AI Software Development Services: How We Built an Intelligent E-Commerce Platform for a Delhi-Based Retailer
Industry: B2C E-Commerce (Home & Lifestyle Products) | Project Type: AI Software Development | Location: New Delhi, India | Timeline: 8 months
Artificial intelligence is no longer the exclusive domain of tech giants. A mid-sized B2C e-commerce company — selling home décor and lifestyle products online across India — came to us with a set of very concrete business problems: high cart abandonment, an overwhelmed customer support team, and a product catalogue so large that customers struggled to find what they were looking for. Their conversion rate had plateaued, and their customer lifetime value was stagnant.
They needed AI software development services that delivered real, measurable outcomes — not research projects. Over eight months, we built three integrated AI-powered capabilities into their existing platform: a personalised product recommendation engine, a conversational AI support chatbot, and a predictive demand forecasting system. Here is the full story.
The Business Problems We Were Solving
Problem 1: No Personalisation
The platform showed the same homepage, the same featured products, and the same promotional banners to every visitor — whether they were a first-time browser or a loyal customer who had made 15 purchases. Every user saw the same catalogue sorted by the same default ranking. The result was poor relevance, low click-through rates on product listings, and a bounce rate of 64% on the homepage.
Problem 2: Overloaded Customer Support
The customer support team of 11 agents was handling over 800 tickets per day. Analysis of the ticket data revealed that approximately 62% of all tickets were one of five repeatable query types: order status, return initiation, delivery timeline, product availability, and refund status. These were queries that could be resolved without human intervention — if the right system existed.
Problem 3: Inventory Blind Spots
Demand forecasting was done manually by the merchandising team based on gut feel and last year's sales. This led to overstocking of slow-moving items (tying up working capital) and stock-outs during peak periods like Diwali and Holi (resulting in lost sales and customer frustration).
Our AI Development Approach
We began with a four-week AI Readiness Assessment — an evaluation of the client's existing data assets, infrastructure, and organisational readiness. We analysed 18 months of transaction data (3.2 million orders), user clickstream data, product catalogue structure, customer service tickets, and inventory records. This assessment produced a prioritised roadmap of AI initiatives ranked by expected business impact and implementation feasibility.
Three initiatives were approved for the first phase of development:
- Personalised Product Recommendation Engine (Highest expected revenue impact)
- Conversational AI Support Engine (Highest expected cost reduction)
- Demand Forecasting & Replenishment System (Working capital impact)
For a broader understanding of how AI is transforming software, see our article: How AI Is Transforming Software Development.
System Architecture
All three AI capabilities share a unified data infrastructure and are exposed through a single API gateway. The diagram below shows the full solution architecture — client interfaces, microservices, data stores, and the MLOps pipeline.
Each AI module runs as an independent Python microservice on AWS EC2, consuming from a centralised Microsoft SQL Server 2019 data warehouse. Apache Kafka handles real-time clickstream ingestion, while daily batch ETL jobs refresh the training datasets for weekly model retraining cycles.
Architecture Overview: End-to-End Data Flow
The following diagram illustrates how data moves through the platform — from ingestion through processing, model serving, and end-user consumption:
Module 1: Personalised Product Recommendation Engine
We built a hybrid recommendation engine combining collaborative filtering (what similar users bought) with content-based filtering (product attribute matching) and real-time session signals (what the current user has browsed in this session). The engine served four recommendation surfaces across the platform:
- Homepage personalisation: Each user's homepage hero banner and featured products now adapt based on their browsing and purchase history.
- Product detail page — "You may also like": Contextually relevant product suggestions surfaced based on the viewed item.
- Cart page — "Complete the look": Complementary product recommendations shown at checkout, increasing basket size.
- Post-purchase email recommendations: Personalised product suggestions in transactional emails, driving repeat purchase intent.
The model was retrained weekly on fresh transaction data. A/B testing was run for eight weeks comparing the AI-personalised experience against the original non-personalised experience. The recommendation engine delivered a 48% improvement in product listing click-through rate and a 32% increase in average order value for personalised sessions.
