Hi, I'm Emad.
I build intelligent systems.

Designed and built an on-demand model-retraining pipeline on SageMaker Pipelines, turning researchers' feature-generation, training, and evaluation code into reusable DAG steps. Enabled parallel experiments tracked in MLflow, cutting experiment turnaround by an estimated ~40%. Built proof-of-concept Snowflake processing and training pipelines and assessed the gaps for migrating AWS research workflows to Snowflake.

Contributed to CLEAR Investigate, Thomson Reuters' commercially launched agentic AI product for investigations. Built a central PydanticAI configuration layer for system prompts, tool descriptions and availability, and agent/subagent wiring, enabling experimenters to reconfigure agents from one place instead of editing scattered definitions. Migrated the team's existing LLM-as-judge evaluation capability into an organized repository for reuse and future CI/CD integration.

Built the LLM summarization workflow for CLEAR Business entity reports, live in production processing hundreds of reports daily. Designed selective XML extraction for targeted per-section context covering business overview, personnel, adverse media, and legal records. Integrated Bing Search and LLM verification to reduce same-name entity mismatches and linked claims to report sources. Refined extraction and prompts through SME annotation rounds against exemplars for 10-15 companies, and compared GPT-4 and Claude Sonnet on quality, latency, and cost to inform model selection.

Built research data infrastructure for CLEAR's entity-resolution system spanning ~800M entities and billions of documents, with a 700 idents/second target. Designed a PII-isolated AWS environment with SageMaker Studio, CloudFormation, and per-researcher profiles; defined a versioned schema reconciling two incompatible annotation sources, with Spark and Python pipelines to merge records and surface label conflicts; and replaced manual runs with scheduled, repeatable SageMaker Pipelines.

Shipped a sentence-embedding query-intent classifier for Checkpoint tax research, distinguishing broad conceptual queries from specific tax lookups to promote higher-level documents. Improved access to conceptual documents by 95% within a ~100ms latency budget. Identified the retrieval gap through SME and customer feedback.

Fine-tuned the open-weights T5 model to generate questions from documents, cutting manual data-curation effort by 40%. Experimented with input representations and evaluated generated text with BLEURT in a research collaboration with university partners.

Trained domain-specific Word2Vec embeddings and LSTM-based NLU models, with interpretability testing to make model decisions transparent.

Developed a conversational question-answering system combining syntactic and semantic document analysis with LSTM-based text tagging. Used semi-automatically generated ontologies to interpret user questions and identify intent, and experimented with knowledge graphs for knowledge representation in an industry research collaboration.