Projects — The Muhammadjon Journal

Active

miuceo.github.io

Problem: Hand-writing HTML for every post across three languages on a static site was slow and error-prone.
Solution: A post written to the Telegram channel is auto-translated and synced to the site via a Cloudflare Worker and GitHub Actions; Astro generates the trilingual pages.
Result: From writing a post to it going live on the site takes minutes, with no hand-written HTML.

Astro · TypeScript · Cloudflare Workers · GitHub Actions · Telegram API

View on GitHub → 0

Active

Post Builder

Problem: Writing a post, translating it into three languages and publishing it required separate tools and many manual steps.
Solution: A block editor for text, images and video, wired to a Cloudflare Worker API that folds translation and publishing into one flow.
Result: A post can be written, reviewed and published directly from one interface — no hand-written HTML.

TypeScript · Astro Islands · Cloudflare Workers · GitHub API

Open tool →

Archived

Neural networks from scratch

Problem: Ready-made libraries like PyTorch hide what's actually happening inside a neural network behind a black box.
Solution: Rebuilt backpropagation and gradient descent from scratch, micrograd-style, using only NumPy and the underlying math.
Result: A deep understanding of the autograd mechanism — which made debugging models in later projects far more reliable.

Jupyter Notebook · Deep Learning · NumPy

View on GitHub → 0

Active

Local RAG Agent

Problem: Document Q&A systems usually depend on cloud AI services — which raises both a privacy problem and an API cost.
Solution: Built a fully local RAG pipeline using Ollama for the local LLM, Chroma as the vector store, and LangChain to wire them together.
Result: Documents can be queried offline, with no API key and no internet connection required.

Python · Ollama · LangChain · Chroma · RAG

View on GitHub → 0

Active

Local Voice Agent

Problem: Voice assistants usually depend on cloud speech-recognition services, which adds latency and a privacy constraint.
Solution: Built a Python pipeline chaining local speech-to-text and text-to-response stages, with no external API calls.
Result: A working offline voice-agent prototype with low response latency.

Python · Speech-to-Text · AI Agent

View on GitHub → 0

Archived

Customer Churn Prediction

Problem: If a company can't tell in advance which customers are about to churn, it can't act to retain them.
Solution: Trained and compared five models — Logistic Regression, Decision Tree, Random Forest, SVM and XGBoost — on historical customer data.
Result: XGBoost gave the most reliable result: ROC-AUC of 0.72, 86% accuracy, and the shortest training time.

Jupyter Notebook · scikit-learn · XGBoost · ML

View on GitHub → 0

Archived

Cirrhosis Survival Prediction

Problem: Estimating a cirrhosis patient's survival probability from clinical measurements alone is a hard classification problem.
Solution: Built a multiclass classification model from clinical features like age, bilirubin and albumin, including the data-cleaning and feature-engineering steps.
Result: Hands-on experience working with medical data and applying multiclass classification metrics.

Jupyter Notebook · Medical ML · Classification

View on GitHub → 0

Active

Gemini Image MCP

Problem: Generating images directly from Claude required a paid API.
Solution: Wrote an MCP server that connects to Gemini's web app through the user's own browser session, with automatic login via Firefox.
Result: Free image generation and editing from Claude Code or Claude Desktop, with every image saved into one HTML gallery.

Python · MCP · Claude · Gemini

View on GitHub → 1

Active

Telegram Account MCP

Problem: Important notes, links and files pile up in Telegram's "Saved Messages", but AI tools can't read them.
Solution: Built an open-source MCP connector on Telethon and FastMCP that lets Claude Desktop read Telegram messages, find files and reply to them.
Result: Claude can read and sort Telegram notes directly; all sessions and keys stay 100% local.

Python · MCP · Telethon · FastMCP · Claude

View on GitHub → 3

Active

MIU secretary (UserBOT)

Problem: Replying quickly to everyone who messages a personal Telegram account is hard, but automating it usually needs a bot token or a Telegram Business subscription.
Solution: Built a Telethon-based userbot that understands context, replies to people who message you, surfaces important messages immediately, and understands images and voice notes.
Result: Never pretends to be you — it operates openly as an AI secretary, managed entirely through your own inbox.

Python · Telethon · Telegram · AI Agent

View on GitHub → 0