Benchmarking, Schema Migrations, Retrieval-Augmented Gen AI on Python & Postgres

Join us for an exciting collaborative talk night with the Puget Sound Python meetup! (aka PuPPy) Cross-Posted at https://www.meetup.com/psppython/events/313614285/

This event starts 30 minutes earlier than usual. Doors open at 5:30 and speakers will begin at 6:00 sharp.

We have a bunch of talks planned about Postgres and Python:

• Junaid Hasan: Benchmarking Database systems on the NYC Taxi Database
• Ivan Schneider: Schema Evolution Automation (SEA)
• Shoumik Gandre: Using Python Postgres for Async Rag Backfilling

We're also looking forward to hearing from Andrew Beyer, a Senior Developer at PATH.

.

===== Talks:
Junaid Hasan: Benchmarking Database systems on the NYC Taxi Database

Data science workflows on local hardware often face a “Mid-Size Data” problem: datasets between 1GB and 100GB that are too large for spreadsheets but inconvenient for distributed clusters. This study benchmarks five data management systems (PostgreSQL, SQLite, Pandas, DuckDB, Polars) on a 41-million row NYC Taxi dataset using a standard Apple M1 laptop. Our results reveal an 18,000x difference in ingestion latency between row-stores and zero-copy columnar engines. Furthermore, forensic analysis of query plans demonstrates that execution architecture (Vectorized vs. Volcano) dominates optimizer intelligence for analytical workloads. Finally, a sensitivity analysis over 20 iterations exposes significant volatility in SQLite’s query planning (σ > 400s) compared to the stability of DuckDB and PostgreSQL.

.

Ivan Schneider: Schema Evolution Automation (SEA)

SEA is an ongoing project designed to take the gruntwork out of database development. Change a database field in a real Python app and you're updating six things: the model, the migration, the API, the CLI, the tests, the front-end. SEA is a CLI that automates that propagation — you describe the change in plain English, it classifies the intent, flags what's dangerous, and generates the code. The immediate goal for the project is to make schema changes fast enough that they don't interrupt your flow. The reach goal is a SEA change: extend that to downstream consumers so the entire propagation chain, from database to front-end to API contracts, is something SEA coordinates rather than something you track manually.

.

Shoumik Gandre: Using Python Postgres for Async Rag Backfilling

I'll be discussing this repository where I collaborated to create an easy to deploy rag which uses Postgres to track the status of the async request (scheduled in rabbitmq waiting to be stored in vector database) https://github.com/Gauri-Khanolkar1/rag-in-a-box

.

===== Speakers:
Junaid Hasan is a Math and Data Science PhD student at UW Seattle, graduating this June. Most of his day-to-day research focuses on AI interpretability and computational number theory, and managing all the data got him interested in benchmarking databases. Before this, he spent some time working on cryptography and algebraic geometry. He's currently looking for roles in the Seattle area after graduation. Outside of work, he's usually hiking or watching soccer, cricket, and F1.

Ivan Schneider is the creator of Factum Erit, a conversational task management platform that connects AI (Claude) to a structured database via MCP — so tasks captured in natural language land in real state, not just a chat window. He started his career as a programmer/analyst (Carnegie Mellon, BS Information and Decision Systems), spent 20 years covering financial technology as an editor and analyst, and returned to building software in 2024 with AI-augmented development.

Shoumik Gandre is a Software Engineer at Amazon Web Services in Seattle where he works on AWS CloudFormation. Outside of work, he enjoys applying machine learning to problems in his own life. Recently that's meant experimenting with deep learning and computer vision to see if he can build a system that helps him learn partner dancing. Earlier in his career, he worked as an AI Engineer at a startup analyzing social networks - building systems that linked people and concepts mentioned on social media to Wikidata, grounding language model outputs in real-world entities.