Ask the Warehouse
Production text-to-SQL assistant analysts actually use: semantic-layer allowlist, self-correction, eval-gated deploys — every answer shows its SQL.
LangGraph, MCP, Text-to-SQL, dbt Semantic, Evals
Read case studyDATA ENGINEER + APPLIED AI
Ameya Kashid.
I build data pipelines
and applied AI systems.
I’m based in Tampa. My work spans SQL, Python, dimensional modeling, and AI workflows. Here are the projects, architecture decisions, and technical details behind it.
01 / PROJECTS
My projects
Start with the problem.
Follow the decisions into the details.
Real-time Streaming Lakehouse
Migrated nightly batch to a Kafka → Flink → Iceberg lakehouse with dbt on streaming tables; dashboards run on the current hour.
Kafka, Flink, Iceberg, dbt
Read case studyDocument RAG with Measured Retrieval
Document RAG reviewers trust because every answer is cited and retrieval quality is a measured gate, not a vibe — exposed to agents over MCP.
Embeddings, Vector Search, Reranking, Evals, MCP
Read case studyLLM Evaluation Harness
Golden sets, CI regression gates, and observability that turned LLM releases from leaps of faith into gated deploys.
Golden Sets, CI Gates, Observability, Prompt Versioning
Read case studyChange Data Capture, Exactly Once
Replaced nightly full-table syncs with Debezium → Kafka → Iceberg upserts; idempotent consumers, managed schema evolution, replayable history.
Debezium, Kafka, Iceberg, Schema Registry
Read case studyData Reliability & FinOps
Moved breakage discovery from board meetings to CI: data contracts, dbt tests, anomaly baselines, per-model cost attribution.
Data Contracts, dbt Tests, Anomaly Detection, FinOps
Read case studyGeoAI Wildfire Risk, California
Spatial ML over terrain, fuels, and weather rasters — California wildfire risk as a living map, not another spreadsheet.
Geospatial ML, Satellite Imagery, Raster, Python
Read case studyMultimodal Document Intelligence
Production extraction at Lid Vizion: CV preprocessing + OCR + LoRA-tuned VLM, with a confidence-ranked queue so reviewers judge exceptions, not key fields.
OCR, VLM, LoRA, Human-in-the-Loop
Read case study02 / BACKGROUND
My experience
Data engineering and applied AI, with an interest in how systems are modeled, evaluated, and operated.
How I approach the workNov 2024 — Present
Lid Vizion
Apr 2021 — May 2023
Kate Industries
Apr 2020 — Mar 2021
Bizizz
03 / CONTACT
Let’s talk about the work.
Write to me about data engineering, applied AI, or dimensional modeling.