Module 2: Conversational AI Support Engine
We did not use a generic chatbot SaaS platform. Instead, we built a custom NLP-based chatbot trained on the client's own support ticket history — 18 months of real conversations between agents and customers. The chatbot was designed to handle six intent categories:
- Order status enquiry (integrated with order management API for real-time status)
- Return & exchange initiation (automated workflow trigger)
- Delivery timeline queries (integrated with logistics partner API)
- Product availability and stock queries
- Refund status tracking
- General product information queries
For anything outside these six categories, the chatbot handed the conversation to a human agent with full context — preserving customer experience while routing only genuinely complex cases to human support. The chatbot was deployed on the website, the mobile app, and via WhatsApp Business API.
Post-deployment results: The chatbot handled 63% of all incoming support queries without human intervention. The human support team's daily ticket load dropped from 800+ to under 300. Average first-response time for customers dropped from 4 hours (during peak traffic) to under 30 seconds.
Module 3: Demand Forecasting & Intelligent Replenishment
We built a demand forecasting system using Facebook's Prophet time-series model combined with XGBoost for feature-rich product-level predictions. The model incorporated:
- Historical sales data (24 months)
- Seasonal indices (Indian festive calendar, monsoon impact on specific categories)
- Promotional calendar (planned discounts and marketing campaigns)
- Current inventory levels and supplier lead times
The system produced weekly SKU-level demand forecasts with confidence intervals, and generated automated replenishment purchase recommendations for the merchandising team. Over-stocking of slow-moving items was flagged automatically with markdown recommendations. Peak-period forecasts for Diwali and New Year were validated against actuals and showed an 89% forecast accuracy at the SKU level — compared to the previous manual process which the merchandising manager estimated at 50–60% accuracy. Read more on this topic: How Predictive Analytics & AI Are Revolutionising Business Forecasting.
Before vs. After: Business Impact
| Metric | Before AI Development | After AI Development | Change |
|---|---|---|---|
| Average order value | Baseline | +32% improvement | +32% |
| Homepage bounce rate | 64% | 41% | 36% reduction |
| Product listing CTR | Baseline | +48% improvement | +48% |
| Cart abandonment rate | 76% | 57% | 25% reduction |
| Customer support tickets per day | 800+ | Under 300 | 60%+ reduction |
| First customer response time | 4 hours (peak) | Under 30 seconds | "99% faster |
| Inventory overstock (as % of SKU catalogue) | 31% of SKUs overstocked | 12% of SKUs overstocked | 61% improvement |
| Demand forecast accuracy | "55% (manual) | 89% (ML model) | +34 percentage points |
Key Implementation Lessons
Data Quality Is Everything
The single biggest predictor of AI project success is data quality. We spent 3 of the 8 project months on data cleaning, deduplication, and feature engineering before any model training began. Businesses that rush to model training on dirty data consistently produce poor-performing AI systems.
Start With High-Impact, Narrow Use Cases
Rather than attempting to "AI-ify" everything at once, we scoped each initiative tightly. The chatbot handled exactly six intent categories — not an open-ended general assistant. The recommendation engine optimised for one metric (order value) per surface. Focused scope = faster deployment = faster ROI.
Human Oversight Is Not Optional
Every AI capability we built had a human override mechanism. Support agents could always take over chatbot conversations. The merchandising team received recommendations, not mandates, from the forecasting system. This is both ethically correct and practically important for building internal trust in the system.
Why Custom AI Development Outperforms Generic AI SaaS
Several generic AI recommendation and chatbot SaaS platforms exist. The client had actually tested one recommendation SaaS for six months before approaching us — and saw negligible improvement because the generic model had no understanding of their specific product catalogue structure, their customer base's regional preferences, or their seasonal demand patterns. A custom-built AI system trained on your own data, built for your specific business logic, consistently outperforms generic tools. The additional upfront investment pays back in superior outcomes.
For a deeper dive into AI's role in ERP and business software: How AI Is Transforming ERP Systems in 2026 and Beyond.
About Net Soft Solutions' AI Software Development Services
Net Soft Solutions provides end-to-end AI software development services for Indian businesses — from AI readiness assessment and data strategy through model development, integration, deployment, and ongoing monitoring. We work across industries including e-commerce, manufacturing, education, healthcare, and SaaS. Contact us for a free AI readiness consultation.
Related Resources
- Role of AI in E-commerce Development
- Top Data Analytics Tools Transforming Business Decision-Making
- How AI Is Revolutionising Mobile App Development
- How Artificial Intelligence Is Transforming the World of Web Design
Project Timeline
The engagement spanned eight months with a structured phase breakdown: