Danial Khilji

Designed and optimized a semantic knowledge representation system, modelling relationships between audience using semantic similarity and weighted graph similarity, enabling >10K users in just one quarter to map source taxonomies to ad-tech platforms like Meta and Google.

Partnered with engineers to automate mappings pipelines using Airflow, reducing turnaround time from weeks to under one week through automated similarity, KPI calculations, and reporting.

Built and maintain a Python package for taxonomy similarity, adopted across teams, with features for data cleaning, KPI generation, model evaluation, results evaluation and configurable weighting of embedding models.

Integrated pre-trained and fine-tuned BERT-based models, developing a weighted similarity algorithm, and continuously improving mapping performance.

Developed an evaluation framework with text corruption strategies, ground-truth validation, and confusion matrices to measure model accuracy.

Implemented LLM-based features (GPT-4, Gemini 1.5 pro) for taxonomy enrichment and lightweight RAG, improving semantic accuracy without on external databases.

Built an LLM supervision layer, using GPT-4/4o, to pre-validate mappings before human review, significantly reducing manual efforts improving efficiency.

Investigated geo-targeted advertising inaccuracies in Snowflake, improving device and email matching via probabilistic methods; developed a QA framework for partner data validation and onboarding.

Consulted on email matching optimization for campaigns using probabilistic matching algorithms and normalization techniques.

Built a LangGraph-based multi-agent system for ad-tech platform research to retrieve API details, audience taxonomies, and reach estimates, and automatically generate validated markdown reports with details and code snippets.

Developed MCP server to expose semantic similarity system to AI Agents.

Participated in WPP Data Challenge #4 and #5. Winner of Data Challenge #5

Contributed to the Audience Knowledge Graph (AKG) project, developing taxonomy mapping module, data cleaning steps for data clean rooms, and Looker Studio dashboard for monitoring each module progress among other tasks.

Created python package to expose taxonomy mapping project to different teams within company. Package uses pre-trained BERT language models from Hugging Face library to generate vector embeddings which are used to calculate cosine similarity.

Focus on taxonomy mapping work also known as Named Entity Resolution (NER). Improved the previously developed algorithm by removing loops and reducing time complexity from O (n) to O (1).

− 10 weeks internship in WPP to give an understanding of how, as a group of agencies, WPP work in the creative world, help their clients grow and build advertisement campaigns.