Case Study · Product Design · AI Engineering · AI Builders Challenge with IBM Bob · 2026

Sky to Porch

Real NASA and NOAA observations about hazards — fire, flood, heat, drought, air, volcanoes — explained in plain English by IBM watsonx Granite, and engineered so the AI cannot lie about them.

IBM watsonx Granite · 16 Live Data Sources · Evidence Guardrails · Next.js · Live Product

Live app GitHub repo Watch the demo
Sky to Porch banner: the tagline Space data, plain English above six colored hazard chips and the satellite-over-house logo
Problem
Space agencies publish extraordinary data about our planet, but it is published for scientists. A parent asking "was it dangerously hot at practice yesterday?" meets raw products and jargon — or a chatbot that answers confidently with no data at all.
Built
A live web app covering six hazards with 16 integrated NASA, NOAA, USGS and EPA sources. Every observation is validated before it is shown; IBM watsonx Granite writes the plain-English answer, and guardrails reject anything the evidence does not support.
My Role
Product owner, UX designer, and quality gate. I directed four AI collaborators — ChatGPT, Codex, IBM Bob, and Claude Code — through 55 recorded architecture decisions, and owned every accept or reject call along the way.
6Hazard lines, all live
16Integrated public data sources
1,780Automated tests
0Claims without evidence

DownlinkThe Demo

From orbit to your front porch in three minutes.

The demo walks the whole journey — picking a real event, watching the evidence pipeline retrieve and validate, and reading a Granite answer that traces back to real observations.

Or skip the video and ask it something yourself — it is live.

35,786 kmThe Data

The data exists. Ordinary people cannot use it.

Fire detections, flood imagery, surface temperature, air quality — NASA, NOAA, USGS and EPA publish all of it, every day. But it is published for scientists, not for the people standing underneath it: raw scientific products, coordinate systems, jargon. People without a science background have no clear way to use any of it.

Their questions are concrete. After a storm damages your roof, an insurance claim goes better with real records of what happened at your address. On a very hot day, you want to know whether it is safe to walk your dog. Before a trip, you want to know what the air was like where you are going.

Asking a general chatbot gets an answer — a confident one, with no data behind it. And with hazards, a made-up answer is dangerous in both directions: a false alarm causes panic, and false comfort leaves people unprepared. That asymmetry became the design brief.

705 kmWhat I Built

Pick a place, a hazard, and what you care about.

Sky to Porch asks for three things: where, what hazard, and what matters to you — home, health, pets, travel, power, community. No GIS skills, no API keys, no jargon. Six one-tap demo stories cover real events, from Hurricane Beryl in Houston to the 2023 smoke episode in New York.

Sky to Porch landing page in dark theme: six real-event story cards beside an interactive map and the insight panel
Start from a real event — one tap sets the place, hazard, and dates.
The insight panel mid-query, showing the staged evidence pipeline: checking coverage, fetching observations, validating evidence, writing the answer
The wait is honest too: the pipeline shows its real stages while it works.
Hurricane Beryl query over Houston: NASA IMERG rain imagery on the map, and a Granite-written answer noting rain was observed but no flood extent was detected
Hurricane Beryl over Houston — rain imagery on the map, and an answer honest enough to say what was not detected.
Sky to Porch on a phone-sized screen: the story cards stacked in a single column above the Ask, Map, and Insight tab bar
The same evidence, one-handed — a bottom tab bar replaces the three-column layout.

Every answer carries its provenance on its sleeve: a badge names the model that wrote it, every number traces to a validated observation, and each source links the official checker where you can verify it yourself.

400 kmMission Control

Directing four AI collaborators through 55 recorded decisions.

Built for the AI Builders Challenge with IBM Bob, this was my largest AI-directed build yet — and the most disciplined. Every architecture decision became a numbered record (ADR-0001 through ADR-0055), so nothing rested on "the model said so".

  1. 01

    Brainstorm & research — ChatGPT

    Explored the problem space, pressure-tested the concept, and helped shape the six-hazard scope before any code existed.

  2. 02

    Architect & build — IBM Bob

    The primary development tool: dataset registry, all six evidence adapters with fail-closed transports, the interpretation layer, AI router and guardrails, the query flow and map.

  3. 03

    Review & verification — Codex

    Code review, prompt writing, part of the verification work, and the stage closeouts that kept 14 work packages honestly finished rather than merely merged.

  4. 04

    Final review, security & deployment — Claude Code

    Final-phase project review and polish, the security hardening of the paid AI endpoints, and the verified production deployment on Vercel.

My constant role across all four stages: product owner, designer, and quality gate — the concept, the UX direction, the honesty rules, and every accept or reject decision were mine.

