Vishal Builds

Shipping AI Products in Public 🤖 @vishalsingh2972

Currently cooking Voice AI & LLMs

DM for Work/Freelance

☕ Feed my Cursor Pro :

[Vishal Builds @debugginglife25]

Amma sends a photo of her medical report on WhatsApp 📝

✅ Instant Telugu/Hindi voice explanation for parent
✅ Clear English health summary for NRI child abroad
✅ Critical cases automatic email alerts to doctor
✅ Built by @SarvamAI Bulbul TTS via @SarvamForDevs

Demo Video 👇

[Vishal Builds @debugginglife25]

@SarvamForDevs medical app 2

🩺 Our in-house Telugu doctor - turning complex medical reports into easy-to-understand local Telugu, explained with warmth and everyday slang.

Built using Bulbul V3 model by @SarvamAI

Video demo coming soon 🥳
https://t.co/AAO1W30yVy

[Vishal Builds @debugginglife25]

@Dimillian made this tweet using voice, working so well

https://t.co/ArUt8i9Jff

[Vishal Singh 🥑 @vishalsingh2972]

Planning to start System Design 📚

Learn topics in this exact order 👇:

• Load Balancing
• Caching
• CDN
• Database Indexing
• CAP Theorem
• Replication & Sharding
• Consistent Hashing
• Message Queues
• Rate Limiting
• Microservices

[Vishal Builds @debugginglife25]

@gdb made this tweet using voice, working beautifully

https://t.co/ArUt8i9Jff

[Vishal Singh 🥑 @vishalsingh2972]

Planning to start System Design 📚

Learn topics in this exact order 👇:

• Load Balancing
• Caching
• CDN
• Database Indexing
• CAP Theorem
• Replication & Sharding
• Consistent Hashing
• Message Queues
• Rate Limiting
• Microservices

[Vishal Builds @debugginglife25]

made this tweet using chatgpt Voice, AGI is here ✨🫡

[Vishal Singh 🥑 @vishalsingh2972]

Planning to start System Design 📚

Learn topics in this exact order 👇:

• Load Balancing
• Caching
• CDN
• Database Indexing
• CAP Theorem
• Replication & Sharding
• Consistent Hashing
• Message Queues
• Rate Limiting
• Microservices

[Vishal Singh 🥑 @vishalsingh2972]

Watching how LLMs think is way more fun than just reading its output.

So, I built an interactive 3D visual debugger for transformer AI models.

Tested with the latest @Alibaba_Qwen 3.8 model.

• Live token generation
• Tensor inspection
• Architecture explorer
• Real inference visualization

Open Sourcing soon ❤️

Demo video 👇

[Vishal Singh 🥑 @vishalsingh2972]

Weekend Project: Stop Token Waste with Qwen 3.8 🔥

Meet Jugaad AI, a multi-agent coding setup where the latest @Alibaba_Qwen 3.8 models do the grunt work while the lead model reasons.

Instead of burning a premium model's context reading entire codebase, the lead assigns bounded tasks to local Qwen 3.8 agents working in parallel. Every code citation is mechanically verified before it's trusted.

⚡ Premium model focuses on reasoning
🏠 Sensitive code stays local
💰 Free local models do the heavy lifting
✅ Hallucinated citations rejected

Tech Stack:
• Local Qwen 3.8 (OpenAI compatible)
• GPT-5.6 / Claude Fable (Lead Orchestrator)
• Python
• tmux

My approach:
Let the expensive model think, let local models read, and never trust a citation until it's been verified. Qwen 3.8 worked beautifully for delegated code-reading tasks.

Demo + GitHub below 👇

[Vishal Builds @debugginglife25]

@DynamicWebPaige Building tool for Hackathon using Gemini. we are hitting free-tier rate limits during multi-agent testing.

Any chance we could get a temporary credit boost/rate limit bump to keep things running smoothly for the judges this weekend?

I hope you see this 👋

[Vishal Singh 🥑 @vishalsingh2972]

spent 3 hours going down the rabbit hole of Grok's open-source grok-build repo. here's what i found 👇

honestly, i expected a simple LLM CLI.

instead i found this huge Rust codebase powering a full-blown multi-agent platform.

i'm a junior dev, and i had AI open the whole time helping me understand what i was reading. half the time i was reading code, half the time i was asking "what does this even mean?" 😅

i definitely didn't understand everything, not even close, but the parts i did understand were so cool that i wanted to write them down so i don't forget.

this is a mix of actual takeaways and a bunch of "i need to GPT this in detail later."

