Hugging Face's logo

Spaces:
lablab-ai-amd-developer-hackathon
/
riprap-nyc
Running

App Files Community
riprap-nyc / RESEARCH.md
seriffic's picture
seriffic
Voice pass: strip em-dashes from user-facing docs
f6423e1
preview code
|
raw
history blame contribute delete
18.2 kB

Riprap: landscape research

Captured 2026-05-06 as part of the AMD x lablab.ai hackathon polish phase. This document underpins the pitch deck (slides/deck.md) and the demo-script choices. Re-validate against the live web before re-using any specific figure.

What Riprap is, distinctly

A citation-grounded LLM that writes audit-quality flood-exposure briefings for NYC addresses by fusing live, historical, modeled, and projected data sources. Mellea rejection sampling refuses to publish a numeric claim it can't cite. The output isn't a score. It's a four-section prose briefing with [doc_id] citations on every numeric assertion, where each doc_id resolves to one specific dataset (Sandy 2012 zone, NYC DEP scenario, USGS HWM, Sentinel-2 chip, NOAA gauge reading, NPCC4 SLR projection).

Granite 4.1 8B drives the prose. Granite Embedding 278M plus GLiNER drive policy-doc retrieval. Prithvi-EO 2.0, TerraMind LULC and Buildings, and Granite TTM r2 drive the EO and forecast probes, with three Apache-2.0 NYC fine-tunes trained on AMD MI300X published on HF Hub.

Architectural commitments other tools don't make:

  1. Silence over confabulation. When a probe returns no data, the briefing omits the section rather than papering over it.
  2. Five-stone epistemic structure. The user can see what's empirical vs modeled vs proxy vs synthetic.
  3. Fully open-source pipeline. Apache-2.0 end-to-end on public- record data, no commercial APIs touched at runtime.
  4. Deployable on either local Ollama or AMD MI300X via vLLM with auto-failover.

Stack as of 2026-05-06: SvelteKit UI on HF Spaces (cpu-basic) at the AMD-hackathon org, FastAPI agent FSM, two-container droplet (vLLM plus riprap-models) on MI300X, full address probe suite at 5/5 PASS in 5.8 to 13.1 s end-to-end.

Landscape map

Direct comps: score-based property risk tools

Tool What it gives Who it serves Hidden cost
First Street Risk Factor (Flood Factor) Score 1 to 10 plus 30-yr risk narrative; powers Redfin, Realtor.com (until Dec 2025 also Zillow) Homebuyers; some lenders Closed model; commercial partnerships; Zillow removed it under industry pressure in Dec 2025
ClimateCheck Score 1 to 100 plus around 30-page property report; 2050 projections Homeowners plus REIT/PE diligence Subscription tiers; methodology behind paywall
Jupiter ClimateScore Global Enterprise SaaS / API; financial metrics (CapEx, OpEx, credit risk) Banks, insurers, asset managers Enterprise pricing; not consumer-facing
Cervest / Climate X / ICEYE Variants of above for ESG / reinsurance Corporate finance and insurance Same

Score-based tools all converge on the same shape: one number, one chart, an explainer paragraph. None show what claim is grounded in which dataset. None expose the audit trail.

NYC-specific government tools

  • FloodHelpNY (City plus State, IDEO-designed). Address lookup to flood-zone label plus insurance plus free resiliency audit. Forms-based, consumer-facing, doesn't fuse live signals.
  • NYC Flood Hazard Mapper. ArcGIS web map of FEMA, NPCC, Sandy, and future scenarios. Static visualization, no narrative.
  • NYC OEM Flood Maps page. Index of the above.
  • EJNYC Flood Vulnerability Index (released 2024-04 by Mayor's Office of Climate and EJ). First-ever city FVI, used to direct spending under NY's "Disadvantaged Communities" framework (35% of climate spend by law).
  • FloodNet NYC (NYU plus CUNY plus city). Over 350 ultrasonic sensors at 1-min cadence, growing to 500 by end-2026. Has a public dashboard but no narrative layer.

