Ross Adams

Helios

AddressMemoGeocodeAddress to lat/lngNY boundsReject out-of-stateStopNo paid calls runParcelNY GIS parcel polygonGrid proximityHIFLD lines, substationsEnvironmentalFEMA · NWI · PAD-USTerrainUSGS 3DEP slopeOrdinanceTown code via ClaudeInterconnectionNYISO queue (proxy)SynthesizeScore + cited memoAddressMemoGeocodeAddress to lat/lngNY boundsReject out-of-stateStopNo paid calls runParcelNY GIS parcel polygonGrid proximityHIFLD lines, substationsEnvironmentalFEMA · NWI · PAD-USTerrainUSGS 3DEP slopeOrdinanceTown code via ClaudeInterconnectionNYISO queue (proxy)SynthesizeScore + cited memo
Figure 1. Helios functional block diagram: one New York address in, a cited viability memo out. The four checks after Parcel run concurrently.

Overview

An AI agent for early-stage site diligence on utility-scale solar projects in New York State. Given one address, it resolves the parcel, runs siting, interconnection, and permitting checks against public GIS data and town ordinances, and writes a viability memo with a 0–100 score, a citation for every flag, and an interactive map.

Returns the memo in about 90 seconds, work that takes a junior analyst 30–60 minutes per site.

Design decisions

LLM scoped to judgment calls
Claude handles two tasks: extracting town solar ordinances and scoring their restrictiveness. Both run through forced tool-call schemas. The score is clamped to 10 of 100 points and defaults to 0 on failure. A moratorium is reported only with a section reference and a verbatim quote. Distances, overlaps, slope and every other weight are deterministic, unit-tested code.
LangGraph orchestration
LangGraph was chosen for the branching that ordinance research needed: a nested sub-graph tries eCode360, Municode, the town site, then NYSERDA, and stops at the first hit. Parallel branches write disjoint state keys, so concurrent updates cannot collide.
Static layers in local PostGIS
Flood zones, wetlands, protected lands, transmission lines and the 10 m 3DEP elevation model are preloaded into PostGIS. The data changes on multi-year cycles; live fetching would add network failure points and raster dependencies without improving freshness.
Public access without auth
Runs are unauthenticated and cost API money. A 20-runs/hour per-IP limit stops scripted abuse; a global cap of 5 concurrent runs bounds spend regardless of IP sharing. CAPTCHA was rejected as friction.
Streaming progress over SSE
Runs take about 90s. Each LangGraph node emits a Server-Sent Event on completion, so parallel checks render as they finish. WebSockets were rejected: the channel is one-way.

Stack

Table 1. Helios stack
AgentLangGraph state graph, with a nested sub-graph for ordinance research
LLMClaude Sonnet 4.6: web search for ordinances, structured extraction, restrictiveness score
APIFastAPI, Python 3.14, Server-Sent Events
DataPostgreSQL 16 + PostGIS: HIFLD, FEMA NFHL, NWI, PAD-US, USGS 3DEP, NYISO queue
FrontendReact, Vite, TypeScript; Folium map; Mapbox address search
Quality161 pytest unit tests and a Docker build on every push, through GitHub Actions
InfraOne Docker image on Fly.io, scale-to-zero

Status and links

Status Live. One New York address per run. Scales to zero when idle; this page prewarms it on load.