Over the past week, I’ve learned more about some of the technologies and tools teachers use in their day to day in the US public school system, and how difficult it can be for them to aggregate their data.
I love how easy this is for teachers to use! Anthropic at it again 👀
"It works" is not a metric.
If you can't trace it, eval it, and re-test it automatically, you don't have an AI agent. You have a liability that happens to demo well.
Adults: Let’s ban AI in schools.
Alpha School student: Let's stress test the AI instead.
The future belongs to kids who learn to push AI to its limits - not hide from it.
You can now use Devin on your desktop! You can use a Devin Local has context about the project you're working in, files included.
Downloadable now! https://t.co/btK2QRErBr
This is a must-read.
Education at scale is the biggest problem we need to solve.
Despite my being an AI Engineer, I wholly agree that universities categorizing AI use as legitimate scholarship, as long as it's footnoted, is wrong.
As @yacineMTB said, "You can outsource your thinking, but you cannot outsource your understanding."
Week 7 at @gauntletai
Built ChatBridge — an AI chat platform that runs apps inside the conversation and stays aware of live state.
• tool invocation mid-conversation
• persistent app state across turns
• agent reasoning grounded in real data, not guesses
• sandboxed app execution + validated messaging at the boundary
The challenge wasn’t building features.
It was keeping the system aligned — agent, tools, state, and security — across multiple turns without drifting or breaking.
Grateful for the standard here. Incredible people, real work.
Demo below
https://t.co/kqX0pMbB45
MONDAY at Gauntlet AI (in the heart of downtown Austin):
We’re doing a live demo and walkthrough of @KellyClaudeAI: How she was built, how she operates, and the details required for an AI agent to build companies with zero human involvement.
https://t.co/yQF6xMWbQc
24 hours to build a RAG system for legacy code.
Introducing LegacyLens, my Week 3 project for @GauntletAI.
It helps developers understand COBOL and Fortran through natural language — turning decades-old codebases into something you can actually reason about.
Features:
• Syntax-aware chunking
• Streaming answers with citations
• A UI designed to feel like a modern dev tool
Built with FastAPI, React, pgvector, and Claude 3.5 Sonnet.
Incredibly grateful for everything the Gauntlet team is teaching us.
Amazing what you can build when you're pushed.
Demo:
https://t.co/LpBflbuYhi
#GauntletAI #RAG #AI #LegacyCode #BuildInPublic
Week 2 of @GauntletAI wraps up.
After 14 days of learning from some of the sharpest minds in AI, I'm humbled to share what I've been building.
I built an AI financial assistant on top of @ghostfolio_ , letting you talk to your portfolio in natural language. It understands your holdings, benchmarks your performance, and can execute trades — all from a simple chat interface.
I also extended it with a personal feature: a real estate layer built directly into the platform. For the first time, your property sits alongside your stocks — market value, rental income, and appreciation all in one place.
One honest number: your actual net worth.
Deeply grateful for the push to build something real. The work is just getting started.
#AI #AgentForge #Fintech #OpenSource
Agent demo:
https://t.co/jhYto28Zfn
Real estate feature:
https://t.co/xVEufQ4Ico
The @DCWHockey Wildcats are going to state!
Cats beat the top two seeds — beating No. 7-ranked Northfield 4-2 tonight — to win the Section 1A title for the second time in 6 years!
Week 2 of @GauntletAI and another 24-hour MVP shipped.
This week I built a natural language financial assistant on top of @ghostfolio — ask it anything about your portfolio and it routes to the right tool automatically.
What I used:
— Claude Sonnet 4.6 for the agent brain
— FastAPI on Railway for the backend
— React + Vite on Vercel for the frontend
— Supabase for conversation memory
— A formal eval framework with 15 golden test cases and 27 labeled scenarios
The hardest part wasn't the code. It was building something that actually knows when NOT to answer.
#GauntletAI #BuildInPublic #AI #FinTech
https://t.co/HkiGOpsmVf
Week 1 of @gauntletai is a wrap! I built CollabBoard, a real-time collaborative whiteboard with a voice-controlled AI agent.
Check out the demo!
https://t.co/j9kUrGod6F
This has been the most intense, yet rewarding week of my professional career. Huge thanks to @Austen and the entire Gauntlet team for pushing us to build at the speed of thought.
Tech: @cursor_ai , @anthropic, @langchain, @vercel, @railway
#AI #BuildInPublic #GauntletAI
It's about smart routing, not compression:
- Analyzes task complexity → routes simple tasks to Sonnet (80% cheaper)
- Estimates tokens BEFORE you send → no surprises
- Validates output quality → catches issues before you waste tokens regenerating
The "burn" happens when you use Opus for simple tasks or regenerate bad code. This prevents both.
During this week’s @gauntletai sprint, I noticed a common pattern: Builders were burning through Claude Code tokens faster than expected
To help reduce that friction, I built a small open-source Claude Code token optimizer skill
Repo here 👇
https://t.co/JXJObLJpHJ
#Savetoks