✅ what i think i understood:

- it isn't just one AI session. there's a parent agent that spawns helper agents, each with its own prompt, status, and lifecycle. seeing one agent spin up other agents inside a terminal was wild.
- instead of relying on grep or regex, it uses Tree-sitter to parse actual syntax trees for languages like Rust, TypeScript, Python, and Go. that was a huge lightbulb moment for me, it understands code structure, not just text.
- prompts aren't just strings. they use cached templates, different prompts for different agent roles, and automatically inject project context like AGENTS.md.
- the terminal UI has a dedicated rendering thread, so even if drawing gets slow, the main event loop stays responsive. i've accidentally frozen my own CLI tools before, so seeing this made me go "oh... that's how production apps do it."

🧠 stuff that made me feel super dumb:

- token estimation uses a simple bytes ÷ 4 heuristic with integer math instead of floating point, and it triggers context compaction at exactly 85%. AI explained why integer math is simpler and deterministic for this kind of heuristic, but i still want to understand when that actually matters in production.
- the indexing system builds a code graph using Tree-sitter and runs with actors, weak references, lazy maps, and event coalescing. i understand maybe 20% of that sentence. i've now got a bunch of tabs open about Rust actor patterns and `Weak<T>`.
- apparently repeatedly redrawing a terminal can reset the cursor blink timer. Grok tracks what actually changed on screen and only emits the minimum terminal updates needed, so the cursor keeps blinking naturally. i didn't even know this was a problem until today.
- i also ran into things like separate crates to avoid circular dependencies and terminal back-pressure, concepts i barely knew existed.

i probably walked away understanding 15–20% of the repo.

big thanks to @xAI for making this public. i definitely didn't understand most of it, but i still learned a lot just from exploring the repo.

but that 20% made me realize how much thought goes into the small details. i kept finding code that solved problems i didn't even know were a thing.

if you're also early in your journey, don't be afraid to read codebases that are way above your level. you probably won't understand most of it, but every time something clicks, you come away having learned something new.

to the Rust and systems folks: what was your favorite part of this architecture? and if you know any beginner-friendly resources on actor patterns, `Weak<T>`, or why integer math is preferred here, i'd love the recommendations.

for now... i've got 17 tabs open and a lot of reading to do. rn i feel i know nothing. 🙃

[Vishal Singh 🥑 @vishalsingh2972]

🇮🇳 has 30 million people with visual impairment.

The hardware gap is largely solved, millions carry a smartphone with internet access.

So, I built an AI agent that runs entirely in the mobile browser.

📷 Uses your phone camera
🚶 Detects obstacles & estimates distance
🗣️ Real-time voice guidance powered by @SarvamAI

Open sourcing it soon ❤️

Demo Video in Kannada 👇

[Vishal Singh 🥑 @vishalsingh2972]

Weekend AI Project 🔥

Meet Research Saathi - a multi-agent AI research team that actually works.

Instead of fighting one massive prompt, I orchestrated 4 specialized agents using LangGraph:

🔍 Search: Finds credible, recent sources.
📄 Reader: Extracts signal from the noise.
✍️ Writer: Synthesizes a structured report.
🧐 Critic: Reviews and kills hallucinations.

The Workflow:
The agents pass state back and forth in a loop, mimicking a real research firm. The quality jump compared to a single-LLM approach is night and day.

Tech Stack:
• LangGraph (Orchestration)
• LangChain (Logic)
• Streamlit (Interface)

Key Takeaway:
Agentic workflows > Clever prompting. Giving models clear roles and a feedback loop is the ultimate unlock for reliability.

Demo video + GitHub in the comments. 👇

[Vishal Singh 🥑 @vishalsingh2972]

Built a real-time Voice Agent in 30 minutes using Sarvam's new MCP server.

