GVVA — say it "Guava" — is an applied-AI engineering firm co-founded and run by Kamogelo Mahlasela and Masego Letsoko — Johannesburg-rooted, serving clients on three continents. This repo is the case study of the platform GVVA built for itself: the same class of multi-agent, serverless system we deliver to clients, running our own intake, delivery, and finance end-to-end.
We believe an automation firm should be its own best case study. We run on what we sell.
About this repo. Case study of GVVA's internal production platform. The source code is proprietary; code in
excerpts/is sanitized and representative. All rights reserved — see the copyright notice below.
Three named systems, one pipeline: a prospect chats with Nectar, the deal and delivery live in Guava Pulp, and the founders steer everything from SEED.
flowchart LR
subgraph Front["Client-facing"]
W["gv-va.com"] --> N["Nectar<br/>AI intake concierge"]
end
subgraph Engine["Guava Pulp - AI ops engine, 25 Lambda functions"]
WH["Webhook intake"] --> LEADS["Leads and conversion"]
LEADS --> INT["AI intelligence endpoints<br/>analyze, plan, estimate, invoice, report, draft"]
INT --> APPR["Human approval queue<br/>no AI email leaves unreviewed"]
AUTO["EventBridge automations<br/>daily reports, payment reminders"]
DOC["Document AI<br/>Textract processing"]
FIN["Finance<br/>invoices, payment webhook"]
end
subgraph Ops["Founder-facing"]
SEED["SEED dashboard<br/>Next.js + Supabase"]
end
N -->|"qualified lead"| WH
SEED --> APPR
Nectar is GVVA's conversational intake agent (Google Gemini 2.0 Flash behind a Node/Express API, also packaged for AWS Lambda). Plenty of agencies have a chatbot. Nectar's difference is the thing we're proudest of:
"Rooted in Africa, native to the world." Nectar doesn't just translate — it adapts manners. The prompt architecture implements six regional cultural protocols and 10+ languages, so the same intake conversation is formal and honorific-aware for a Tokyo executive, high-energy and ROI-first for a US founder, relationship-led for a Gulf trader, and celebrates hustle in Lagos slang for a Nigerian entrepreneur.
That cultural engine extends to substance, not just tone:
- Regional payment intelligence — recommends M-Pesa in Kenya, Paystack in Nigeria, Yoco/SnapScan in South Africa, Pix in Brazil, Stripe in the US: 25+ payment systems mapped to region, so advice is never generically Western.
- Regulatory awareness — speaks GDPR to Europe, POPIA to South Africa, CCPA to California, HIPAA to US healthcare.
- A structured 7-step intake flow underneath the warmth: business profile, pain points, scoping, and next steps land in the database as structured lead data, with notification emails to both sides.
- Production hygiene — rate limiting, security headers, CORS allow-listing, structured logging, graceful degradation.
The back office is a 25-function serverless engine (Node.js 20 on AWS Lambda, SAM-deployed, HTTP API with Cognito JWT auth) over three DynamoDB tables with purpose-built access patterns (a founder-dashboard index and a client index over a single-table project design).
What it runs:
- Lead lifecycle — webhook intake from Nectar, list/get/convert; a lead becomes a project without re-keying.
- AI intelligence endpoints (Anthropic Claude) — per-project
analyze-lead,generate-plan,estimate-costs,generate-invoice,generate-report, anddraft-email: the judgment-heavy paperwork of a services firm, generated on demand. - Human-in-the-loop by design — AI-drafted client communications land in an approval queue; a founder approves before anything sends. Same principle as our client work: AI drafts, humans authorize.
- Document AI — client documents upload via presigned S3 URLs into a versioned bucket and are processed with Amazon Textract.
- Finance rails — invoice generation and tracking, a payment webhook, and EventBridge-scheduled automations: daily operating reports and automated payment reminders.
A Next.js + TypeScript + shadcn/ui admin panel on Supabase (Postgres with row-level security): live lead pipeline (NEW → WON/LOST), project detail pages with AI-generated roadmaps, and interactive task management. Two founders see the whole company on one screen.
The same engineering shows up in GVVA's client deliveries — most visibly a production multi-agent AI platform for a U.S. personal-injury law firm that runs the firm's entire lead-intake operation: three cooperating AI workers, TCPA-compliant outbound voice with a full evidentiary audit trail, server-side ad attribution, and an operations dashboard. It's documented in three dedicated case studies:
| Case study | What it covers |
|---|---|
| legal-intake-ai-platform | The full three-agent platform on AWS serverless |
| tcpa-compliant-voice-engine | Deterministic IVR, machine detection, recording→transcript→PDF audit trail |
| serverless-ad-attribution | Meta CAPI, multi-domain pixel routing, marketing forensics |
Beyond that flagship, GVVA has delivered client back-ends in food-delivery/logistics and hospitality across South Africa and Asia.
- AI drafts, humans authorize. Every outward-facing artifact — an email, a legal letter, an outbound call — passes a human or policy gate before it touches a real person.
- Own your tools. Clients get systems they own outright — code, infrastructure, and data — not rented seats.
- Compliance is a feature. POPIA, GDPR, TCPA: designed in from the first diagram, not bolted on for the audit.
- Enterprise discipline at boutique scale. Versioned backups before every change, evidence-based post-mortems, feature-flagged rollouts that ship dark and enable deliberately.
| File | Pattern it demonstrates |
|---|---|
excerpts/regional_protocol.js |
The cultural-protocol engine: region detection → tone, payments, and compliance mapping |
excerpts/approval_queue.js |
Human-in-the-loop gate: AI-drafted communications held for founder approval |
Built by Kamogelo Mahlasela and Masego Letsoko.
© 2026 Kamogelo Mahlasela and Masego Letsoko (GVVA). All rights reserved.
This repository is published for viewing only, so prospective employers, clients, and collaborators can evaluate our work. No license is granted. Beyond viewing on GitHub (and the limited on-platform rights GitHub's Terms of Service provide), no part of this repository — text, architecture diagrams, or code excerpts — may be copied, reproduced, modified, distributed, or used to create derivative works without our prior written permission.