PulseGrid

Inspiration

Modern cities are reactive by default. When a water main fails during a heatwave, or a power outage hits a campus with a medical facility, operators are flooded with simultaneous alerts — and forced to manually decide what matters most. There's no intelligent layer that says: this one first, here's why, here's what to do. We built PulseGrid to be that layer.

What it does

PulseGrid ingests urban incident signals, normalizes them into a structured format, and ranks them using a deterministic 8-factor weighted scoring model. Each issue gets an AI-generated context enrichment, a full response plan, targeted alert drafts, and a shift-handoff note — all in under 10 seconds

Critical severity classification stays fully deterministic — no hallucination risk. NVIDIA NIM handles only context, explanation, and language generation.

How we built it

The pipeline runs as: Signal → Normalize → Rank → Enrich (AI) → Plan (AI) → Alert (AI) → Handoff

All AI calls use structured prompts with Return ONLY valid JSON constraints and fall back to deterministic outputs if the model fails or the API is unavailable — ensuring the app never breaks during a demo.

Challenges we ran into

Scoring calibration — Getting the 8 weights to produce intuitive results across wildly different incident types (a flooded street vs. a medical urgency vs. a waste buildup) required careful tuning. A log10 population curve prevents a single large-population low-severity issue from dominating the rankings.

AI reliability in a demo context — LLMs don't always return clean JSON. We built a two-pass parseJSONResponse function that attempts full parse first, then falls back to bracket extraction, then to pre-generated fallback data — so the UI never shows a broken state.

Cascading risk modeling — Some incidents look small in isolation but trigger compounding failures. A power cut during a heatwave is more dangerous than either alone. We encode this with an independent cascadingRisk field (0–10) per category rather than deriving it from severity.

What we learned

Deterministic + AI hybrid systems are more robust than pure LLM pipelines for safety-critical decision support

Zod schema validation at the AI output boundary is essential — it catches malformed responses before they reach the UI

Weighted scoring models need careful normalization to avoid any single factor dominating (especially population, which spans orders of magnitude)

Built With

• llama-3
• next.js
• node.js
• react
• tailwindcss
• typescript
• vercel
• zod