MVP Roadmap
Approach
This is a background workstream run in parallel with Helio's manufacturing business, so the roadmap is sequenced by milestone rather than calendar dates. Each milestone is demoable on its own; the POC (M0–M3) is the gate for deciding whether the pivot graduates to a funded track.
M0 — Foundation
Goal: spec-driven workspace ready; every later line of code traces to a spec.
| Deliverable | Definition of done |
|---|---|
Isolated repo (helio-ocean) | Standalone git repo, zero production dependencies |
| PDF planning deliverables | PRD, roadmap, user stories, feature specs rendered via the adapted pdf-generator skill |
| Requirement ID scheme | PRD-Rx → FS-x → AC-x.y, mirrored in docs/traceability.md |
M1 — Ingestion & Normalization
Goal: three heterogeneous sources land in one schema. Implements FS-1, FS-2, FS-3, FS-4.
| Deliverable | Definition of done |
|---|---|
| Sentinel-2 FDI adapter | Live STAC pull for the SoCal AOI; FDI math unit-tested against the Biermann et al. (2020) formulation |
| NOAA MDMAP adapter | Parses MDMAP export CSV format |
| Debris Tracker adapter | Parses export CSV; free-text notes extracted to structured fields, ambiguity flagged |
| Normalization pipeline | All sources → Observation schema → SQLite; idempotent re-runs; provenance retained |
M2 — Intelligence
Goal: the fused data says something a single source can't. Implements FS-5, FS-7, FS-8.
| Deliverable | Definition of done |
|---|---|
| Zone classification | Confirmed / suspected / reported zones emitted as GeoJSON with the evidence behind each |
| NL query | Claude (Bedrock) tool-use over the store; labeled offline fallback |
| Post-storm briefing | Generated for the Feb 2024 event window from live pipeline output |
M3 — Map POC
Goal: the 10-minute demo. Implements FS-6; exercises PRD success criteria SC-1..SC-6.
| Deliverable | Definition of done |
|---|---|
| Map-first app | MapLibre map with zone/detection/report layers, pre/post-storm toggle, popups, legend |
| Query panel | NL question box wired to FS-7 |
| Verification report | pytest suite green; traceability matrix complete; honest real-vs-fixture accounting per source |
Post-MVP Backlog (specced, not built)
Ordered by expected value for the pivot decision:
- Municipal stormwater record extraction — the hardest-data showcase: LA County MS4 monitoring PDFs → structured discharge/debris records (extends FS-3's extraction pattern).
- Drift modeling — Copernicus Marine currents/wind to project where detected debris will travel; turns zones into forecasts.
- MARIDA-trained classifier — replace FDI thresholding with a model trained on the open MARIDA labels; expected large precision gain.
- Copernicus Data Space adapter — credentialed CDSE ingestion slotting into the FS-1 interface.
- Alerting — post-storm detection spikes push notifications to subscribed users (Helio's alerting pattern).
- Multi-region + onboarding — AOI picker, per-region pipelines.
- Productionization — deploy on Helio's existing AWS patterns (Next.js frontend, containerized API, Bedrock agents); auth & multi-tenancy.
Explicit De-scopes
No calendar commitments, no hiring, no marketing site, no customer pilots until the M3 gate. Run cost target for the POC: $0 standing infrastructure (everything local; Bedrock invoked only on demand).