Experience.

At Handshake AI, I write evaluation pipelines for Python, C++, and Go repositories. The work combines hard test design, multi-file debugging, and checking that fixes survive the full regression suite.

Evaluation pipelines

Develop LLM evaluation pipelines across Python, C++, and Go codebases, with Docker-based reproducible environments and fail-to-pass test suites that probe model behavior on production code.

Adversarial evaluation

Design prompts that surface weaknesses in long-context reasoning, multi-file edits, and tool use, then translate the findings into fine-tuning data and reusable prompt patterns.

Repository-scale debugging

Trace and resolve multi-file bugs in production-scale open-source repositories, validate fixes against regression suites, and return edge cases to evaluation tasks.

• Python
• C++
• Go
• Docker
• LLM evaluation

At Amazon, I replaced weekly manual reporting work with a daily pipeline, added data-quality checks, and made common reporting questions self-service.

Analytics orchestration

Migrated manual analytics workflows to a Python- and Airflow-orchestrated pipeline on PostgreSQL, reducing weekly data preparation from 12 hours to under 2 and enabling daily reporting.

Data quality

Built modular Python validation with dbt-style tests to detect schema drift, anomalies, and outliers, eliminating recurring downstream report errors and improving training-data integrity.

Self-service reporting

Standardized organizational reporting through reusable SQL models and Tableau dashboards, replacing ad hoc analyst requests with a consistent self-service data layer.

• Python
• Airflow
• PostgreSQL
• SQL
• Tableau
• Data validation

At Accenture, I built NLP, forecasting, churn, and recommendation models, then turned them into AWS services that could be versioned, tested, and monitored.

NLP systems

Designed end-to-end NLP pipelines with PyTorch, Hugging Face Transformers, and spaCy for text classification, named entity extraction, and intent recognition, shipping models into production applications.

Predictive ML

Built scikit-learn and XGBoost systems for forecasting, churn prediction, and recommendations, improving prediction accuracy by double digits over rule-based baselines.

Production ML

Productionized models on AWS with Docker, SageMaker, and FastAPI, adding versioning, A/B testing, and Prometheus monitoring to reduce iteration cycles from weeks to days.

• PyTorch
• Hugging Face
• spaCy
• scikit-learn
• XGBoost
• AWS
• SageMaker
• Docker
• FastAPI
• Prometheus

I supported academic research at KL University while completing my undergraduate studies.