your perspective on this,@DarioAmodei how do you see the balance between platform and builders evolving?

What do you think? Is it "survival of the fittest" or a "monopoly move"? 👇 #AI #Anthropic #BuildInPublic #SaaS #Vertical

[Soma @SomanForever]

@lydiahallie Great to see support for open source. This will help a lot of builders.

As platforms expand further up the stack into vertical use cases, it does raise an important question around how developers should think about long-term differentiation.

[Soma @SomanForever]

@itsMaheshVaikri @AnthropicAI Exactly..

[Soma @SomanForever]

Anthropic is rugging the entire ‘AI wrapper economy. 🧵
The "Platform vs. Builder" contract is officially broken.

Here’s why your favorite AI lab is now your biggest competitor:
1/ We paid $200/mo subscriptions and thousands in API credits to "beta test" the market for them. We come up the use cases in BFSI or Health. Now, they’re just building vertical solutions and "Sherlocking" their own devs.

2/ Anthropic's "Claude for Customer Support, Education, Financial Services, Government, Healthcare, Life Sciences, and Nonprofits" isn't an infrastructure update - it's a hostile takeover of the application layer. Did they use our prompt data as a cheat sheet to see what actually works, then built it "natively."?

3/ Why build a vertical startup if the model provider can just flip a switch and launch a competing product with a 100x distribution advantage? The ‘infrastructure-only’ positioning didn’t hold.

4/ If you’re building "AI for [X Industry]" on top of Claude you aren't a founder, you’re an unpaid R&D department for a $100B giant. How can a small builder compete when the platform owns the "brain," the data, and the distribution? While you’re dreaming of building on their platform, their roadmap may already be moving into your space.

6/ They’ve shifted from being the "horizontal foundation" to being the "vertical monopoly." If they keep moving up the stack, there is nothing left for us to build.

Curious to hear your perspective on this,@DarioAmodei how do you see the balance between platform and builders evolving?

What do you think? Is it "survival of the fittest" or a "monopoly move"? 👇 #AI #Anthropic #BuildInPublic #SaaS #Vertical

[Soma @SomanForever]

@tomerlondon @GustoHQ sent

[Soma @SomanForever]

@pmitu or you should ask How does LinkedIn use you?

[Soma @SomanForever]

𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗩𝗼𝗶𝗰𝗲 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀? 𝗛𝗲𝗿𝗲’𝘀 𝗪𝗵𝗮𝘁 𝗔𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗠𝗮𝘁𝘁𝗲𝗿𝘀.

After building real-time voice agent systems with Deepgram, ElevenLabs, OpenAI, and Telnyx, here are the lessons that made the biggest difference:

𝟭. 𝗖𝗮𝘀𝗰𝗮𝗱𝗲, 𝗱𝗼𝗻'𝘁 𝗯𝘂𝗻𝗱𝗹𝗲.
Most providers offer all-in-one voice agents (STT + LLM + TTS bundled together). This is convenient, but you're locked in. Design your pipeline so each component is swappable. Today you might use Deepgram Nova-3 for STT and ElevenLabs for TTS. Tomorrow a faster option may appear, and you can swap one piece instead of rebuilding everything.

𝟮. 𝗟𝗮𝘁𝗲𝗻𝗰𝘆 𝗶𝘀 𝘁𝗵𝗲 𝗼𝗻𝗹𝘆 𝗺𝗲𝘁𝗿𝗶𝗰 𝘆𝗼𝘂𝗿 𝘂𝘀𝗲𝗿𝘀 𝗳𝗲𝗲𝗹.
Nobody cares about your architecture diagram. They care that the AI responds before the silence gets awkward. Cache aggressively. Stream everything. Measure time-to-first-byte on every component.

𝟯. 𝗧𝗵𝗲 𝗲𝘅𝗽𝗲𝗻𝘀𝗶𝘃𝗲 𝗽𝗮𝗿𝘁 𝗶𝘀𝗻'𝘁 𝘄𝗵𝗮𝘁 𝘆𝗼𝘂 𝘁𝗵𝗶𝗻𝗸.
LLM tokens get most of the attention, but real cost comes from idle connection time and repeated API calls. Add caching at the retrieval layer. Use smaller models where possible. Not every turn needs GPT-4o when a lighter model can handle it.

𝟰. 𝗧𝘂𝗿𝗻 𝗱𝗲𝘁𝗲𝗰𝘁𝗶𝗼𝗻 𝗺𝗮𝗸𝗲𝘀 𝗼𝗿 𝗯𝗿𝗲𝗮𝗸𝘀 𝘁𝗵𝗲 𝗲𝘅𝗽𝗲𝗿𝗶𝗲𝗻𝗰𝗲.
Knowing when the user is done speaking versus just pausing is harder than it sounds. Get endpointing and barge-in handling right early. No one wants to talk over an AI that will not stop talking.

𝟱. 𝗗𝗲𝘀𝗶𝗴𝗻 𝗳𝗼𝗿 𝗺𝘂𝗹𝘁𝗶𝗹𝗶𝗻𝗴𝘂𝗮𝗹 𝗳𝗿𝗼𝗺 𝗱𝗮𝘆 𝗼𝗻𝗲.
Adding language support later means redesigning your prompts, TTS voice selection, and STT model settings. Treat language as a core parameter across every pipeline component. It should not be an afterthought.