Federal / authoritative

  • FEMA Flood Map Service Center / NFHL. Official; covers 90%+ of population; static GIS layer plus PDFs. The disclosure-of- record but not a synthesis tool.

Real-estate platforms (the volatile zone)

  • Redfin. Still shows First Street Flood Factor on every listing.
  • Realtor.com. Still shows it on 110M+ listings.
  • Zillow. Removed climate risk display in December 2025 under California Regional MLS pressure. Still links out, but it's hidden. This created a vacuum that an open citation-grounded alternative could fill.

Closest academic / AI comps

  • Flood-LLM (Brisbane, MDPI Sustainability 2026). Multi-source LLM for property-level flood risk, validated on Brisbane against official labels. Academic, not deployed; no Mellea-style citation discipline; no live signals.
  • GIS-Integrated Flood LLM (Tandfonline 2024). LLM constrained by a flood knowledge graph plus GIS interaction. Research artefact.
  • FloodLense (arXiv 2024). UNet/RDN/ViT plus LLM for satellite flood detection. Research; image-only.

Where Riprap fits: differentiators that demo well

Ranked by visibility in a 3-minute demo:

  1. Citation prose vs scores. Riprap returns "Hurricane Sandy flooded this address on October 29 to 30, 2012, according to the empirical inundation zone [sandy]. 19 flood-related 311 service requests were logged within 200 m over five years [nyc311]." Every number cites a doc; each doc resolves to a footer source row. First Street returns "Flood Factor 8/10". This gap is the demo.
  2. Live, historical, modeled, projected: in one paragraph. Sandy 2012 (empirical), DEP 2080 stormwater scenarios (modeled), 311 last 5 years (proxy), FloodNet last 3 years (empirical, hyperlocal), NPCC4 SLR (projected), Granite TTM r2 surge nowcast (96-h forecast). No comp combines all four temporal modes.
  3. Open-source NYC fine-tunes. Three Apache-2.0 models (Prithvi-EO-2.0-NYC-Pluvial, TerraMind-NYC-Adapters, Granite-TTM-r2-Battery-Surge) trained on AMD MI300X. Anyone can reproduce, fork to other cities, or audit. First Street's model is closed; ClimateCheck's methodology is behind a paywall.
  4. AMD hardware story. The whole stack runs on MI300X via vLLM (LLM) plus a sibling ROCm container (probes). All Apache-2.0. This is the AMD hackathon track's preferred narrative: open models, open infra, open data, real GPU acceleration.
  5. Mellea grounding receipts. The four checks (numerics_grounded, no_placeholder_tokens, citations_dense, citations_resolve) are the audit. The meta card surfaces "4/4 grounding checks passed, 1 reroll". That's audit credibility no consumer comp shows.
  6. Self-aware silence. Touchstone shows "FloodNet sensor: 0 events in 3 years" with silent_by_design. Lodestone shows "TTM Battery surge forecast: peak |residual| < 0.3 m, omitted." Most tools always render a value. Riprap's silence is a feature.

Stakeholder demos to craft

Six concrete personas, each with a query that exercises a different part of the system. These are the demo arcs to rehearse.

1. Resident / homebuyer (the FloodHelpNY swap-in)

"I'm thinking about renting an apartment at 80 Pioneer Street, Brooklyn. Should I worry?"

Demo arc. Type the address. Watch the planner classify single_address, then 19 step events fire across the four data Stones in around 13 s. Briefing names Sandy 2012 inundation, 65 311 complaints, 2 FloodNet sensors with 4 events including a 51 mm peak on a specific date, Ida 2021 HWM 130 m away, microtopo HAND 3.81 m plus TWI 14.79 (very high saturation propensity). Footer shows 7+ named primary sources.

Demo hook. "Compare what we just generated to First Street's number-and-bar-chart for the same address. Which would you trust to make a $4,000/month decision?"

