The Challenge: Your Agents. Your Models. Your Edge.
I recently participated in the NVIDIA Hack for Impact in London—a 48-hour intensive hackathon bringing together frontier builders and ML engineers. The challenge was to design an agentic application powered by open models, deployed entirely on a local-first NVIDIA DGX Spark (powered by the GB10 Grace Blackwell Superchip).
We were tasked with grounding our agents in real-world urban life using open data from the City of London, focusing on tracks like Economic Systems, Public Services, and Urban Operations.
What We Built: "Ask London"
My team developed Ask London, a voice-enabled chat agent that acts as a definitive, hallucination-free oracle for London's 33 boroughs.
Instead of relying on generic LLM knowledge, we built a system with two strict reasoning lanes:
- Retrieval (RAG + Graph): Cross-references cited facts and rankings from fragmented datasets (Deprivation, Health, Income).
- Computation (Text-to-SQL): Uses PydanticAI to calculate exact aggregates directly from a DuckDB backend.
Running locally on the NVIDIA Nemotron model via NIM, the agent assembles a mosaic of answers regarding housing affordability, life expectancy, and taxation. Most importantly, it never invents a number, if the data cannot answer the question, the agent safely returns "compute unavailable," with every figure traced directly to its source chunk or SQL query.
View the source code and deployment instructions on GitHub →
Gallery
