Vianova

The query API

Customers ask for their data as typed JSON, and the API compiles it to Snowflake SQL. A colleague wrote the first endpoint. I built the query language on top of it, and I've been its main developer ever since.

• Filters, aggregations, computed fields, joins, pivots, custom metrics and geospatial operations
• Raw points snapped onto roads, areas or H3 cells
• Customer files turned into queryable, map-ready datasets, with dashboards created automatically
• About 80% of the commits on this code, and 83% of the lines still there today

AI in the product

Since September 2024, you can type a question in plain language. Claude writes the query, the API validates it, and it can become a chart. I built this from the first commit, and I still maintain it.

• Tested like code: CI fails if query generation drops below 85%, or suggested prompts below 90%
• Explanations of a generated query, and suggested prompts for each dataset
• One interface over Anthropic, Bedrock and OpenAI, now behind nine use cases
• PowerPoint reports built from a dashboard, with AI-written analysis in several languages, from prototype to production in five months
• AI dataset categorization, from hackathon idea to default on upload in a month

When production pushes back

Real clients and real data find what tests don't. These are the ones I'm happiest about.

• Year-long queries didn't fit the 29-second API timeout. Routing them to a daily rollup cut about 40M rows to about 1.7M, with no change for callers
• One slow client could freeze the API for everyone. I traced it to cached routes blocking the event loop, fixed it, and added request deadlines and query cancellation
• A live traffic-data pipeline ran on a colleague's laptop. I moved it to scheduled cloud jobs with alerts, five weeks before a vendor deadline
• Pydantic v1 to v2 across the backend in about four weeks: 217 files, with the API contract unchanged

Data

The data reaches the API through a dbt project of 483 models, and I've been its top committer since 2022.

• Brought unit testing to the project in 2022, and wrote about 42% of today's unit tests
• Rebuilt regulation compliance history so each hour belongs to the rule version in force, checked against 176,000 production version pairs
• Built the first dbt pipeline into the stream serving table, then made it safe: partial data can't overwrite full days, and empty merges fail loudly
• Automated 30-day deletion of citizen report photos and comments, backups included, with a test that fails if anything is left