Akshay Sharma · Data engineering · Applied AI research

Evidence over hype

My research asks practical questions about AI in data systems and answers them with executable benchmarks: does the translated pipeline produce the same data, does the fix leave the tables correct, and when should an agent hand off to a person.

• Domain-Driven Data and AI Platforms: Governing Analytics and Machine Learning Pipelines at Scale IET Software · 2026 A domain-driven reference architecture for governing analytics and ML pipelines in large enterprises.
• LEGATO: Agentic Large Language Model Translation of Legacy ETL Workflows to Modern Dataflow Code with Test-Based Semantic Verification Preprint · 2026 An agent that translates legacy ETL into Spark and checks every translation against executable data tests.
• Fix or Page? Silent Harm and Safe Escalation When LLM Agents Remediate Data Pipeline Incidents Preprint · 2026 64 broken pipelines, three model sizes: data quality monitors approved every harmful fix, while the agents' own confidence was the useful safety signal.
• When the Rules Collide: Explainability, Fairness, and Model Risk Governance in Credit Machine Learning Preprint · 2026 A review of how fair lending law, model risk guidance and capital rules constrain one another in credit ML.
• An Event-Driven Medallion Architecture with Domain-Driven Design for Training-Serving Consistent Credit Risk Feature Engineering Preprint · 2026 Event-driven medallion pipelines that keep credit risk features consistent between training and serving.
• Agent-Oriented Reference Architecture for Domain-Driven Enterprise Data Platforms Preprint · 2026 A reference architecture for running AI agents over domain-owned enterprise data.
• Domain-Driven Governance for Regulated Lakehouses: A Contract-Based Architecture and Reproducible Evaluation on Snowflake Preprint · 2026 Governance that propagates across domains, measured with a reproducible Snowflake benchmark.
• Proactive Credit Line Increase Using Risk-Constrained Explainable Machine Learning Preprint · 2026 Credit line decisions as constrained profit maximization, with calibrated models, SHAP explanations and fairness limits. A. Sharma, G. Sharma
• FraudRAG: Knowledge Graph-Augmented Retrieval for Real-Time Financial Statement Fraud Detection Using a Domain-Driven Medallion Architecture Preprint · 2026 A Graph-RAG pipeline combining a Neo4j knowledge graph and vector retrieval to check financial statements.
• A Domain-Driven Reference Architecture for Scalable and Governed Data Systems in Regulated Enterprises Preprint · 2026 The reference architecture behind domain-owned, governed data platforms in banking.
• Risk-Aware FinOps Optimization Using Machine Learning in Regulated Multi-Cloud Environments Under review · IIPEM 2026 Calibrated cost forecasts and tail-risk optimization for cloud spend under data residency rules. G. Sharma, A. Sharma, S. A. Syed
• Risk-Aware Self-Healing Cloud Infrastructure Using Agentic AI and Policy-Based Remediation for Regulated Financial Systems Under review · ISASF 2026 Agents propose repairs; a deterministic policy gate decides what may run. G. Sharma, A. Sharma, D. Singh

Open benchmarks and code

15 years of production pipelines

Batch and streaming platforms, lakehouses, legacy ETL modernization and machine learning pipelines, in industries where wrong data has real consequences.

• 2026 to now
  Salesforce · Lead Member of Technical Staff
  Customer Data Graphs
• 2025 to 2026
  Capital One · Principal Data Engineer
  Domain-driven medallion data architecture
• Until 2025
  Discover Financial Services · Principal Data Engineer
  Credit risk and fraud data platforms
• Earlier
  Tata Consultancy Services and Wipro
  Data engineering across financial services, retail and healthcare

Education: M.S. in Data Science, Bellevue University. Earlier hands-on projects are on the projects page.

A newsletter on AI agents that keep data correct

Every two weeks: one finding, the numbers behind it, and something a data team can do on Monday.

• Your Data Monitors Will Approve the Wrong Fix Oct 2026 What I learned watching AI agents repair 64 broken data pipelines.

Articles in other publications

• Benchmarking Claude Haiku, Sonnet, and Opus on 64 Broken Data Pipelines HackerNoon · Nov 2026 Which model tier is worth paying for when AI agents repair data pipelines, and why every tier still needs verification.

Reviewing, judging and speaking

Talks, reviews and collaboration

I'm open to speaking, reviewing and research collaboration on data engineering and AI agents.

• Emailakshay.sharma1088@gmail.com
• LinkedInlinkedin.com/in/akshay-sharma-1088
• GitHubgithub.com/akshaysharma1088
• Substackakshaysharma10.substack.com
• LocationChicago, Illinois

Find my work

• ScholarGoogle Scholar profile
• ORCID0009-0009-5718-2187
• ResearchGateresearchgate.net/profile/Akshay-Sharma-76
• SessionizeSpeaker profile
• DZoneAuthor profile
• HackerNoonAuthor profile
• SSRNPreprints on SSRN, linked by DOI above
• ZenodoCode and data archived with DOIs, listed under Open benchmarks