Grounded AI systems | Safe data access | Evaluation-first delivery
Portfolio · LinkedIn · Resume · Email
I build AI applications where the model is one component of a larger, testable system. My work connects data engineering fundamentals with RAG, Text-to-SQL, workflow automation, operational analysis, and deployment validation.
| Project | What it demonstrates | Evidence | Links |
|---|---|---|---|
| Safe Text-to-SQL | Schema and KPI context, SQL policy validation, bounded read-only execution | 10 tests; synthetic evaluator: 90% execution, 80% semantic correctness | App · Code |
| Enterprise RAG Knowledge Assistant | Grounded retrieval, document context, source-aware answers, and enterprise knowledge workflows | Architecture and retrieval behavior documented | Code |
| Meeting Notes | Structured extraction, reviewable drafts, approval before communication | 11 tests; public health endpoint reachable | API · Code |
| AI Log Analyzer | Parsing, incident retrieval, evidence, hypotheses, confidence, action gates | 6 local tests; deployable, not publicly hosted | Code |
| 01 · Ground | 02 · Guard | 03 · Measure |
|---|---|---|
| Retrieve schemas, documents, metrics, or source notes before generation. | Validate model output, restrict tools, use least privilege, and gate external actions. | Track correctness, groundedness, safety, latency, cost, and recovery behavior. |
| Project | Core lesson | Production question |
|---|---|---|
| Text-to-SQL | Valid SQL is not automatically correct SQL. | How do we protect the database and verify business meaning? |
| Enterprise RAG | Fluent answers need current, authorized evidence. | Was the right context retrieved, and can the answer be supported? |
| Meeting Notes | Structured extraction is safer than opaque rewriting. | What prevents invented commitments or an incorrect recipient? |
| AI Log Analyzer | Diagnosis should expose evidence and uncertainty. | How does an engineer verify the hypothesis before acting? |
A useful AI feature needs more than a model call. It needs trustworthy data, explicit contracts, validation at system boundaries, safe failure behavior, measurable quality, and an owner when something goes wrong.
- Grounding: retrieve schemas, documents, metrics, or source notes before generation.
- Safety: treat generated SQL, tools, and external actions as untrusted until authorized.
- Evaluation: separate correctness, groundedness, safety, latency, cost, and availability.
- Operations: distinguish health checks, smoke tests, local tests, and semantic evaluation.
- Data: carry meaning, freshness, lineage, and access scope into the AI workflow.
Evidence and status definitions
| Label | Meaning |
|---|---|
Reachable |
A public page or endpoint responded successfully. |
Smoke tested |
A representative hosted interaction completed. |
Locally validated |
Tests or compilation passed in the recorded environment. |
Deployable |
Deployment configuration exists, but a current public check is not claimed. |
These labels are intentionally narrower than “production-ready.” A URL, a passing test suite, and a successful model response each prove different things.
Python · SQL · FastAPI · Streamlit · PostgreSQL · SQLite · SQLAlchemy · Pydantic · sqlglot · LangChain · FAISS · Docker · pytest
Evaluation datasets · RAG freshness · SQL safety · AI observability · deployment contracts · incident debugging · human approval workflows
SQL / ETL / BI → Data Engineering → Python → ML
→ LLM Apps / RAG
→ Safe Agents / Evaluation
Streamlit Chatbot · RAG Q&A Demo · Machine Learning · Deep Learning · SQL Practice · Power BI