My mission? Solve the most important problem in Bengaluru: finding a good Masala Dosa 🥘

I wanted to test Sarvam’s localized models in a live environment. I don't speak Kannada, so I built and tested it in Telugu. It was wild how smoothly it handled language switching, real-time tool calling, and multi-turn conversations without noticeable lag.

Tech stack 🔥
• @SarvamAI (saaras + bulbul)
• LiveKit Agents
• Cline

Built in Telugu, but since Sarvam supports 22+ Indian languages, feel free to fork the repo and try it in your own language!

GitHub repo link in the comments.

Demo video 👇

[Vishal Builds @debugginglife25]

@HarveenChadha Awesome opportunity ✨

[Vishal Builds @debugginglife25]

@AlexFinn EXACTLY!!!

https://t.co/PiJdFY0mcr

[Vishal Builds @debugginglife25]

@TheCoderShow @cerebras Hyderabad DMed you!

[Vishal Builds @debugginglife25]

@bhavyaaa_here The UI is so beautiful 😍❤️

[Vishal Builds @debugginglife25]

@Sambhav_Gandhi Hey you're DMs are closed, could you please ping me so I could reply?

[Sarvam @SarvamAI]

Kisan Samman, a voice-first AI agent that helps farmers discover government schemes, check eligibility, and get application guidance in their own language in under 30 seconds, built using Sarvam's Speech to Text and Text to Speech models.

Built by: @vishalsingh2972

https://t.co/SK3lfZhPtX

[Vishal Singh 🥑 @vishalsingh2972]

Building AI for 🇮🇳:

Build an AI Agent that helps farmers find, apply to government schemes in under 30 seconds.

Indian government spends ₹1.8 lakh crore+ every year on farmer welfare.

Yet millions of eligible farmers still miss out on benefits.

Not because schemes don't exist.

Not because funds aren't allocated.

But because finding the right scheme is often harder than farming itself.

Farmers face:
• Lack of awareness about available schemes
• Complex government portals
• Language barriers
• Confusing eligibility requirements
• No clear guidance on how to apply

So, over the weekend, I built Kisan Samman, a Voice-First AI Agent that helps farmers discover and access government schemes in seconds.

A farmer simply speaks in their own language:
🎙️ "क्या मेरे राज्य में ड्रिप सिंचाई के लिए कोई सब्सिडी है?"

Kisan Samman automatically:
✅ Finds relevant government schemes
✅ Checks eligibility using the farmer's profile
✅ Retrieves the official application portal
✅ Locates the nearest Common Service Centre
✅ Explains everything back in the farmer's language using voice
✅ question → eligibility → application all guidance in under 30 seconds.

🧠 Tech Stack:
• Grok+ Llama 3 70B
• LangChain
• Sarvam AI (ASR + TTS)
• Flask + Python
• Leaflet.js + OpenStreetMap

Goal of this project:
Make government welfare schemes as accessible as a conversation.

Demo video below 👇

[Vishal Builds @debugginglife25]

Amma sends a photo of her medical report on WhatsApp 📝

✅ Instant Telugu/Hindi voice explanation for parent
✅ Clear English health summary for NRI child abroad
✅ Critical cases automatic email alerts to doctor
✅ Built by @SarvamAI Bulbul TTS via @SarvamForDevs

Demo Video 👇

[Vishal Singh 🥑 @vishalsingh2972]

🇮🇳 "Amma, mee report vachindi. Nenu chepthanu vinandi..."

This is Dear Comrade — a health bridge for every NRI child who worries about their parents back home.

How it works:
📸 Parent sends lab report photo on WhatsApp
🧠 Sarvam 105B understands medical data
🎙️ Bulbul TTS speaks warm Telugu/Hindi voice note to parent
📱 English summary + critical alerts to child abroad
👨‍⚕️ Automatic email to family doctor if values are critical
🖥️ Real-time dashboard with charts, AI summaries, PDF export, email service.

Built with the amazing @SarvamAI models (105B + Bulbul TTS + Mayura), Next.js, Supabase, and Twilio.

The goal?
No parent should feel lost reading medical jargon.
No child should feel helpless 8,000 miles away.
Use AI to build for India. @SarvamForDevs

GitHub repo in comments.

Demo video below 👇

[Vishal Builds @debugginglife25]

@curlysaarthak Any specific examples particular job roles...