Software Engineer working across data and AI — designing the architecture first, shipping it to production, and staying calm when it breaks at 2am.
I work across three disciplines that reinforce each other: software engineering as the foundation, then data engineering and AI engineering built on top of it. That means I design typed, tested, well-bounded systems — and then apply them to pipelines and models that have to hold up in production, not just in a notebook.
I care about the parts most demos skip — idempotent pipelines, tested transformations, typed tools behind real boundaries, measured accuracy, and human-in-the-loop control before anything touches production data.
- Software Engineering: Python and TypeScript across the stack, typed contracts and clean module boundaries, FastAPI and REST API design, React front ends, pytest suites with enforced coverage, CI/CD quality gates, Docker packaging, and refactoring that treats readability as a feature.
- Data Engineering: Medallion lakehouse design (S3 → Snowflake → dbt → Airflow), dimensional and SCD2 modelling, incremental/MERGE fact loading, keyless cloud storage integrations, and read/write boundary separation by role.
- AI Engineering: LangGraph state machines, bounded tool-calling agents, hybrid RAG (pgvector + knowledge graph), QLoRA fine-tuning with execution-based eval harnesses, and LLM enrichment turned into queryable warehouse columns.
- Production Reliability & Incident Response: MLflow experiment tracking, drift and data-quality monitoring, structured tracing, human-in-the-loop approval gates on anything that mutates production data, and root-cause analysis that reports
UNRESOLVEDrather than guessing.
| Domain | Core Technologies & Tooling |
|---|---|
| Software Engineering | |
| Data Engineering & Warehousing | |
| AI Engineering | |
| MLOps, Cloud & Reliability |
Deep-Dive: Comprehensive Technical Stack & Tooling
- Languages: Python (3.12+), TypeScript, JavaScript, SQL (advanced — window functions, MERGE, recursive CTEs), Bash, Java (basics).
- Software Engineering: Typed interfaces and clean module boundaries, FastAPI and REST API design, JWT auth, SQLAlchemy, React 19 / Vite / Tailwind front ends, pytest suites with enforced coverage, strict typing (mypy) and linting (ruff), code review and refactoring, Docker packaging, semantic versioning and Makefile-driven developer workflows.
- Data Engineering: Medallion (Bronze/Silver/Gold) lakehouse design, star schemas, SCD Type 2 history, incremental fact loading, partitioning and clustering strategy, data contracts, idempotent DAG design, keyless S3↔Snowflake storage integrations, least-privilege role separation.
- Transformation & Orchestration: dbt (models, tests, macros, incremental strategies, exposures), Apache Airflow (DAGs, sensors, retries, backfills), Apache Spark / PySpark, Databricks notebooks and workflows, Delta Lake.
- AI Engineering: RAG (hybrid dense + lexical retrieval, knowledge graphs), LangGraph state machines, typed tool-calling agents, QLoRA / LoRA fine-tuning, vLLM serving, prompt and eval harnesses, execution-accuracy benchmarking, LLM enrichment into warehouse columns, text-to-SQL, guardrails and human-in-the-loop gates.
- MLOps & Observability: MLflow tracking and registry, CI/CD quality gates on model accuracy, containerized serving, drift and data-quality monitoring, structured tracing (LangSmith).
- Cloud & Infrastructure: AWS (S3, IAM, Lambda, EC2), Snowflake, Vercel, PostgreSQL + pgvector, SQLite, GitHub Actions, Linux.
Certifications & continuous learning: Databricks Fundamentals, Claude AI Fluency.
"A pipeline you can't re-run isn't a pipeline. An agent you can't measure isn't a system. Good engineering is what makes both of them maintainable."
- Software Craft First: Clear interfaces, low coupling and high cohesion, typed contracts end to end. The same discipline that makes an API maintainable is what makes a pipeline re-runnable and a model reproducible.
- Measure, Don't Assert: Every AI claim belongs behind an automatic evaluation harness. If accuracy can't be reproduced on a held-out set, it isn't a result.
- Idempotent & Observable by Default: Re-runnable pipelines, tested transformations, structured traces and honest failure modes — including saying "unresolved" rather than guessing.
- Built for Failure, Not Just Success: Incident response, human-in-the-loop approval gates and graceful degradation are designed up front, not bolted on after the first outage.


