Lameda: a commerce platform that lives where Nigerian merchants already sell

Nigerian merchants manage multi-million naira businesses through WhatsApp voice notes, manual bank transfers, and Excel sheets. Orders arrive as voice notes. Payments are screenshots forwarded to a group chat. Every tool built for this market, Shopify, WooCommerce, every Western SaaS, was designed for English-speaking, card-paying desktop users. None of them fit.

Lameda is my answer as a founder: a full commerce platform delivered through WhatsApp, Instagram, and Telegram. Merchants onboard in five minutes, and customers shop, pay, and track orders through the same chat window they use for everything else. The platform covers the complete cycle, AI conversation, product search, cart and checkout, payments with direct merchant settlement, delivery tracking, human handoff, lead follow-up, and a merchant dashboard, and every part of it has been built and tested internally.

Nigeria's informal commerce sector runs on WhatsApp. Small merchants selling fashion, food, electronics, and home goods take orders in DMs, negotiate prices in group chats, and collect payment via Opay, PalmPay, or manual bank transfer. This works until it doesn't: there is no inventory system, no payment records, no order history, and no way to run a promotion without typing the same message to three hundred contacts one by one.

The tools that exist were built for other markets. Shopify requires a card-accepting payment gateway, a storefront, and a browser. WhatsApp Business has no catalog search, no cart, and no checkout flow. The formal e-commerce platforms either price out small merchants or require onboarding flows that assume a desktop user with a stable internet connection.

• Platform rules are load-bearingWhatsApp and Instagram business messaging runs on Meta's approval processes and quality signals. An account that messages too aggressively gets restricted, so the platform has to protect merchants from themselves automatically.
• Unit economics at Nigerian price pointsAI cost per message determines whether the pricing model works at all. Every AI call has to be priced to the job it does, not routed to the most capable model by default.
• Data protection as a baselineLameda operates under Nigeria's data protection regulation. Merchant details and customer delivery addresses are encrypted at rest; nothing personally identifying is stored as plain, readable text.
• Consent, not assumptionMarketing requires an explicit yes, asked after first purchase, never inferred from silence. Opt-outs are respected everywhere, immediately, and every conversation opens by disclosing the AI assistant.

Three groups live inside the product: merchants running their businesses through the assistant and dashboard, their customers shopping in chat, and live agents receiving human handoffs with response-time enforcement. Around them sit the platform dependencies: Meta's Cloud API for WhatsApp and Instagram, Telegram's Bot API, Paystack for payments and settlement, and the AI providers.

As founder I also built for a fourth role: the operator. A separate, tightly secured Lameda admin portal handles merchant approvals and platform oversight, kept deliberately apart from the merchant-facing surface.

The product thesis came from watching how the market actually transacts: orders as voice notes, payment proof as screenshots, promotions typed contact by contact. The gap was not awareness of e-commerce tools; it was that every existing tool demanded the merchant change channels, change habits, and change how their customers buy.

The pilot itself became the research instrument. The original onboarding flow asked merchants for business details, delivery zones, a logo, and pricing before the assistant went live, and fewer than 40% of signups reached an active assistant. That number, not an opinion, is what reshaped onboarding into five questions with everything else editable later.

Build the commerce stack inside the channel merchants already use. The merchant stays on WhatsApp. Their customers stay on WhatsApp. The infrastructure, catalog, cart, payments, delivery, CRM, runs underneath, invisible to both.

Architecturally, that meant one backend for every channel: WhatsApp, Instagram, and Telegram customers all talk to the same assistant and commerce logic, with only the messaging layer changing per channel. Commercially, it meant going through Meta's official approval processes from the start, because a platform handling other people's customer conversations and payments has to sit on infrastructure that cannot break without warning.

• option 01Unofficial messaging workarounds vs Meta's official channelsRouting around Meta's approval process would have been faster. Chosen instead: the official Cloud API with its paperwork and review cycles, because merchants' livelihoods cannot sit on infrastructure that could be shut off without warning.
• option 02AI lead scoring vs simple rulesAn LLM could judge how "interested" each customer seems, at added cost and unpredictability. Chosen instead: transparent rules. Hot means items in cart or checkout reached; warm means interest shown. Free to compute, and a merchant can always see exactly why a customer is flagged.
• option 03One flagship model vs task-matched modelsRouting every AI call to the most capable model is easier to build. Chosen instead: a fast, lightweight model for text conversation and a more capable one for reading product photos, with automatic fallback to a second provider during outages. This cut AI cost per message by more than half.
• option 04Per-channel builds vs one backend with adaptersBuilding each channel as its own stack would ship channel one faster. Chosen instead: a single conversation engine and commerce core behind thin channel adapters. When Instagram support was added, it slotted in without touching the engine, the catalog, or checkout.

