software engineer · co-founder of Exequtech — shipping field service management SaaS  · Linux, Neovim · final-year BComp

github

Jeremia Fourie  • Peerlist

<h3>In-cockpit pilot-fatigue detection on a Raspberry Pi.</h3><p>This project focuses on developing an edge device for in-cockpit pilot-fatigue detection using a Raspberry Pi. The system aims to monitor physiological signals to identify fatigue before it impairs performance.</p><ul><li><strong>Vision:</strong> Utilizes a real-time face pipeline to compute Eye Aspect Ratio (EAR) for blink and microsleep detection, and tracks yawning frequency. Aviation-informed thresholds are used to flag microsleep events.</li><li><strong>Biometrics:</strong> Integrates heart rate and heart-rate variability (RMSSD) to generate a stress index, escalating the fatigue assessment based on sustained anomalies.</li><li><strong>Environment:</strong> Incorporates cabin conditions like altitude, recognizing that hypoxia can compound fatigue risk.</li><li><strong>Identity:</strong> Employs face embeddings for pilot enrollment to establish individual baseline comparisons.</li></ul><p>The fused indicators are published continuously over TLS MQTT, a protocol chosen for its efficiency and suitability for intermittent cockpit connectivity. The system prioritizes edge computing to minimize bandwidth, latency, and privacy concerns, ensuring detection capabilities even without a stable network link. The core challenge was achieving real-time computer vision within the Raspberry Pi's resource constraints.</p><p>The project involved leading the edge team, which included building the necessary circuitry, guiding the device architecture, and sourcing the hardware. The Python/OpenCV pipeline was developed by teammates, while the ground station's operator frontend and pilot-enrolment API were also part of the lead's contributions.</p><p>The complete system, comprising edge devices and a cloud ground station, was a final-year capstone project that achieved a score of 97%. This project highlighted the critical role of hardware limitations in shaping software design.</p>

CogniFlight Edge

<h3>Exequ-Jobs</h3>
<p>Exequ-Jobs is an offline-first Android application designed for field technicians. It addresses the critical need for functionality in areas with no mobile signal, ensuring that work orders, photo capture, signature collection, and time tracking can continue uninterrupted. The app is built with an architecture that prioritizes local data integrity, with a robust synchronization mechanism to reconcile data with the cloud once connectivity is restored.</p>
<ul>
<li><strong>Offline-First Architecture:</strong> The encrypted local database (Room over SQLCipher) serves as the single source of truth, guaranteeing data persistence even without a network connection.</li>
<li><strong>Seamless Synchronization:</strong> Background workers like SyncOrchestratorWorker and PhotoSyncWorker manage data reconciliation, job downloads, and photo uploads, ensuring data consistency between the device and the cloud.</li>
<li><strong>Real-time Updates:</strong> Server-sent events facilitate the flow of real-time updates for jobs and status changes directly into the local database.</li>
<li><strong>Multi-tenant Support:</strong> A single binary supports multiple tenants, with dynamic request routing based on short-lived tenant JWTs issued by Exequtech's central console.</li>
<li><strong>Production-Ready Features:</strong> Includes end-to-end encryption at rest, native crash symbolication via Crashlytics, ML Kit for barcode scanning, biometric authentication, and distinct staging/production environments.</li>
<li><strong>AI-Assisted Development:</strong> The development process leverages AI tools, with the lead engineer maintaining full ownership of architecture, reviews, and releases.</li>
</ul>
<p>Key technologies powering Exequ-Jobs include Kotlin, Jetpack Compose for the UI, Room for local data persistence, SQLCipher for encryption, Apollo GraphQL for API interactions, Hilt for dependency injection, WorkManager for background tasks, and Firebase services. The application is continuously developed and shipped, currently at v1.4+.</p>

Exequ-Jobs

<h3>The Product</h3><p>Exequtech OS is the field service management (FSM) platform my company builds for South African trade businesses — electricians, plumbers, HVAC and solar installers. Job cards flow into quotations and invoices; technicians clock time and log trips; stock and assets are tracked across sites. One Go API serves three clients: the operator web app, a marketing/auth console, and the <a href="/projects/exequ-jobs/" target="_blank" rel="noopener noreferrer">Exequ-Jobs Android app</a>.</p><h3>The Interface is an Operating System</h3><p>The operator app doesn't use the usual sidebar-and-tables SaaS layout. It is a <strong>desktop</strong>: draggable, resizable, snapping windows; a taskbar; a start-menu app grid — a pattern I first built for <a href="/projects/cogniflight-cloud/" target="_blank" rel="noopener noreferrer">CogniFlight's ground station</a>. Billing, Jobs, Assets, Contacts, Reports and Settings are "apps" you open side by side — because a dispatcher genuinely works that way, quoting in one window while checking a technician's clock-ins in another.</p><p>Under the hood it's React 19 + TypeScript (strict) with the React Compiler, Apollo Client with full codegen typing against the GraphQL schema, virtualized data grids for large tenants, in-browser PDF generation and barcode scanning. A Playwright-based profiling harness tracks heap, CPU and leak baselines so the window manager stays fast as the platform grows.</p><h3>Tenant Isolation at the Database Layer</h3><p>Multi-tenancy is enforced where it can't be forgotten: in PostgreSQL itself. This is Brian Felgate's design — he leads our Go backend — and living with it has permanently changed how I think about isolation. Every request opens a transaction and stamps it with the caller's identity via set_config; row-level-security policies and audit triggers then scope every query and record every change.</p>// Every transaction is stamped with the caller's tenant before any query
// runs; RLS policies enforce isolation below the application layer.
func setAuditContext(ctx context.Context, tx pgx.Tx, data *Ctx) error {
    if data.Tenant != nil {
        if _, err := tx.Exec(ctx, "SELECT set_config('audit.tenant_id', $1, true)", data.Tenant.ID.String()); err != nil {
            return fmt.Errorf("failed to set audit.tenant_id: %w", err)
        }
    }
    // member_id, request_id and reason follow the same pattern,
    // feeding the trigger-based audit trail.
    return nil
}
<p>A forgotten WHERE tenant_id = ? in application code therefore leaks nothing — the database refuses to return foreign rows. The same session variables drive an immutable, trigger-based audit log tied to request identity.</p><h3>Other Pieces of What We've Built</h3><ul><li><strong>Schema-first GraphQL with two surfaces</strong> — a console API for account/tenant management and a tenant API for the business domain, both generated with gqlgen from 18 schema files.</li><li><strong>A server-side document engine</strong> that renders quotations, invoices and job-card reports to PDF through a themeable template system.</li><li><strong>An MCP server wrapping the API</strong>, so AI agents can operate the platform through the same permission model as humans.</li></ul><h3>My Role</h3><p>I co-founded Exequtech and lead the product's web side: the operator app and the console are primarily my commits (561 on the web platform), and I built the Android client, AI-assisted. My co-founder Gerhard van Staden builds on the web platform with me — and with years of web development and client consulting behind him, he's also the one who mentors me in both. Hosting and deployment are Brian's domain — he manages the production environment hands-on and double-checks the security himself, which is deliberate: it's how we keep the risk managed while working with AI this heavily. We build Exequtech in our free time — a small crew testing our limits — but the users are real, so the standards are too. I work with Claude Code daily; the deal is simple: it types fast, I review, decide, and answer for what I ship.</p>