riprap-nyc / slides /VIDEO_TRANSCRIPT.md
seriffic's picture
ship: v0.5.0 code changes — compare UI + cleanup pass
caa28aa

Riprap — Demo Video Transcript

AMD × lablab.ai Developer Hackathon · May 4–10 2026

Target: ~5 minutes


[SLIDE 1 — Title card] · ~0:00–0:10

SCREEN: Slide 1. Riprap logo. "Citation-grounded NYC flood-exposure briefings, on AMD MI300X."

Climate risk is one of the most consequential datasets in real estate and urban planning right now. But the tools that exist today give you a score. A number from one to ten. No explanation. No sources. Just a black box. We built Riprap to be the audit trail behind that number.


[SLIDE 2 — The problem] · ~0:10–0:30

SCREEN: Slide 2. "Climate risk data is a black box." Two boxes: market scores vs Zillow pulling climate data.

First Street gives you a flood factor. ClimateCheck gives you a percentile. Jupiter charges enterprise rates for a proprietary model. In November 2025, Zillow removed climate risk scores from listings entirely — under pressure from the real-estate industry. When a number meets resistance, the only defense is the audit trail. Riprap is the audit trail.


[SLIDE 3 — Solution] · ~0:30–0:40

SCREEN: Slide 3. Screenshot of the Riprap UI — briefing prose with citation chips, map panel, stone trace.

Type any address in New York City. Get back a written briefing where every numeric claim — every flood depth, every complaint count, every risk percentage — links to its primary public-record source. Federal data. City data. Apache-2.0 models. Nothing proprietary.


[SLIDE 4 — Civic-tech case] · ~0:40–1:00

SCREEN: Slide 4. Four boxes: NY Disclosure Law, DEP Stormwater Plan, EJNYC FVI, No commercial APIs.

New York's property disclosure law — March 2024 — requires sellers to disclose flood history. Riprap is the citable narrative that makes that disclosure meaningful. The DEP's $30 billion stormwater priority list covers 86 sites. Riprap provides the per-neighborhood evidence layer that backs up that ranking. And because every model is Apache-2.0 and every dataset is public record, environmental justice advocates can audit the same system that a developer uses. No commercial gatekeeping.


[SLIDE 5 — Architecture] · ~1:00–1:30

SCREEN: Slide 5. "Five Stones fan out. One cited briefing comes back." Four evidence cards (Cornerstone, Keystone, Touchstone, Lodestone) + Capstone bar at bottom.

The architecture is called Five Stones. A natural-language query hits the Planner — Granite 4.1 3B — which classifies intent and selects a specialist roster. Each Stone is a class of evidence. Cornerstone reads the hazard record: Sandy inundation zones, FEMA flood maps, USGS high-water marks, Prithvi satellite imagery. Keystone reads what's exposed: MTA stations, schools, hospitals, building footprints from our TerraMind NYC fine-tune. Touchstone reads what's happening now: live FloodNet sensors, 311 flood complaints, NOAA tide gauges. Lodestone looks forward: NPCC4 sea-level projections, our Granite TTM Battery surge nowcast. Then Capstone — Granite 4.1 8B on vLLM — synthesizes everything into a four-section briefing. Every numeric claim must cite its source, or the Mellea rejection sampler rerolls it. The briefing doesn't publish until all four grounding checks pass.


[SLIDE 6 — Fine-tuning] · ~1:30–1:50

SCREEN: Slide 6. Three fine-tune cards: Prithvi-EO-2.0-NYC-Pluvial · TerraMind-NYC-Adapters · Granite-TTM-r2-Battery-Surge.

We trained three NYC-specialized models on AMD MI300X hardware, all published Apache-2.0 on Hugging Face Hub. Prithvi-EO-2.0-NYC-Pluvial detects pluvial flooding from Sentinel-2 imagery — 0.60 IoU on the Ida test set, a 6× lift over the baseline. TerraMind-NYC-Adapters adds LoRA adapters for building footprint and land-use classification, plus 6 points of mIoU in 18 minutes of training. And Granite TTM r2 fine-tuned on the Battery tide gauge gives us a 9.6-hour surge residual nowcast at 35% lower RMSE than persistence. These aren't experiments. They're in production in every briefing.


[SLIDE 7 — Demo intro] · ~1:50–2:00

SCREEN: Slide 7. "Live demo." Query text: "I'm thinking about renting an apartment at 80 Pioneer Street, Brooklyn. Should I worry?"

Let's run it live. Three queries, three different intents.


[DEMO CLIP 1 — Pioneer Street, single address] · ~2:00–2:40

SCREEN: Cut to recording riprap-demo-20260506-234537.webm at t≈62s.

  • Left panel: briefing fully rendered. Title "Flood-exposure briefing · 80 Pioneer Street, Red Hook."
  • Sections 01 Status through 04 Policy context visible with inline [1] [2] [3] citation chips.
  • Right panel: Sandy flood map showing Pioneer Street pinned inside the inundation zone (blue overlay).
  • Status bar: intent: single_address · 19 specialists · attempt 1 · done

Thirteen seconds end-to-end. Nineteen specialists fired. The briefing tells you: Pioneer Street sits inside Hurricane Sandy's 2012 inundation zone, 0.82 metres above the nearest drainage channel, in the 78th percentile for water accumulation risk. FloodNet sensor FN-BK-018 — two blocks away — has logged four flood events since 2023. The DEP's high-intensity scenario puts the site under six inches of standing water. Every number has a footnote. Every footnote resolves to a named public dataset.