2. Real-estate attorney / disclosure compliance

"Does 100 Gold Street, Manhattan need to disclose flood risk under RPL §462(2)?"

Demo arc. Same single_address path. Briefing produces a citable narrative covering FEMA designation, prior flood claims (where present), terrain, recent complaints. Mellea grounding check is the qualifier: "this prose is grounded against four invariants and passed 4/4."

Demo hook. New York's March-2024 amended Property Condition Disclosure Statement requires sellers to disclose flood history and FEMA-floodplain status. RPL §231-b requires every residential lease to disclose prior flood damage. Riprap is the citable narrative tool. Show how the briefing maps line-by-line to the disclosure requirements.

3. NYC OEM / DEP planner

"Hollis, Queens"

Demo arc. Neighborhood intent fires (9 step events), produces an NTA-level briefing. 434 flood-related 311 over 3 years (87 catch- basin clogged, 42 street-flooding), 4.3% of neighborhood projected to flood under DEP moderate-2050 scenario, 25% of cells with HAND<1 m. RAG retrieval pulls relevant DEP/NPCC4 policy paragraphs.

Demo hook. DEP just announced a $30B stormwater priority list (86 locations) and a $68M Brooklyn Bluebelt expansion in Prospect Park. Riprap supports the prioritization argument with citable per- NTA evidence. Pair with the EJNYC Flood Vulnerability Index for the EJ-spending overlay (35%-to-disadvantaged-communities legal mandate).

4. Insurance underwriter / actuary

"442 East Houston Street, Manhattan"

Demo arc. Same as resident demo, but emphasize the provenance trace UI. Every Stone row, every doc_id, every source URL, vintage, and tier glyph.

Demo hook. When an underwriter writes a risk memo, the audit chain matters. First Street's "we used a proprietary catastrophe model" doesn't survive a regulator review the way "we used FEMA Sandy 2012 polygon, NYC DEP 2021 stormwater scenario, USGS Ida HWM Event 312, NOAA gauge 8518750, NWS station KNYC, Granite TTM r2 fine-tune (test MAE 0.1091 m vs 0.1467 zero-shot, citable)" does.

5. Climate journalist / advocacy

"Coney Island, Brooklyn"

Demo arc. Neighborhood briefing. 87.5% of NTA in 2012 Sandy zone, 382 flood complaints over 3 years, 7.8% projected flooded under 2050 moderate, 38.9% of DEM cells with HAND<1 m, DEP extreme- 2080 at 44.2% flooded.

Demo hook. ProPublica/NYTimes/THE CITY-style data journalism. Every claim in a Riprap briefing is reproducible. Anyone can paste the same query and get a near-identical narrative. The journalist can publish the briefing as the methods section.

6. Architect / developer

"What are they building in Gowanus and is it risky"

Demo arc. Planner classifies development_check. FSM pulls DOB filings plus flood layers for the project sites. Briefing comments on which proposed buildings sit inside Sandy 2012, which intersect DEP extreme-2080, what the microtopo says.

Demo hook. Pre-design siting check. The Gowanus rezoning is one of NYC's largest active development zones, well known to flood. Show how the tool surfaces flood concerns before architects pour concrete.

Lateral and unexpected use cases

Ten bets, ordered roughly from most-buildable to most-speculative.

