Skip to content

← back

Lovely Interiors

AI paint consultation with computer vision

GitHub →

What it does

  • Takes a photo of your room and reads the lighting nuances, existing finishes, and undertones using Gemini 2.5 Pro vision.
  • Runs a structured consultation that tracks your requirements room by room so recommendations compound across a whole house, not just a single wall.
  • Searches a specialized 300-color paint catalog with technical specs, matching undertones and design-harmony rules rather than keyword similarity.
  • Validates each recommendation against LRV (light reflectance value) targets and basic design principles before it shows up in the shortlist.
  • Optionally hands off to a browser-automation agent that orders the physical sample swatches from the retailer's site, with retry and cart-state recovery.

Why I built it

Consider this my LinkedIn comeback entry. I have always been fascinated by interior design, even if accounting became my career, but decision fatigue around paint palettes kept me from bringing that passion into my own home. Instead of using a generic ChatGPT or Gemini prompt, I built something purpose-built: a specialized color database, a structured consultation that tracks my requirements, a room-by-room paint plan, and automated sample ordering. Sometimes the best way to learn new tech is to solve your own problem.

How it works

Room photoinputPhoto analysisgemini visionColor matchingsemantic searchLRV validationdesign rulesCoordinatorgoogle adkRecsfarrow & ballBrowserorder samples
Four specialized agents collaborate through a coordinator. The browser agent actually orders the physical paint samples for you.

What was hard

  • Lighting reads differently in every photo and no single prompt handles it all. The photo-analysis agent had to be narrowed to "read lighting and undertones" instead of "describe the room," otherwise it hallucinates materials it cannot actually see.
  • Premium paint catalogs name colors evocatively, not descriptively (think dusty lavender or coastal fog, not a hex code). Semantic search on the name alone mismatches more than it matches. The color agent falls back to LRV and hue when name similarity is misleading.
  • Browser automation against a real retail site is fragile. Error recovery (retries, re-login, cart state, out-of-stock handling) turned out to be more work than the recommendation logic itself.
  • The general pattern that came out of this: agents are the right shape when each agent has a different success criterion. Photo analysis succeeds when it reads lighting correctly. Color matching succeeds when it returns harmonious options. Ordering succeeds when the sample arrives. Forcing all three into one agent just produces a confused generalist.

Outcome

  • A shortlist of paint colors matched to your photo, with per-color rationale that references the lighting and undertones the vision agent saw.
  • LRV + design-rule validation per recommendation so you can see why a color made the list.
  • A room-by-room paint plan when you consult on multiple spaces.
  • Physical samples ordered to your door, optional.
  • Professional handoff export for a contractor. Not a consumer toy, structured enough to hand to a pro.

Tech

Google ADK · Gemini 2.5 Pro · Browser-Use · Computer vision · Python