Project Overview

I built a production-grade data ecosystem to transform millions of raw NYC Taxi (TLC) records into a scalable analytics platform. This project implements a full-lifecycle ELT pipeline, from automated ingestion and Medallion-style transformations to executive BI insights and an AI-integrated semantic layer for natural language querying.

The core mission was to bridge the gap between complex engineering and business operations. By implementing a semantic layer via an MCP Server, It enabled non-technical stakeholders to "chat" with their data—asking complex questions and receiving real-time answers without ever writing a line of SQL or code.

The Tech Stack

• Orchestration

  Dagster for end-to-end pipeline visibility.

  

• Ingestion

  dlt (data load tool) for "Pythonic" extraction and loading.

  

• Transformation

  dbt + MotherDuck/DuckDB to power a Bronze-Silver-Gold Medallion architecture.

  

  

• AI Integration

  MCP Server to enable LLMs to query the data warehouse directly.

  

• Consumption

  Evidence for BI-as-code and Streamlit for interactive data apps.

  

  

Project Overview

Transportation planners often need to identify network bottlenecks or evaluate scenario performance instantly, but are frequently siloed by the need for complex SQL queries across massive relational databases. I built Travel Model Assistant to democratize this data, creating an Agentic Semantic Layer that allows analysts to "talk" to their models.

The core innovation is a specialized Context Layer designed to handle industry-specific jargon like "V/C Ratios" and "PCE Volumes." Unlike standard chatbots, this system utilizes Functional Tool-Calling; the AI agent autonomously invokes a custom spatial engine to generate thematic Folium maps, applying hex-code color scales to visualize network performance and scenario deltas in real-time.

Example Use Case

The Tech Stack

• LLM Orchestration

  LangChain + Gemini for agentic reasoning and complex SQL tool-calling.

  

  

• Context Engineering

  Engineered a Domain-Specific Prompt Layer that resolves transportation anaphora and ambiguous intent into precise SQL operations across DuckDB.

  

• Frontend & Visualization

  Streamlit for the UI and Folium for reactive geospatial mapping.

  

  

Project Overview

Managing the "curb" is one of the most data-intensive challenges in modern urban mobility. At my current firm, I engineered an automated Computer Vision Pipeline that transforms raw Street View imagery into machine-readable parking regulations, fully compliant with the Curb Data Specification (CDS).

By utilizing Gemini as the multimodal intelligence engine, I bypassed traditional OCR limitations to interpret complex, overlapping parking signs in varying environmental conditions. To bridge the gap between LLM creativity and database reliability, I implemented a Pydantic-enforced Semantic Layer, ensuring that every extracted regulation strictly adheres to the CDS nested JSON schema.

Multimodal Extraction Example

To ensure regulatory compliance, the pipeline must translate visual cues (colors, arrows, time-windows) into strict Curb Data Specification (CDS) attributes. Below is an example of the agent identifying a multi-regulation sign:

Raw Street View Input

AI-Extracted Regulations

The Tech Stack

• Intelligence & Extraction

  Gemini for multimodal sign interpretation and Pydantic for enforcing strict Curb Data Specification (CDS) schemas.

  

  

• Data Infrastructure

  Google Cloud Storage (GCS) for scalable raw image hosting and PostgreSQL for querying extracted regulatory metadata.

  

  

• QAQC & Monitoring

  Streamlit internal dashboard for human-in-the-loop verification, accuracy scoring, and rapid prompt iteration.