Our portfolio — real work, real outcomes

📚 Case 1 · Multi-branch School ERP

Client: private school group · 3 branches · ~2,400 students · Haryana, India.

Problem: fees across 3 branches reconciled from 4 bank accounts into Excel; attendance on paper registers; parent communication over scattered WhatsApp groups; payroll run manually on the 1st of every month.

What we built: unified multi-tenant ERP — admissions, fee heads & concessions, attendance (biometric + mobile), timetable, exam & report cards, payroll, parent app (Android + iOS), bulk SMS/email, Razorpay gateway, role-based access per branch, offline attendance sync for low-network classrooms.

Stack: React 18, Node.js (NestJS), PostgreSQL + Prisma, Flutter (parent app), AWS (EC2, RDS, S3, CloudFront), Razorpay, MSG91 (SMS).

Outcome: fee-reconciliation time cut from 3 working days/month to under 2 hours. Parent-query volume reduced ~60% in 3 months. Rolled out to 2 additional branches without new engineering effort.

Timeline: 10 weeks to MVP, 16 weeks to full rollout. Model: fixed-milestone, 5 sprints.

📹 Case 2 · AI CCTV detection for warehouses

Client: logistics operator · 2 warehouses · 40+ existing IP cameras · North India.

Problem: security guard on duty missed overnight fire incidents and after-hours intrusion. Replacing 40 cameras with a "smart" solution would have cost ₹18L+.

What we built: on-prem AI inference server (one GPU box per warehouse) that connects to the existing RTSP camera feeds. Detects fire, smoke, intrusion after hours, weapons, falls and crowding. Alerts go to WhatsApp, email and a web dashboard within ~2 seconds. Zero new cameras purchased.

Stack: Python, YOLOv8 (fine-tuned on client's own footage), OpenCV, FastAPI, Redis, Docker Compose, Twilio for WhatsApp, Nvidia RTX A2000 inference boxes.

Trade-off: on-prem over cloud for lower bandwidth cost and sub-second local alerts; model updates pushed over the air monthly.

Outcome: average detection latency 1.6 seconds. One real fire incident caught at 02:14 AM, acted on within 4 minutes, damage estimated under ₹20,000 vs. an insurance quote of ₹8L+ had it spread. ROI in the first quarter.

Timeline: 8 weeks. Model: fixed hardware + software milestones, AMC on monthly retainer.

💬 Case 3 · WhatsApp AI automation for a service business

Client: home-services company · ~1,200 enquiries/month on WhatsApp.

Problem: leads arrived 24/7 on WhatsApp but were only answered 10:00–19:00. Analytics showed ~30% of leads went cold because no one replied in the first hour.

What we built: AI chatbot on the official WhatsApp Business Cloud API. Intent classification, lead qualification (service type, location, budget, timeline), appointment-slot booking, CRM push to the client's existing HubSpot. Human handover triggers automatically when the model's confidence drops below a threshold, when the user explicitly asks, or after a complaint keyword.

Stack: WhatsApp Cloud API, Node.js, OpenAI GPT-4o-mini (classification + replies), Redis (conversation state), PostgreSQL (CRM mirror), HubSpot API.

Safety: no auto-replies on refund, legal or complaint intents — always routed to a human. Full conversation log stored for 90 days for QA.

Outcome: first-response time dropped from avg 4 hours to under 30 seconds. Qualified-lead rate up ~2.4×. ~22% of enquiries now book an appointment without human involvement, freeing the team for high-value calls.

Timeline: 5 weeks. Model: fixed build + monthly LLM-cost pass-through.