• Approval time over launch speedMeta's business messaging approval takes real time and paperwork, and the platform absorbed that delay deliberately, including a full migration sprint onto official infrastructure. Trust is the product; there was no shortcut worth taking through it.
• Upfront routing engineering over easy promptingThe task-matched AI routing took real effort to build while a one-model setup would have worked fine at pilot volume. At growth volume, cost per merchant is what decides whether the pricing model survives, so the effort was spent early.
• Boring infrastructure over controlSupabase over self-hosted Postgres, a managed scheduler over a custom one. Every hour not spent maintaining infrastructure is an hour spent on the product a merchant actually feels.
• Tested guardrails over assumed onesA safety mechanism that has never been triggered by real traffic is a hypothesis, not a guarantee. Guardrails, consent enforcement, quality-signal pausing, provider fallback, are tested against live behaviour, which costs time and buys certainty.

Lameda shipped across thirteen-plus sprints, each targeting a coherent layer: the AI conversation engine and core commerce loop first, then merchant tooling, growth features, monetisation, platform infrastructure, and compliance. Every sprint shipped to the demo environment; there were no long staging branches.

1. Sprints 1-2 · FoundationMerchant onboarding, the AI assistant engine, product catalog, cart
2. Sprints 3-4 · CommercePayments, order lifecycle, delivery tracking, customer data encryption
3. Sprint 5 · Agent LayerHuman handoff inbox with live updates and response-time enforcement
4. Sprint 6 · GrowthBroadcast messaging with opt-out built in, referral system, analytics dashboard
5. Sprints 7-7.5 · PlatformSubscription billing, settlement tools, assistant health monitoring, team management
6. Sprint 8 · Commerce DepthDelivery lifecycle, restaurant categories, the public discovery assistant
7. Sprint 9 · WhatsApp LaunchWhatsApp channel goes live, AI cost optimisation, automatic AI provider fallback
8. Sprint 10 · Lameda Admin PortalA separate, tightly secured operator app for merchant approvals and platform oversight
9. Sprint 11 · Meta MigrationWhatsApp and Instagram move onto Meta's official infrastructure
10. Sprint 12 · Trust and ComplianceExplicit consent capture, always-on AI disclosure, automatic protection against messaging-quality issues
11. Sprint 13 · Lead CaptureHot, warm, and cold lead tiering with one-tap follow-up

The AI layer has three jobs, each on the model that fits it. Every incoming message is read for intent first, greeting, product search, checkout, complaint, in English, Nigerian Pidgin, or both mid-sentence. Replies and product descriptions run on a fast, lightweight model tuned to each merchant's business. Product photos run on a more capable model that matches fabric, cut, colour, and brand against the catalog. If the primary provider has an outage, the assistant falls back to a second provider rather than going silent. This routing keeps the blended cost around ₦1.58 per message, less than half of a one-size-fits-all setup, with no drop in reply quality.

The full commerce cycle has been built and tested internally: a customer can discover a merchant, shop by text or photo, pay through Paystack with automatic settlement and commission split, track delivery, and be followed up as a lead, all without leaving the chat. Privacy is built into the data layer rather than bolted on: encryption by default, explicit consent tracking, always-on AI disclosure, automatic protection of merchant accounts when messaging-quality signals drop, and no third-party data sharing.

• ›Onboarding friction kills conversion faster than product gaps. Fewer than 40% of signups survived the original setup flow. Cutting it to five questions, with everything else editable later, changed how merchants trusted the product: working assistant first, configuration after.
• ›Unit economics belong in sprint one. Routing every AI call to the flagship model is the path of least resistance, and at growth volume the cost per merchant is what determines whether the pricing model actually works. The task-matched routing paid for its upfront effort.
• ›Test the safety net, don't just trust that it's there. A compliance guardrail can look complete in code review and still not trigger correctly against a real message from a real platform. Guardrails get tested against live traffic before they are trusted.