100 kmThe Kármán Line

The line the AI is never allowed to cross.

A single LLM asked about hazards will confidently invent numbers. Sky to Porch draws a hard line: deterministic code interprets the evidence, and the AI only phrases what has already been proven.

The evidence pipeline — every stage is allowed to refuse

Stage 1 · Retrieve

Real observations only

NASA · NOAA · USGS · EPA · server-side

Satellite imagery and ground-station records for the exact area and date. Every source carries a machine-readable coverage profile; a timeout or size overrun ends as a stated failure, never a quiet default.

Stage 2 · Validate & interpret

Deterministic conclusions

Plain TypeScript under unit test

Validated numbers become structured conclusions — a station's peak heat index becomes an official NWS severity category. The same evidence always produces the same conclusion.

Stage 3 · Explain

Granite writes the answer

IBM watsonx · ibm/granite-4-h-small

Granite receives conclusions to phrase, not raw data to interpret — which is what keeps the numbers real and the tone human. A labeled OpenAI fallback covers provider outages.

Stage 4 · Guardrails

The output is read back

Per-fragment rejection

Any number not in the evidence, any invented URL, any unsupported claim is rejected — the offending sentence is replaced by deterministic text and the reason is logged. Safe prose survives.

Two fallbacks, two reasons: provider fallback (watsonx to OpenAI) is availability insurance; deterministic fallback — a real answer with real numbers, no AI at all — is honesty insurance.

Claim separation: what a source shows is stored apart from what it proves. A smoke polygon is smoke analysis, not the air in your lungs — and "no observation" is never rendered as "no danger".

What I chose — and what I rejected

Honest degradation over fake coverage

Canadian and Mexican sources are registered and gated, shown as prepared — not quietly claimed as live. When a source fails, the UI says so and shows what remains.

Two kinds of evidence, never merged

Satellites cover everywhere and prove little about one street; ground stations prove a lot about one point and cover almost nowhere. Showing both, clearly labeled, is the honest option.

The waiting time tells the truth

A live query takes ten to twenty seconds. Instead of a spinner, the panel shows the pipeline's real stages — the wait itself becomes a proof of work.

Regional truth over false precision

One station or one satellite pixel is never offered as proof about one building, person, or pet. Answers name their real scale, and the analysis radius is always visible.

10 mThe Guardrails

A public app spending paid AI calls is an abuse target.

The last two architecture decisions before launch were pure defense. An audit of all twelve API routes found one unguarded paid endpoint — it was deleted, not patched. Then the rate limiting moved off the server's memory entirely:

  • Distributed rate limiting on Upstash Redis — serverless instances each forget their own counters, so the counters live in shared storage: burst, sustained, daily, and a global daily budget across all users.
  • Privacy-minimal identity — rate-limit keys are HMAC digests, never IP addresses or message content, and every key expires on its own TTL.
  • Origin gate — only the app's own domain may spend the paid endpoints; other websites' front-ends are refused before any money moves.
  • Fail degraded, never open — if Redis cannot answer, the app refuses the paid call and answers deterministically. No verdict is not permission to spend.

Deployment day proved the design: the first production build went live before its environment variables existed — and the app did exactly what it was built to do: served real evidence, spent nothing, and said honestly that AI was not configured.

1 mField Notes

How I used AI in this project

This project was built in collaboration with AI; the design decisions, content, and direction are mine. I did not write it line by line — I directed four tools through distinct stages and owned the quality bar throughout.

  • AI did — concept sparring and research (ChatGPT); architecture, adapters, guardrails, and tests (IBM Bob); code review, prompt writing, verification passes, and stage closeouts (Codex); final review, security hardening, and the production deployment (Claude Code).
  • I did — the product concept, the six-hazard scope, the honesty rules, the UX and visual direction, the demo-story selection, verification and acceptance of every merged PR, credentials and spending controls, and the contest submission.
  • What I learned — the guardrails caught real failures in production: Granite output was rejected mid-demo and the deterministic fallback answered instead, invisibly to the user. Watching a safety design work for real — rather than in a test — changed how I think about "done".

0 mThe Porch

Live, verified, and honest about what comes next.

Sky to Porch shipped as a deployed, rate-limited, security-verified product before the AI Builders Challenge deadline — 1,780 automated tests, six live hazard lines, and a browser-verified production rollout.

Next: coverage beyond the United States, and more languages — starting with Spanish and Traditional Chinese, because hazard information is hardest to reach exactly where English is not the first language. The mission fits on its own tagline: real data, honest answers, for the people who live under this sky.

Ask it about your sky.

Try the live app, watch the demo, or read the code.

Live app GitHub repo Video demo