2 karma · joined February 16, 2026
Over the past three years I’ve worked on RAG-based LLM systems, semantic search pipelines, and computer vision services for KYC and content moderation. Most of my work revolves around taking models from prototype to reliable production systems — optimizing retrieval, reducing hallucinations, improving latency, and making AI services actually usable in real-world environments.
I enjoy working on system design challenges around context memory, vector search, caching strategies, and scalable inference APIs. Currently exploring more advanced agent orchestration and multi-agent workflows.
Router Branching: Uses LangGraph for conditional flows—classifies as conversational, planning-needed, or direct-execute. Handles tool calls safely with confirmations for risky stuff. Safe Package Mgmt: Installs dependencies locally without internet, avoiding supply-chain risks. REPL-like Execution: Stateful code interp for persistent sessions.
It's built with Python/LangChain, runs on any Linux box. Here's a quick example: Ask it to "analyze this PDF and plot data"—it plans, browses the file, executes via tools, and outputs. Repo: https://github.com/zkzkgamal/zkzk-agent (stars/feedback welcome!) What do you think—any ways to harden the router against edge cases? Or integrations you'd want (e.g., more bio/chem tools)? This framing emphasizes progress from your old posts, adds tech depth, and invites input. If it doesn't take off, analyze the comments and iterate—many viral projects started with flops. Good luck; HN can be unpredictable, but building genuinely useful stuff is the real win! If you share more project details, I can refine this further.
This chapter dives into LangGraph for real agent orchestration: - Build ReAct-style agents with tool calling - Visualize graphs (nodes/edges/conditional routing) using Graphviz – super visual and satisfying - Handle state updates live, multi-agent collaboration patterns, self-refinement loops - Full executed notebook showing a simple agent reasoning through queries (e.g., math tools, search integration)
All local-first with Ollama, plus OpenAI/Gemini support.
Repo: https://github.com/zkzkGamal/Agentic-AI-Tutorial (Chapter4 folder has the notebook + code)
Chapter 5 (Personal Finance Tracker Agent with FastAPI + ChromaDB RAG) is next and almost ready.
Feedback super welcome – especially on the LangGraph patterns or what real-world agents you'd like to see built!
Thanks for any upvotes/comments – still early days :)
Highlights in v2:
Network-aware actions & Wi-Fi auto-reconnect
Image search & download from DuckDuckGo
Streaming token-by-token responses
Quick install script for fast setup
Optional voice input/output (Whisper + TTS)
Human confirmation for dangerous operations
All processing happens locally — nothing goes to the cloud. It can manage files, processes, deployments, and system commands with AI assistance.
GitHub: https://github.com/zkzkGamal/zkzkAgent
Would love feedback, ideas, or discussion on self-hosted AI agents.
I built StudyWithMiku, an open-source AI study assistant that:
Ingests PDFs automatically (drop into a folder)
Embeds them into ChromaDB
Uses RAG for grounded answers
Maintains conversational memory with LangGraph
Outputs synthesized character voice using Coqui TTS + DiffSinger
Runs locally (Ollama supported) or with cloud LLMs
The goal was to combine:
• Retrieval-augmented generation • Stateful agent orchestration • Tool execution (web/system tools) • Persistent memory • Character-specific TTS
into one cohesive system.
Engineering challenges included:
Chunking strategy and retrieval accuracy
Avoiding recursive agent tool loops
Managing CUDA/PyTorch/protobuf conflicts
Keeping latency reasonable with local models + TTS
It’s Linux-first (Ubuntu recommended) and GPU helps, but CPU works.
I’m particularly interested in feedback on:
RAG optimization approaches
LangGraph/agent architecture patterns
TTS expressiveness tuning
Local LLM deployment tradeoffs
Happy to answer technical questions.