Tuba Siddiqui

where data meets intent and systems learn to think

Most people use frameworks. I'm learning to build them.
Most people read papers. I'm implementing them.
Most people know what works. I'm discovering why.

This is a self learning journey into the architecture of intelligence itself. From the mathematics of attention to the pragmatics of production. From transformer implementations in NumPy to RAG systems that ship.

Credentials: MSc Data Science @ Sapienza University of Rome
Experience: Production systems @ Fastweb (Italy) & HBL (Pakistan)
Philosophy: Every abstraction hides a lesson. Every framework makes assumptions. Know both.

An AI companion that remembers. Integrated with WhatsApp for daily reflections, emotional check ins, habit tracking. Not just responding but understanding continuity.

Stack: FastAPI • WhatsApp API • Vector Memory Read how it thinks

Automated intelligence that captures scattered feedback, finds patterns, generates summaries. The system that thinks about what users think.

Stack: n8n • Gemini • Workflow Orchestration

Read how it thinks

Currently inside: "Attention Is All You Need" (Vaswani et al.)
Implementing every component from scratch. No tutorials. Just math, paper, and NumPy.

Also reading:

• GPT-3 paper (architecture evolution)
• RAG paper (Lewis et al., 2020)
• LLaMA technical report (understanding what changed)

This month's experiments:

• ✓ Self-attention mechanism (50 lines, pure NumPy)
• ⟳ Building RAG without frameworks (5 different approaches)
• → Comparing LangChain vs pure implementations

Latest writing: The Feedback Collector That Thinks
On building systems that understand user intent without human intervention

What broke recently: Discovered my RAG was confidently hallucinating because chunking strategy was too naive. Fixed with semantic splitting and reranking. Latency increased 40% but accuracy improved 60%. Worth it.

Current question: When do frameworks actually help vs just hide complexity? Building 5 RAG implementations to answer with data.

Papers implemented from scratch        → 3
RAG systems built this month           → 2 (working on 3 more)
Lines of attention code written        → ~847
Failed experiments documented          → 12
Production tradeoffs measured          → cost • latency • accuracy

Fetal Brain Structure Segmentation • Master's Thesis
Attention U-Net meets Grad-CAM. Deep learning for medical anomaly detection in ultrasound imagery.
→ Repository

VisionRoad • Bachelor's Engineering Thesis
YOLOv3 and Raspberry Pi teaching traffic lights to adapt.
→ Project

Voice Recognition System
MATLAB, DSP, and the mathematics of identifying speakers.
→ Details

Address Data Cleaning Tool
Fuzzy matching and postal validation. Making messy data make sense.

Languages: Python • R • MATLAB
Learning: PyTorch • TensorFlow • NumPy • FAISS
Building: FastAPI • Streamlit • Docker
Deploying: AWS • n8n workflows
Exploring: OpenAI • Anthropic • Groq
Analyzing: PowerBI • OpenCV

Open to discussing:

• RAG optimization strategies (chunking, reranking, hybrid search)
• Production LLM systems (cost vs latency vs accuracy tradeoffs)
• Building AI systems without heavy frameworks
• Technical writing about deep learning journeys

Would love to collaborate on:

• Projects that require understanding systems from first principles
• Building tools for other learners diving deep into AI
• Writing about the gap between research papers and production

Best way to reach me: Email or LinkedIn for thoughtful technical discussions