SOLPOP

Multimodal AI Inspection Platform for Solar Fleets

RoleSolo build — Founder, Designer, Full-stack & AI Engineer. End-to-end ownership: research, schema, backend pipeline, two-stage AI orchestration, frontend UX, design system, audit infrastructure, deploy.

Built byShaurya Punj

LocationIndia

Year2026

Overview

SOLPOP is a multimodal AI platform that turns a single solar panel photo into an inspector-grade O&M report in under a minute. Drop one image or batch up to 24 — the system auto-detects every distinct module in the frame, grades each against a 21-defect taxonomy, and synthesizes a fleet-level report with severity heatmap, prioritized 7/30/90-day actions, and quantified kWh/kW loss. Built for solar farm operators, asset owners, drone inspectors, and insurance claims adjusters who need fast, evidence-backed condition data without hiring a thermal-imaging crew.

Access on Request

Not publicly deployed.

Book a 20-minute private walkthrough and I will run the whole system end-to-end: architecture, code path, live UI, and the edge cases that shaped it.

Book a private demo

The Problem

Solar fleets quietly lose 5 to 15% of annual output to defects visible from the ground — soiling, microcracks, hotspots, partial shading. Diagnosing it today means hiring a thermal-drone service, waiting weeks, and paying four to five figures per site. Existing analytics platforms charge enterprise contracts and ignore the long tail of installers and asset owners running 100 kW to 10 MW fleets.

The Solution

A two-stage AI pipeline. Gemini 2.5 Flash sees the array — detects every panel, grounds defects with bounding boxes at cell-region precision, and emits strict JSON. Llama 3.3 70B on Groq aggregates that into an executive O&M report with health scores, ranked risks, and quantified energy loss. Results stream as NDJSON so panels light up live. Every AI call is audit-logged with pinned model versions for warranty-grade reproducibility. Free tier with BYO keys, scaling into a multi-tenant SaaS for solar farms.

Project Walkthrough

Agent Architecture

Panel-Gate Classifier

InputNormalized image buffer (≤768px long edge)

EngineGemini 2.5 Flash (JSON mode + fallback chain)

Output{ isSolarPanel, confidence, reason, imageDescription }

Speed: ~1 s per image (preflight, kills hallucinated reports)

Multi-Panel Detector

InputWider drone / array image

EngineGemini 2.5 Flash (grounding mode, NMS dedup)

OutputArray of normalized bboxes [ymin, xmin, ymax, xmax] + confidence

Speed: ~2 s per source image

Per-Panel Vision Inspector

InputSharp-cropped per-panel JPEG

EngineGemini 2.5 Flash (Zod-validated structured output)

OutputPanelAnalysis JSON — type, scores, defects + bboxes, actions

Speed: ~3 s per panel, parallelized at concurrency 2

Fleet Synthesis Analyst

InputArray of PanelAnalysis JSON

EngineLlama 3.3 70B Versatile via Groq (JSON-mode forced)

OutputSystemReport JSON — fleet score, severity dist, risks, recs

Speed: ~3-5 s for fleets up to 24 panels

Inspector Chat (fleet copilot)

InputSession JSON + chat history (last 30 turns)

EngineLlama 3.3 70B Versatile, streaming

OutputMarkdown stream with panel-IDs, severity tags, ROI numbers

Speed: ~250 tok/s, 4 personas × 7 locales

Defect Educator

InputDefect type + severity

EngineLlama 3.3 70B Versatile, streaming

Output4-section markdown (what / why / matters / fix), <180 words

Speed: ~250 tok/s, in-memory cached per defect type

System Architecture

Technology Stack

frontend

Next.js 16 (App Router, Turbopack, React 19)
TypeScript end-to-end
Tailwind v4 + custom industrial-brutalist design system
Framer Motion (springs, useScroll + useSpring parallax, count-ups)
Lenis smooth-scroll
Geist Sans + Geist Mono + Instrument Serif
react-dropzone, react-markdown, react-pdf, lucide-react

backend

Next.js Route Handlers (Node runtime, maxDuration 300 s)
NDJSON streaming (start / progress / detected / panel / synthesizing / report / done / rejected)
Zod schemas validate every model response before state
sharp (image normalize, EXIF-preserved crop, JPEG re-encode)
exifr (GPS, altitude, heading, takenAt, device extract)
In-memory token-bucket rate-limiting (per-IP, swappable to Redis)
@react-pdf/renderer (server-side PDF claim packet)

ai

Gemini 2.5 Flash for vision (panel-gate, multi-panel detect, per-panel) + Gemini 2.0 Flash fallback
Llama 3.3 70B Versatile on Groq for synthesis, chat, explainer
Prompt-version pinning (SHA-256 hash of system + user prompt)
Per-call audit log: { model, version, prompt-hash, payload-hash, latency, status, error }
JSON-mode forced on all structured outputs
Rate-limit-aware retry w/ Retry-After parsing
NMS dedup on overlapping detection bboxes

data

Drizzle ORM (sqlite default via better-sqlite3, Postgres swap on DATABASE_URL)
18-table schema: orgs, users, memberships, sites, strings, modules, inspections, inspection_images, panel_results, defects, reports, audit_log, model_versions, api_keys, usage_events, shares, jobs, webhook_subs
IndexedDB via idb (per-browser session history, thumbnails)
Append-only JSONL audit log (.solpop-audit/<date>.jsonl)
30-day TTL DB-backed shareable links