  1. Pre-storm cohort briefings. Subscribe Riprap to NWS flood- watch alerts. When a watch lands, fan out one briefing per affected NTA plus push to OEM, press, and advocacy lists. Citable evidence on demand for the press cycle that follows.
  2. Climate-grant evidence sections. HUD CDBG-DR and FEMA BRIC applications need an auditable evidence base. Riprap auto- generates the "vulnerability assessment" section so a community group can apply for resilience funding without hiring a consultant.
  3. Local Law disclosure boilerplate. Plug into a brokerage's listing flow. When an agent enters an address, auto-generate the NY RPL §231-b lease addendum or §462(2) disclosure draft. ROI is high since the law took effect 2024 and many landlords are still figuring out compliance.
  4. MTA station-hardening prioritization. Riprap already has the MTA-entrance probe (KEY-001 in the demo). Run the FSM across all subway entrances; rank by exposure × ridership. The MTA's October-2025 Climate Resilience Roadmap Update is the policy hook.
  5. DOE school siting. When DOE reviews proposed school sites or selects schools for retrofit, Riprap briefings (with expect_311_ge plus Sandy plus DEP overlays) would catch flood exposure that form-style screens miss.
  6. Time-machine variant. Re-run the FSM with snapshot data from a past date. "What would Riprap have said about Hollis on August 31, 2021, the day before Ida?" Useful for retrospective analysis, expert testimony, and stress-testing the system.
  7. Cross-city scaffold. The architecture is NYC-specific by data choice, not by code. Port to Houston (post-Harvey plus Hurricane Beryl 2024), Miami (king tides), Boston (CSO floods), Charleston (chronic tidal), with a per-city probe set plus RAG corpus.
  8. Federation with FloodNet alerts. When a sensor triggers a flood event NOW, fire a Riprap live_now briefing for the surrounding NTA: "what's at stake in the next 6 hours." Connects FloodNet's hyperlocal sensor reads to the OEM decision loop.
  9. EJNYC × Riprap pairing. Rank all 188 NTAs by Riprap-detected exposure, intersect with state DAC designations. Output: a map of "underserved plus underwater". The most underfunded high-exposure neighborhoods.
  10. Court testimony / expert witness. Citable, reproducible flood narrative as a court exhibit. The Mellea passes-record plus provenance trace are the kind of artefact a regulator or judge can audit. Especially relevant after the December-2025 Zillow controversy created public discussion of climate-data integrity.

Risks and framing

  • Real-estate industry pushback. December 2025: Zillow removed First Street's climate scores under MLS pressure because the data was hurting transaction volume. A free, citation-grounded alternative could face the same reflex. Riprap's defence is that it's a narrative tool for professional analytical work, not a buy/don't-buy verdict. Keep the disclaimer footer prominent.
  • Redlining hazard. Exposure narratives can be misused by landlords or insurers to discriminate against high-flood-risk (often disproportionately disadvantaged) neighborhoods. Mitigations: (a) the citation transparency makes biased reasoning auditable, (b) the EJNYC pairing in lateral-use #9 reframes exposure data as a tool for affected communities, not against them, (c) keep "for professional analytical work, not personal property decisions" front and center.
  • Disclosure-status liability. A briefing is evidence but probably not the §462(2) disclosure under New York real-estate law. Don't position it as legal disclosure-of-record without a real-estate-attorney review.
  • Cold-start latency. First query after droplet redeploy is around 30 s while models warm. For demos, ping the Space and run one warm-up query 5 minutes before showtime.
  • Geocoder edge cases. "PS 188, Lower East Side" geocoded to a Brooklyn PS 188 in our test suite. For demos, pick fully-qualified street addresses; document the disambiguation behavior.

Polish punch-list (deck-driven)

Concrete polish items the research surfaces, ranked by demo value:

  1. Sample-query pills on landing. Six clickable pills below the search bar, one per persona above. Let the audience demo themselves.
  2. A "What this is" bar at the top of the landing. Three lines: "Citation-grounded NYC flood briefings. Every number cites a primary source. Open-source, public data, audit-grade synthesis."
  3. Compare-mode link from the briefing. Once Riprap delivers a single_address briefing, surface "compare with another address" as a one-click affordance. The compare intent already exists in the planner.
  4. EJNYC-FVI overlay on the map sidebar (#9 above). Riprap's exposure × DAC designation, two clicks to a powerful editorial demo.
  5. First-query warm-up message during the cold start: "loading probes on AMD MI300X. First query after redeploy takes around 30 s; subsequent queries 5 to 13 s."

Sources