Estagent — AI Property Tours with Spatial Context
2nd Prize, Sea x OpenAI Codex Hackathon Vietnam 2026 | 12-hour build, three-developer team
Estagent helps people understand a property before deciding to visit. Creators turn photos, panoramas, property details, and an optional floorplan into a reviewed tour; visitors explore a 3D dollhouse and panorama checkpoints, listen to guided narration, and ask questions about the room they are viewing. Built in a 12-hour hackathon sprint with OpenAI, LangGraph, Blender, Three.js, and FastAPI.
Timeline
2026
Type
Competition
Status
completed
My work
- •Led the web lane and acted as code integrator for the three-developer team: reviewed, merged, and released teammate branches to main (116 of 156 commits in the public repository)
- •Built the landing page, Creator photo-tour flow, and immersive Viewer with the Three.js dollhouse, panorama checkpoints, and guided narration playback
- •Integrated contextual tour chat, bilingual narration with TTS, and native FFmpeg video export into the release candidate
- •Ran the sprint with contract-first lanes and parallel AI coding agents (Codex workers in isolated Git worktrees), plus a declaration of reused sources and work built during the event
Outcome / Impact
- •Won 2nd Prize at Sea x OpenAI Codex Hackathon Vietnam 2026
- •Shipped a live end-to-end demo: capture upload, AI scene planning with human layout review, Blender-built model, narrated tour, and published-tour Q&A
- •Kept the Creator in control through explicit layout and content reviews, confirmed edit proposals, and immutable published versions
- •Grounded visitor answers in approved tour data and the active checkpoint, with navigation targets validated by the backend
Tech / Skills
Project Media
Selected website screens and product walkthrough.
Case Study
1) Context / Problem
Listing photos rarely explain how rooms connect or what a space feels like, so renters spend time on visits that do not fit. Estagent turns existing captures into a tour that visitors can explore and question before they decide to visit.
2) Your Role
As Web Lead and Code Integrator, I owned the Creator and Viewer web experience, integrated teammate branches into main, and kept the live demo releasable throughout the 12-hour sprint.
3) Approach
OpenAI organizes the captures into a structured scene plan; the Creator clarifies and approves the layout before a local Blender worker builds a textured single-floor model, which Three.js renders as a dollhouse with capture markers and panorama checkpoints. A checkpointed LangGraph workflow, request deduplication, and version history preserve progress across interruptions. The geometry is approximate authored geometry, not metric photogrammetry.
4) Result / Impact
Won 2nd Prize at Sea x OpenAI Codex Hackathon Vietnam 2026 with a deployed demo covering tour creation, guided narration, contextual visitor chat, and shareable published tours.
5) Learnings
In a short sprint, frozen contracts and a single integrator let several developers and AI coding agents work in parallel without breaking main. Showing the human review steps made the AI output easier to trust.
6) Links
See links above.