KaguaJua kabla ya kununua — know before you buy.

The problem: a market where quality is invisible

Most furniture in Kenya is made and sold in the informal jua kali sector — roadside workshops and open-air showrooms where a bed, sofa set, or dining table is a major household purchase, often several weeks' income.

At the point of sale, a buyer cannot tell well-seasoned hardwood with tight mortise-and-tenon joints from green timber butt-nailed together and hidden under filler and varnish. The failures arrive months later: joints loosen, panels split as the wood dries, frames rack and wobble. Because buyers can't observe quality, they won't pay for it — so price becomes the only signal, careful fundis can't charge for craft, and the whole market drifts toward the lemon. This is the classic information-asymmetry failure, and it hurts both sides: buyers waste scarce income on furniture that fails, and skilled producers are trapped at commodity prices.

Why this is solvable now

The missing ingredient has always been cheap expertise at the moment of purchase. A master fundi could spot every one of these defects in two minutes — but you can't bring one shopping. A vision-capable Claude agent can now do a meaningful fraction of that inspection through a phone camera, plus a few physical tests the buyer performs with their own hands. Expertise that used to cost more than the furniture drops to a few shillings of inference.

Who it's for, and the conditions on site

Primary persona

Amina, 34, Eastleigh, Nairobi. Just moved apartments; needs a bed, a dining set, and a sofa over the next three months. Shops at roadside showrooms and Gikomba. Was burned before — a sofa that sagged within a season. She has a KSh 25,000 budget, a mid-range Android phone, and a healthy suspicion of sellers. She doesn't know joinery; she knows disappointment.

Secondary (later): diaspora buyers furnishing family homes remotely — they send a relative to the shop and want the inspection done by someone they trust. High willingness to pay; defer until the core loop works.

Site conditions — design constraints we treat as fixed

Breaking the problem down

The naive version — "upload a photo, get a quality score" — fails. One photo of a chair tells you almost nothing about the joint you can't see. The design insight that makes Kagua work is that the phone is a sensor the agent can direct, and the buyer's hands are instruments the agent can borrow.

The guided protocol is the product

The agent runs an item-specific inspection protocol: it asks for particular shots from particular angles, coaches the user live ("closer on that joint — right where the leg meets the seat"), requests extra evidence when it spots something, and directs physical tests a buyer can perform in a shop without permission or tools:

• Wobble test — press diagonally opposite corners; racking reveals loose joints.
• Marble / bottle-top roll — set it on the surface; drift reveals an unlevel top.
• Sight along the edge — phone camera held low along a surface catches warp and twist.
• Drawer pull — smooth travel and a look at the drawer's corner joints (dovetail vs staple).
• Fingernail press, underside — soft, dent-prone wood or suspiciously heavy "wet" timber suggests green (unseasoned) stock.
• Sit-and-shift — creak and flex under real load, where sitting is expected.

Steps 3–4 loop: the agent re-requests evidence until confidence clears a threshold — or says plainly what it couldn't verify.

Defect classes mapped to the best available sensor

Not everything is visible, and pretending otherwise would destroy trust. We split the defect space by how each class can actually be detected, and let the report's confidence language follow the sensor, not our marketing:

Key questions we had to answer

The product: basic and extended

The interactive mockup walks every screen below with a real worked example (a KSh 3,500 kitchen stool at a roadside shop). Basic is the complete, self-sufficient loop; Extended is the network layer that switches on once assessment volume exists.

The report is the hero screen

Everything funnels into one screen the buyer can act on in the shop, structured as: headline verdict (Sound / Fair — priced-in issues / Serious concerns) → defect cards (annotated photo crop, what it is, what it means in use, cost to fix, what to ask the seller) → what we couldn't verify → price sanity check against the quoted price. The follow-up chat sits directly under the verdict, pre-seeded with the three questions buyers actually ask.

Agent design

Inspection protocols as skills

Each item type ships as a protocol file the agent loads on selection: shot list with framing guides, item-specific physical tests, defect taxonomy with severity rubric, and typical failure patterns for that type as made in Kenyan workshops (e.g., bed frames: under-sized side-rail bolts; sofas: webbing instead of springs — fine — vs cardboard under fabric — not fine). Protocols are data, not code: the fundi panel from Q1 can review and improve them without an engineer.

Structured output, native UI

The agent's verdict is a typed object, not prose — the app renders it as cards, and follow-up chat stays grounded in it:

{
  "verdict": "fair",            // sound | fair | serious_concerns
  "headline": "Solid frame, two issues to price in",
  "confidence_notes": ["Could not verify timber seasoning from photos"],
  "defects": [{
    "type": "joint_gap", "location": "rear stretcher, left tenon",
    "evidence": "photo_5_crop", "severity": "structural_medium",
    "implication": "Will loosen within 6–12 months of daily use",
    "fix": {"what": "Re-glue and clamp", "cost_kes": [100, 200]},
    "ask_seller": "Ask the fundi to re-glue it before you pay."
  }],
  "price_check": {"quoted_kes": 3500, "typical_range_kes": [2800, 3800],
                  "basis": "grade + size, Nairobi roadside typical"}
}

Calibration rules

• Below-threshold confidence → more evidence, not a guess. The agent's first instinct is another photo or test, and its last resort is "unverified," never bravado.
• Implications, not adjectives. "Poor quality" is banned vocabulary; "the back legs will loosen because the joint is end-grain nailed" is the register.
• Mirror the user's language. Swahili in, Swahili out; mixed in, mixed out. Voice notes accepted everywhere text is.

Architecture & build plan

Stack (thin client, smart service)

• Client: Android-first. PWA for pilot speed, thin native wrapper (Capacitor) before scale — camera ergonomics, share-to-WhatsApp intents, and home-screen presence matter here. Client owns capture UX, compression (≤400 KB/shot), offline queue.
• Agent service: Claude Agent SDK server-side; one session per assessment; protocol files loaded per item type; prompt caching on protocols cuts cost ~40–60%.
• Data: object store for photos (lifecycle: purge originals after 90 days, keep crops), Postgres for structured assessments — this table is the future shop index. Auth: phone number + SMS OTP (Africa's Talking).
• Cost control: images downscaled to inspection-sufficient resolution before upload; per-user daily caps; Haiku-class triage for image-quality rejection before the full model runs.

Phased build

Fundi Bora reuses this entire assessment engine — protocols, schema, storage — with a different lens (diagnosis instead of verdict). Build once, wear two hats; the sibling doc covers what's added on top.

Assumptions log

Every assumption this plan rests on, including process decisions made while producing these deliverables. DECIDED = confirmed with you; ASSUMED = our call, reversible; TO TEST = must be validated in the field.

Open questions & next decisions

1. Pilot geography: one dense cluster (Gikomba/Ngong Rd furniture belt) vs. one estate's buyers. Cluster-first gives the shop index density fastest.
2. Verdict vocabulary in Swahili: the exact three-level wording needs native-speaker workshopping — tone carries the trust.
3. Partner strategy: pilot alone, or with a development partner (GIZ-type SME programs, KAM's informal-sector arms) who brings fundi networks and study funding?
4. Item-type order: we assumed chairs/stools, tables, beds first (highest regret-per-shilling). Sofas hide the most defects — argue us out of deferring them.