𝟲. 𝗕𝘂𝗶𝗹𝗱 𝗳𝗼𝗿 𝗽𝗿𝗼𝘃𝗶𝗱𝗲𝗿 𝗳𝗮𝗶𝗹𝘂𝗿𝗲 𝗼𝗻 𝗱𝗮𝘆 𝗼𝗻𝗲.
Voice APIs go down. WebSocket connections drop mid-call. Design adapter patterns so your system handles failures smoothly. Whether you're routing through Twilio or Telnyx, the caller should not notice anything changed.

𝟳. 𝗞𝗲𝗲𝗽 𝘁𝗵𝗲 𝗔𝗜 𝗰𝗼𝗻𝗰𝗶𝘀𝗲.
The biggest complaint with voice agents isn't wrong answers. It is long answers. Users will tolerate a good answer in 3 seconds over a perfect answer in 10 seconds. In voice, silence kills the experience.

What's been your biggest challenge building voice AI systems?

#VoiceAI #AI #Deepgram #ElevenLabs #OpenAI #Telnyx #Twilio #SpeechToText #TextToSpeech #ConversationalAI #AIEngineering

[Soma @SomanForever]

@Anjaliydv1706 ha..ha.. I was reacting to the sudden surge post OpenClaw. Meaningful results are rarely in the moment. Panic tends to produce more volume but not value.

[Soma @SomanForever]

We’re in the “panic-build AI agentic systems” phase of the cycle.
Nobody knows the use case.
Everybody knows the deadline is imaginary.
Still shipping. Before it all turns to dust.

[Mahesh Vaikri @itsMaheshVaikri]

🚀 Unlock the Future of AI with SnapLLM! 🌟

Are you tired of waiting forever to switch between large language models? Say goodbye to those painful delays! 🙌
SnapLLM is here to revolutionize LLM inference with sub-millisecond model switching. Imagine loading multiple models once and switching between them in the blink of an eye—literally under 1 millisecond! 💫

Why SnapLLM?

Blazing Fast Switching: Traditional methods take seconds to minutes. SnapLLM does it in 0.02ms. That's lightning speed! ⚡

Multi-Model Mastery: Manage and switch between various models effortlessly. From Llama 1/2/3 to Qwen, we've got you covered. 🧩

GPU/CPU Hybrid: Automatically distributes layers based on available VRAM for optimal performance. 💻

OpenAI/Anthropic Compatible: A drop-in replacement for the OpenAI API, making integration a breeze. 🔄

Ready to Experience the Future?

1️⃣ Clone the Repo: Get started with git clone https://t.co/DjtI0iZXX7

2️⃣ Build & Run: Follow the simple steps in the README to build and start the server.

3️⃣ Explore: Dive into the bundled web UI or use the CLI/API for seamless model management.

Join the Revolution!

🌟 Star the Repository: Help others discover SnapLLM and join the movement towards faster, more efficient AI.

💬 Engage with the Community: Contribute, discuss, and shape the future of AI together.

Don't miss out on this groundbreaking technology! SnapLLM is the future of Local LLM inference, and it's ready for you today. Let's make AI faster, together! 🚀

[Soma @SomanForever]

#SarvamAI is getting attention, but would you choose it for your use case?

Choosing the right LLM isn’t about hype. It’s about fit.

Start with:
- What is the task?
- What are the constraints (cost, latency, scale)?

Then evaluate:
- Capability: can it handle the complexity?
- Modality: do you need text-only or multi-modal (image, audio, video)?
- Context window: can it fit your input?
- Latency: is it fast enough?
- Cost: will it scale?
- Output control: can it follow structure reliably?

Guidelines:
- Smaller models → simple tasks
- Stronger models → complex reasoning
- Multi-modal models → when inputs are beyond text
- Use RAG for external knowledge
- Keep temperature low for structured outputs

Start with the use case, choose the model that fits, then tune parameters.

#Sarvam #LLM

[Soma @SomanForever]

@Yadnyawalkya00 @sumanthraman STT and TTS are just lossy compression of texts.. LLM cannot reason your emotions in this approach. But native audio can capture your emotions angry speech, enthusiasm or sarcasm etc.. that’s the difference.. Based on your mood it will change its tone…

[Soma @SomanForever]

There are two things @sumanthraman.

One is native audio and another one is Text-to-speech(TTS) and Speech-to-text(STT). TTS and STT are very old technologies.
Now AI models can understand native audio(assume you speak as input) and process it and generate TTS (like output). #Gemini can already do this. But not excelling in that like #Sarvam do. Because #Sarvam has fine-tuned and optimized specific models to handle more diversity of accents. But Gemini is very general purpose. It is not optimzed to the same level of understanding all variants of accents in Telugu, Tamil(like nellai baashai, chennai baashai or nagai or kongu baashai) or other Indian languages upto the level of Sarvam(Atleast that's what Sarvam is claiming for).

Also there is something called Benchmark testing for AI Models. If you benchmark Sarvam and Gemini handling more variants of regional languages, Sarvam claims it has outperformed Gemini. Gemini support Indian lanaguages but it is a global model lagging in the specifics where Sarvam has detailed focus.

Hope this helps!!

[Soma @SomanForever]

@kharis_micheal Bartender. AI cannot get certified. Same goes for any licensed jobs.