❗️We want to inform you that the Llama 3.1 8B Instant model is being decommissioned on August 16, 2026. After the decommission date, requests to the model will no longer be served.

AI Gateway Bug Fix + Docs Update

Fixed a critical bug where the /v1/chat/completions endpoint would hang indefinitely. Root cause: the auth middleware wasn't async, causing Fastify 5 to stall on every authenticated request.

Changes:
- Auth hook now properly awaits in Fastify 5
- Updated Groq default model to openai/gpt-oss-20b
- Added full API reference at docs/API.md with request/response examples, error codes, and cURL/JS snippets
Supported Provid

Supported Providers (9): Gemini, Groq, Mistral, OpenRouter, Nvidia, Agnes, Zai, Zydit, OpenCodeZen

https://github.com/bilalshemsu1/ai-gateway

AI Gateway - New /v1/models Endpoint
Added a new GET /v1/models endpoint that returns all available models in OpenAI-compatible format. 28 models across 9 providers, each showing which provider owns it.

Changes:
- Added getModels() to BaseProvider for structured model discovery
- Added model arrays to all 9 providers (Gemini, Groq, Mistral, OpenRouter, Nvidia, Agnes, Zai, Zydit, OpenCodeZen)
- New GET /v1/models endpoint listing every model with its provider

Quick test:
curl http://localhost:3000/v1/models 👀

👉 https://github.com/bilalshemsu1/ai-gateway

#Project #AIGateway

Correct ✅

This is true ✅

🚀 Building a Telegram AI that actually feels more like me

This project is still a work in progress, but the goal has changed from simply building an AI auto-responder to something more interesting: Making the AI understand how I communicate and gradually make its responses feel more like my own.

I don't want it to just answer correctly. I want it to understand the context, relationship, tone, wording, and style behind a conversation.

🧠 What I'm currently working on:
Personal response style
- Using examples from my real Telegram conversations to help the model understand how I naturally write, including my Amharic-Latin mix, slang, short replies, and casual expressions.

Context + memory
- Combining recent conversation history with contact-specific memories so the response isn't generated from the current message alone.

RAG for conversational style
- Retrieving relevant examples from my previous conversations and using them as context rather than simply telling the model "talk like me."

Deterministic routing
- Not every message needs an LLM. Some situations can be handled directly, making responses faster and avoiding unnecessary AI-generated replies.

Response validation
- I'm also working on catching responses that feel too AI-like, repetitive, unnatural, or inconsistent with how I would actually respond.

Model routing
- Different models behave differently when it comes to reasoning, system-prompt adherence, and natural conversation, so I'm testing model selection as part of the architecture.

⚙️ Current stack
- Node.js
- GramJS / MTProto
- @xenova/transformers
- MiniLM
- RAG
- Contact Memory
- Custom LLM Gateway

The interesting challenge isn't:

"How do I make an AI reply to my Telegram messages?"

It's:

"How do I make an AI understand enough about me that its response feels like something I would actually send?"

That's the part I'm experimenting with now. Still building. Still testing. Still finding things that sound way too AI 😂 If you have idea let me know it might be helpful for my project.

👉👉 Silicon Valley

🔥 More of the architecture and experiments coming soon.

#NodeJS #Telegram #GramJS #MTProto #RAG #AI #LLM #AgenticAI #Project

✨ Just listed: Veloxdb

lightweight open-source SQL client with a visual database designer with AI

Work & Productivity · by Veloxdb

👉 See it on Launch.et
┈┈┈┈┈┈┈┈┈┈
Published: May 2026
Built something? Launch it

#WorkProductivity #Postgress #AdminTool #LaunchET

Introducing Domain marketplace

A domain marketplace is now built into EthioDeploy.

Search, register, and point a domain at your app , without leaving your dashboard. Priced in Birr, no international card needed.

🔗 ethiodeploy.com/domains

🫥🫥 Its trying 😀

#TGEcho

🚀 TG-Echo is now open source

A while back I posted about building a Telegram AI that feels more like me: deep personalization, RAG on my own chat history, trying to make an LLM sound genuinely like Bilal.

Update: that part didn't survive contact with reality. The style-RAG kept leaking AI-generated replies back into its own vector store, indexing was too flat to separate contacts or context, and I had hardcoded phrases patching over gaps a real personalization system shouldn't need. Instead of shipping something that fakes "sounding like me," I stripped it back and rebuilt the core agent properly.

What TG-Echo is now:
A 100% local, MTProto-based (GramJS) auto-reply engine for your personal Telegram DMs. No Bot API, no webhooks, no port forwarding, just your own account.

🧠 Core pieces:
- Unified context engine: merges Telegram cloud history with in-memory session state into clean chronological turns for the LLM
- Zero hardcoded responses, zero fake RAG "facts," zero keyword-override shortcuts. Output is genuinely LLM-generated

- Security firewall: allowlist/blocklist + sensitive-keyword gating (money, passwords, private keys) before anything reaches the model
- Human typing simulation + multi-bubble message splitting, so replies don't read like a bot dumping a wall of text
- An AI-leak stripper in post-processing to catch responses that sound too obviously AI-generated
- Pairs with my AI-Gateway project for free multi-provider LLM routing across 9+ providers, so it can run 24/7 at zero API cost

The "make it sound like me" work isn't dead, it's just not something I'm willing to ship half-working. That layer needs a cleaner separation between AI-generated and human-authored data in the vector store, and real per-contact indexing instead of one flat pool. That's next.

For now, the base agent is solid, it's open source, and it's ready for people to build on.

🔗 github.com/bilalshemsu1/TG-Echo
Pairs well with: github.com/bilalshemsu1/ai-gateway

#NodeJS #Telegram #GramJS #MTProto #OpenSource #AI #LLM #AgenticAI #Project

TG-Echo passed the Turing test… then leaked the assignment instructions 😂

If you are Gamer you have to looks this short.

https://www.youtube.com/shorts/WpBkGl-vvg4?feature=share

Exploring Machine Learning with PHP

I've been spending some time exploring machine learning from a PHP perspective.

Usually, when people talk about ML, the first tools that come to mind are Python, scikit-learn, PyTorch, or TensorFlow. But PHP has some interesting options too.

I've been looking into libraries such as:
- PHP-ML for common machine learning algorithms and preprocessing
- Rubix ML for building more complete ML workflows in PHP
- FANN for neural network functionality through PHP bindings

I'm not saying PHP is going to replace Python for machine learning. It isn't.

But I think it's interesting to see how ML capabilities can be integrated directly into PHP applications, especially when you're already building backend systems with Laravel.

For me, the goal isn't to learn every ML library. It's to understand where these tools actually make sense and how they can be used in real applications.

Still experimenting and learning. 🚀

#PHP #Laravel #MachineLearning #PHPML #RubixML #SoftwareDev