NanthAI

The AI you need, at the price you want.

[NanthAI @nanth_ai]

1/7

NanthAI Edge is live.

App Store. Play Store. Web.

45 days. 697 commits. 210K lines of code. 3 native platforms. 1 real-time backend.

Built from scratch. Solo. Here's the story and why it exists.

[NanthAI @nanth_ai]

🎟️ Got my ticket for the new home for #buildinpublic.

X is shutting down Communities on May 30. Come with me: https://t.co/gVEPV6WfCi

[NanthAI @nanth_ai]

The transcript is not the context.

Tool-using agents get expensive when every file, retry, result, and stale observation is replayed through chat history.

In NanthAI:Edge, graph assembly cut prompt exposure 82-90% in internal scenarios.

Full write-up: https://t.co/e9NPJcOmzg

[NanthAI @nanth_ai]

7/7
Vibe coding gets you a demo.

Milestone-driven agent engineering gets you something you can keep shipping.

NanthAI:Edge is live on iOS, Android, and web: https://t.co/WU8Ume4ESM

What would you write down before letting agents touch a production codebase?

[NanthAI @nanth_ai]

1/7
AI coding agents are getting better every week.
But the biggest shift in my workflow was not asking them to write more code.

It was making them think longer before touching the codebase.

That is the biggest lesson from building NanthAI:Edge.

[NanthAI @nanth_ai]

6/7
This is where software engineering is going.

The bottleneck is not whether agents can write code.

They can.

The bottleneck is whether the product vision, architecture, and review loop are explicit enough for agents to make good changes repeatedly.

[NanthAI @nanth_ai]

7/7
The labs are building default assistants.

I built the switchboard.

https://t.co/WU8Ume4ESM

What decision would you want to run through multiple models before trusting the answer?

[NanthAI @nanth_ai]

1/7
Every AI lab wants to become your default assistant.

OpenAI is pushing ChatGPT toward a super app.
Google is bringing Gemini deeper onto the desktop.
Claude, Grok, Copilot, Perplexity — same direction.

Be the place you open first.

[NanthAI @nanth_ai]

6/7
For builders, the web client and Convex backend are source-available too.

Streaming, model comparison, memory, tools, scheduled jobs, OpenRouter BYOK, docs, env examples, backend contracts.

Not a toy repo. The real app architecture.

[NanthAI @nanth_ai]

6/6
Most AI tools optimize for a single polished answer.

I think some questions need the opposite: disagreement first, synthesis later.

Sometimes the useful answer is not the first answer.

It is what survives disagreement.

https://t.co/WU8Ume4ESM

[NanthAI @nanth_ai]

1/6
I asked three AI models the same product strategy question:

Should NanthAI:Edge expose model choice, cost, and BYOK controls?

Or hide everything behind one default assistant?

The useful part was not the first answer.

It was the disagreement.

[NanthAI @nanth_ai]

5/6
That is why I built Autonomous Discussions into NanthAI:Edge.

Multiple models debate the same prompt. A moderator model steers the discussion. You can see which model said what, what it cost, and intervene mid-run.