SCREEN: Slow scroll of left briefing panel while voiceover continues. Citation chips [1] [2] [3] visible inline. Bottom of panel shows section 04 "Policy context" with RAG passages from NPCC4.

The map on the right isn't decorative — it's live. The layers are grouped by Stone, so you can see exactly which evidence tier each visual comes from.


[DEMO CLIP 2 — Mellea 4/4 grounding card] · ~2:40–3:05

SCREEN: Recording at t≈270s. Right panel scrolled to Capstone section.

  • Capstone card: "grounding checks: 4/4 passed", rerolls=0, passed=4, attempt=1.
  • Four check items: numerics_grounded · no_placeholder_tokens · citations_dense · citations_resolve

Here's the proof. Mellea ran four grounding checks on the completed briefing: every non-trivial number appears verbatim in a source document; no template fragments leaked through; every number has a citation in the same sentence; every cited ID resolves to an actual input document. Four of four. First attempt. Zero rerolls. This is what "every number cites its source" looks like as a machine-verifiable claim, not a marketing promise.


[DEMO CLIP 3 — Hollis, Queens · neighborhood intent] · ~3:05–3:30

SCREEN: Recording at t≈510s. New query: "Hollis, Queens."

  • Status bar: intent: neighborhood · 9 specialists · attempt 1 · done
  • Left panel: neighborhood briefing — NTA-level statistics, DEP stormwater scenario percentages, 311 flood complaint counts.
  • Right panel: Cornerstone section with Sandy inundation percentage for the NTA + FEMA layer.

Same system, different intent. "Hollis, Queens" is a neighborhood query — nine specialists instead of nineteen, NTA-level aggregates instead of point data. The planner classified it in under a second and dispatched the right Stone roster automatically. Hollis is a stormwater-flooding neighborhood, not a coastal one. The briefing reflects that: Sandy inundation is low; the DEP moderate-intensity scenario covers 22% of impervious surface; 311 flood complaints cluster around the 180th Street drainage corridor. Different geography, different risk profile, same citation standard.


[DEMO CLIP 4 — Compare · Pioneer vs Gold Street] · ~3:30–4:00

SCREEN: Screenshot compare-hf.jpg — the live HF Space compare result.

  • Title: "COMPARE 80 PIONEER STREET BROOKLYN TO 100 GOLD STREET MANHATTAN"
  • Key differences bar at top: Status: 80 vs 100 · Empirical: 65 vs 26 · Modeled Drainage (HAND): 3.81m vs 38.2m
  • Side-by-side Status sections — Pioneer: "exposed to flood risk, Sandy inundation zone, TWI 14.79." Gold St: "moderate flood exposure, HAND 6.42m, mid-slope position."
  • Status bar: intent: compare · 11 specialists · attempt 1 · done

One more. "Compare 80 Pioneer Street Brooklyn to 100 Gold Street Manhattan." The planner routes this as a compare intent — two full specialist runs, results merged side by side. The key differences bar surfaces the contrast immediately: Pioneer Street sits 3.81 metres above its nearest drainage channel. Gold Street at 100 is 38.2 metres. Pioneer has 65 empirical flood signals in the record; Gold Street has 26. Same city. Same storm history. Radically different exposure. This is the query a developer, an insurer, or a disclosure attorney actually wants to run.


[SLIDE 8 — What's next] · ~4:00–4:20

SCREEN: Slide 8. Three boxes: Break out the Stones · Other flood-impacted cities · Historical-event mode.

The architecture is NYC-specific by data choice, not by code. The five-Stone pattern generalizes: Houston, Miami, Jakarta — swap the probe sets and RAG corpus, the FSM is the same. Each Stone is already isolated enough to ship as a standalone package. And we want to add historical-event mode: re-run the FSM against snapshot data from before Sandy, before Ida. Validation against measured outcomes as a first-class feature, not an afterthought.


[SLIDE 9 — CTA] · ~4:20–4:30

SCREEN: Slide 9. Dark background. "github.com/msradam/riprap-nyc" large. "Apache-2.0 · public data · AMD MI300X · IBM Granite 4.1 · Mellea grounding."

Everything is open. Apache-2.0, public data, MIT and Apache models. Riprap on AMD MI300X. Try it at the link in the description.


Segment map

Segment Source Timestamp / asset
Slides 1–7 slides/deck.pdf screen-record slide deck
Demo clip 1 — Pioneer briefing + map assets/video/riprap-demo-20260506-234537.webm t≈62–90s
Demo clip 2 — Mellea 4/4 card assets/video/riprap-demo-20260506-234537.webm t≈265–290s
Demo clip 3 — Hollis neighborhood assets/video/riprap-demo-20260506-234537.webm t≈505–545s
Demo clip 4 — Compare result compare-hf.jpg (static screenshot or re-record) n/a
Slides 8–9 slides/deck.pdf screen-record slide deck

Total runtime estimate

~4:30 — comfortable under 5 min with natural pauses.