Muhammad Umar Saleem
AI Agent & Chatbot Developer | Data Scientist, NLP, ML , CV
Profile in progress
Muhammad Umar is building their profile!
I just wrapped up my latest project: an Intelligent Log Classifier. The core challenge wasn't just "making it work"—it was making it scalable and cost-effective.
In many AI applications, developers default to using Large Language Models (LLMs) for everything. This is expensive and slow. Instead, I built a Cascade Architecture that uses the "cheapest" brain power first:
🚀 Phase 1: Regex (The Specialist) Standard, recurring log patterns are caught instantly using Regex. Cost: $0.00.
🧠 Phase 2: BERT (The Generalist) If the log is unique but we have enough training data, our local BERT model takes over. It understands the "vibe" of the log without needing a cloud API.
✨ Phase 3: Google Gemini (The Expert) Only when the data sample is too small for BERT to be confident does the system call the LLM. It acts as a safety net for rare edge cases. Cost: Managed & minimal.
Why this matters: By building this "logic gate," I reduced API dependency by nearly 90%, ensuring the system is high-speed while keeping the cloud bill near zero.
Tech Stack: FastAPI | React | BERT | Google Gemini | Hugging Face | Python
1
24
I built an AI agent that manages inventory for me — and I'm sharing a quick demo video.
The problem I wanted to solve: when you're walking around a store or garage checking stock, you don't want to go back to a laptop just to look something up. So this agent works from your phone too.
You can type something like "I need Dettol 100g" or just take a photo of a product label. The agent then:
➡️Reads the product name and quantity from your text or image
➡️Checks live stock in the database
➡️If stock is low and a supplier is already saved, it drafts a purchase order email on its own
➡️If no supplier is saved, it searches the web and shows the best options with price and rating
➡️Before anything is sent, it shows you the result so you can approve, revise, or search again
Built with LangGraph for the agent workflow, Gemini for reading text/images, Groq for writing the emails, Tavily for supplier search, and Streamlit for the interface.
This was a good project to actually understand how agents work when they have to make decisions and loop back based on real data, not just answer one prompt and stop.
Still a prototype with more features on the way, but I'm incredibly happy with how it turned out. Watch the quick demo video to see it in action! 👇
1
20
🩺 AI-Powered Skin Health Assistant
I built a platform that identifies six common skin conditions from a photo — and then goes a step further with a built-in medical assistant that can actually talk you through it, remembering context across the conversation instead of answering questions one-off.
The Workflow:
📸 Image Classification — a ResNet50 model, fine-tuned via transfer learning, identifies the skin condition from a user-submitted photo.
🧠 Conversational Memory — built with LangChain, the assistant remembers earlier parts of the conversation, so users can ask follow-ups ("what about treatment options?") without repeating context.
💬 Symptom & Treatment Guidance — the assistant provides informational guidance on symptoms and general treatment options based on the identified condition.
⚡ Full-Stack Build — React frontend for a smooth user experience, FastAPI backend handling model inference and chatbot logic.
Tech Stack: ResNet50 (Transfer Learning), LangChain, React, FastAPI, Python
⚠️ Built as an informational tool, not a diagnostic replacement — always intended to point users toward professional medical care, not substitute for it.
🎬 Watch the demo below.
1
46
🚀 Turning Natural Language into Data Insights
I built a Business Intelligence (BI) Chatbot that bridges the gap between raw datasets and business decisions — no SQL, no manual data prep, no dashboard-building required from the user. Just ask a question in plain English and get an answer.
The Workflow:
🔐 Secure Access — Gmail OAuth login, private user IDs, full data control (delete anytime).
🛠️ Hybrid Processing — combines manual data cleaning with automated feature engineering behind the scenes to surface hidden trends, so the user never has to touch raw data.
💬 Natural Language → SQL — ask "Show me monthly sales" and get instant results, no query-writing needed.
📊 Visual Intelligence — automatically generates both tables and charts for every query.
Tech Stack: Python, NLP, Streamlit, MySQL.
Currently iterating on: smarter automatic chart-type selection, so every query gets the most useful visualization by default.
Data shouldn't be a headache — it should be a conversation.
🎬 Watch the demo below.
1
81