Field notes from a research engineer shipping AI systems. Real AI problem, what I tried, what I learned. 18+ yrs · 35+ patents · 17+ papers 📩 Weekly notebook ↓
Joined February 2024
- Tweets8.1K
- Following989
- Followers2.8K
- Likes1.9K
AI EdTalks retweeted
NCM vs ChatGPT: 31x Faster AI Explained youtu.be/DrsLhITurkc?si=gtTJ…
AI EdTalks retweeted
The Only Claude Mods You Need to Install This Year
youtu.be/RoF2c-p8QpQ?si=ZmO6…
AI EdTalks retweeted
Last week, a reader replied:
"This fixed a bug we'd been stuck on for 2 weeks."
That's why I write it.
Practical AI engineering from a research engineer: 18+ yrs building production systems, 35+ patents, 17+ papers.
Subscribe to the Newsletter 👇
AI EdTalks retweeted
Why Your AI Agent Needs a Tool Router (And How to Build It)
Giving your LLM access to every function doesn't make it smarter; it creates a confused AI agent that struggles to execute. When you dump the entire toolset into the prompt, you increase the likelihood of hallucinated parameters and execution errors.
Instead of overwhelming the model, you need a gatekeeper. We look at how to implement tool routing to restrict the available functions based on the user's specific request.
This approach relies on semantic retrieval to dynamically surface only the relevant tools for a given task.
By narrowing the scope, you improve LLM performance and reduce the noise in your agent's decision-making process. This is the difference between a brittle prototype and a system that actually works in production.
Subscribe for weekly backend engineering primers, and tell me if you have started implementing dynamic routing in your production agents.
youtu.be/JD9SV5ZoZDg?si=p1jN… v
AI EdTalks retweeted
Which AI skill should you learn first? Andrew Ng's answer: #aiagents #ai...
youtube.com/shorts/ckjL0wZU2… v
AI EdTalks retweeted
Anthropic Sept 10 2026 report:
How attackers are bypassing API key security to drain your LLM production credits.
The new Anthropic report outlines exactly how prompt injection exploits are being used to leak keys for resale on Telegram.
We break down the technical reality of these AI security vulnerabilities and how to audit your own system traces.
Subscribe for weekly backend engineering primers, and tell me if you have started restricting your model permissions yet.
youtube.com/shorts/zPA6gwVVD…
AI EdTalks retweeted
Stop building point-to-point integrations and learn why the Model Context Protocol is the new standard for AI engineering.
If your agentic AI setup feels like spaghetti code, you are likely suffering from the M-by-N connection problem.
This protocol acts like a USB-C port for your LLMs, simplifying how your applications handle complex tool integration.
youtube.com/shorts/ZQ1BDe5CH…
AI EdTalks retweeted
Model routing sends each LLM request to the cheapest model that can actually handle it. Most AI apps skip this and send everything to one frontier model — including the intent classification, the extraction, and the summarising.
One engineer logged every call his app made and found 70% of them were grunt work running at frontier prices. He put a small classifier in front, routed simple requests to a cheap model, and his bill went from around $800/mo to under $160. (One team's result, not a guarantee — but the shape holds everywhere.)
This video covers the full pattern: how routing works, why it's just a load balancer with opinions, what multi-agent fan-out does to your token spend, and the two traps that make routing backfire — including a cascade setup that can cost more than always calling your best model.
The AI cost trap: Why you’re using the wrong model youtu.be/1FA0LLM8QQw?si=kxtv…
AI EdTalks retweeted
LLM Observability From Scratch — Debug Any AI Agent youtu.be/NH6wxVUlH10?si=kyE9… via @YouTube
AI EdTalks retweeted
The only 2 words you need to debug AI #aiengineering #aiagents #llmobser... youtube.com/shorts/6pkRZU88R… via @YouTube
AI EdTalks retweeted
Your MCP Server Breaks Behind a Load Balancer Here's the Fix youtu.be/djvAoRh9kEM?si=4QOC…
AI EdTalks retweeted
OpenTelemetry Don't get locked into one AI observability tool #aiagent... youtube.com/shorts/51lu9yyN4…
AI EdTalks retweeted
Join the newsletter.
Upgrade yourself every Friday:
One real AI problem,
what I tried,
what I learned.
AI EdTalks retweeted
LLM Observability From Scratch — Debug Any AI Agent youtu.be/NH6wxVUlH10?si=yHdN…
AI EdTalks retweeted
Production observability checklist for LangGraph agents.
If you're running an agent for real users, work down this list. 👇