Ahmed Fraz

Key Skills

Software

Google Cloud Vertex AI

Top Subject Matter

No subject matter listed

Top Data Types

Computer Code Programming

Top Label Types

Text Generation

Freelancer Overview

I am a Full-Stack AI Automation Engineer with a strong foundation in designing and implementing data-driven systems across real estate, finance, and infrastructure domains. My experience includes building autonomous AI agents and "Human-in-the-Loop" workflows where I engineered processes to analyze raw lead intent, extract structured data using regex and JSON enforcement, and automate data annotation pipelines. I have hands-on expertise with tools such as n8n, Python (Pandas, Requests), SQL, and leading LLM APIs (OpenAI, Google Gemini, Anthropic Claude), as well as experience developing RAG pipelines for parsing and vectorizing unstructured documents for downstream querying and analysis. My approach combines rigorous data infrastructure design with practical automation, ensuring high-quality, reliable training data for AI systems. I am adept at optimizing data workflows, migrating and transforming datasets, and implementing robust error handling to support scalable AI and data labeling operations.

Labeling Experience

Real Estate Intent Classification & RLHF

Google Cloud Vertex AIComputer Code ProgrammingText Generation

Developed a classification dataset for Real Estate lead scoring. Task: Classify raw natural language inputs into intents: "High-Intent Buyer," "Investor," or "Casual Window Shopper." Method: Designed "Chain-of-Thought" system prompts to force the LLM to reason before classifying. Correction: Implemented a feedback loop where low-confidence scores were routed for human review, correcting the model's logic to improve future precision (RLHF approach). Outcome: Improved intent recognition accuracy by 40% through iterative prompt refinement and strict JSON schema enforcement.

Developed a classification dataset for Real Estate lead scoring. Task: Classify raw natural language inputs into intents: "High-Intent Buyer," "Investor," or "Casual Window Shopper." Method: Designed "Chain-of-Thought" system prompts to force the LLM to reason before classifying. Correction: Implemented a feedback loop where low-confidence scores were routed for human review, correcting the model's logic to improve future precision (RLHF approach). Outcome: Improved intent recognition accuracy by 40% through iterative prompt refinement and strict JSON schema enforcement.

2025 - 2025

Education

Work History