Turn a site walk into a trusted project record.
SiteThread is a phone-first construction field agent that turns ordinary site walkthrough video, imagery, and spoken observations into structured, evidence-backed project records.
It helps construction professionals move from raw field media to reviewable findings, action items, and site reports without adding a heavy capture workflow.
Construction teams already capture photos, videos, and spoken notes during site walks. The difficult work comes afterward: finding what matters, connecting observations to locations and timestamps, and turning scattered media into a reliable project record.
SiteThread prepares that record while keeping the source evidence attached to each finding.
Site walkthrough
↓
Media processing
↓
Transcript + visual evidence
↓
Structured observations
↓
Confirm / Edit / Dismiss
↓
Evidence-backed site report
Users record or upload a walkthrough and narrate what they see. SiteThread combines the transcript with relevant visual evidence to prepare structured observations. Each observation can be confirmed, edited, or dismissed by a construction professional. Only confirmed or edited information becomes part of the trusted report.
AI prepares. Human confirms.
Findings remain traceable to their source where available, including transcript segments, timestamps, frames, and clips. SiteThread supports professional review; it does not independently certify safety, compliance, engineering acceptance, or completion percentages.
Livepeer's official creative MCP is used for bounded transcription. SiteThread creates six-second FFmpeg windows, gives the creative transcribe tool a short-lived private R2 URL, and retains its own source-window bounds because provider timing may be absent. Google Gemini provides bounded visual semantics. Structured observation reasoning is selected independently through the SiteThread-owned ObservationReasoner interface; the hackathon config uses Groq with strict JSON Schema output. The reasoner sees only labeled SiteThread evidence, validates results again with the existing Zod schema, and rejects references outside that evidence bundle. Provider provenance records the configured and served model when available; there is no raw Livepeer gemini-text route or provider/model fallback.
See the Livepeer integration and technical spike for the detailed integration decisions, provider boundaries, capability validation, and processing strategy.
SiteThread uses a Next.js application for the product experience and API boundaries, durable background jobs for long-running media processing, private object storage for media, and structured PostgreSQL data for project records. FFmpeg handles deterministic media operations; Livepeer handles media intelligence.
See the architecture and engineering plan for system boundaries, data flow, repository structure, testing, and implementation decisions.
| Area | Technology |
|---|---|
| Application | TypeScript, Node.js, Next.js, React |
| UI | Tailwind CSS, shadcn/ui |
| Validation | Zod |
| Database | Neon PostgreSQL, Prisma |
| Media storage | Cloudflare R2 |
| Background processing | Trigger.dev |
| Media intelligence | Livepeer Agent, Google Gemini |
| Observation reasoning | Provider-neutral ObservationReasoner (Groq adapter configured for the hackathon) |
| Media utilities | FFmpeg |
| Hosting | Vercel |
| Testing | Vitest, Playwright |
| CI | GitHub Actions |
- Golden Path & Demo Acceptance Criteria (COD-13)
- Architecture & Engineering Plan
- Livepeer Integration & Technical Spike
- Repository Agent Guidance
The repository uses the following branch model:
dev— protected integration branchmain— protected release branch- short-lived feature branches — implementation work
Changes should be made on a feature branch and submitted through a pull request into dev. The application is still being built; the detailed architecture and integration documents describe the current engineering direction.