Rafael Santos
Geonix — Principal
TODO_USER_COPY: 5–8 word outcome-anchored headline
TODO_USER_COPY: 15–22 word sub-tagline naming stack + speed-of-delivery
TODO_USER_COPY: real benchmark (e.g., 4 TB → tiles in 12 min, 1 node)
live
Sample: anonymized Tokyo open-data probe traces
If any of this sounds familiar:
- TODO_USER_COPY: problem statement 1 (TB-scale pipeline pain)
- TODO_USER_COPY: problem statement 2 (visualization / query bottleneck)
- TODO_USER_COPY: problem statement 3 (agentic / NL-query pain)
How I work
Three packages, fixed scope, visible price floors.
Foundation
TODO_USER_COPY: from $XProduction-ready ingestion + calibration pipeline on the data you already collect.
- Schema + storage layout (Parquet / GeoParquet)
- Streaming or batch ingestion with backpressure + retries
- Calibration sprint on a sample dataset
- Observability: per-stage timings + correctness checks
Visualization & Query
TODO_USER_COPY: from $XTB-scale data made browsable: tile servers + NL query over your warehouse.
- MVT tile server (Rust or Go) over GeoParquet
- Browser app with MapLibre + deck.gl
- NL → SQL layer on DuckDB or your warehouse
- Latency budget + caching strategy
Agentic Layer
TODO_USER_COPY: from $XDomain-specific agents and the orchestration substrate they run on.
- MCP server exposing your domain tools
- Agent orchestration (typed tool I/O, retries, traces)
- Eval harness for agent behaviors
- Cost + latency telemetry
Selected evidence
TODO_USER_COPY: real benchmark headline
Methodology →Engagements under NDA
Japanese national ministry · Global automaker R&D · Research universities
About
TODO_USER_COPY: 2-sentence intro framed institutionally for EN audience.
More →Let's talk
20 minutes, video call, no sales script.