Enterprise Internal Tools

A manufacturing company with 800+ employees across 4 facilities was drowning in manual processes. Attendance was tracked through fingerprint scanners that broke frequently in the factory environment (dust, moisture, and physical wear), resulting in missing records that HR spent 15+ hours per week reconciling. Access control to restricted areas (chemical storage, server rooms, executive floors) relied on physical key cards that were routinely shared, lost, or duplicated. The operations team had no centralized dashboard — they relied on a patchwork of spreadsheets, email threads, and a decade-old ERP system that only two people in the company understood how to query.

The client wanted a unified internal tool platform that could replace these fragmented systems with something modern, reliable, and usable by non-technical staff. The attendance system needed to handle shift changes, overtime calculations, and leave management. The access control system needed zone-based permissions (different employees authorized for different areas) with real-time audit logs. And the operations dashboard needed to pull data from multiple sources — the ERP, attendance system, production logs, and inventory databases — and present it through a single, queryable interface.

A unique requirement was a voice-powered query interface. The operations manager wanted to be able to ask questions like "How many units did we produce last week?" or "Who was late more than 3 times this month?" and get instant answers without writing SQL or navigating through menus. This needed to work reliably with Indian English accents and in noisy factory environments.

We designed the platform as a modular monolith — a single deployable application with clearly separated internal modules (attendance, access, dashboard, voice) that communicated through internal events rather than HTTP calls. This gave us the organizational benefits of microservices (independent development, clear boundaries) without the operational overhead (no service mesh, no distributed tracing needed for what was fundamentally a single-location deployment). The backend was built with Node.js and Express, with PostgreSQL as the primary datastore and Redis for session management and real-time event publishing.

The attendance module replaced fingerprint scanners with face recognition cameras (using the same ArcFace pipeline we developed for our hospitality client, adapted for the factory environment). We added hard hat detection as a preprocessing step — in factory zones, employees were required to wear hard hats, and the system would log a safety compliance event alongside the attendance record. The access control module used a combination of face recognition and NFC badges (for areas where camera installation wasn't practical), with a zone-permission matrix managed through the admin dashboard. Every access event was logged with a timestamp, employee ID, zone, and authorization status, creating a complete audit trail.

The voice query interface was built on OpenAI's Whisper for speech-to-text (deployed locally on a GPU server for privacy — no audio data left the premises) and GPT-4 for natural language to SQL conversion. We created a curated schema description and a library of 200+ example queries that the model used as few-shot examples. The system supported follow-up questions ("What about this month?") by maintaining conversational context, and included a confirmation step for any query that would modify data. The admin dashboard was a React application with role-based views: HR saw attendance and leave, operations saw production metrics, and facility management saw access logs and safety compliance.