How Small Businesses in Chhattisgarh Are Starting to Use AI

There is a perception that artificial intelligence is something only large technology companies use. That it requires massive budgets, PhD researchers, and data centres full of GPUs. A few years ago, that was largely true.

It is not true anymore.

In the last two years, we have built AI-powered systems for businesses in Chhattisgarh — manufacturers, retailers, service companies — that have never had a technology team, let alone an AI strategy. The common thread is not that they decided to "adopt AI." It is that they had a specific, expensive problem, and AI happened to be the most practical solution.

Attendance fraud in a manufacturing facility

A manufacturing facility near Raipur had a persistent problem: buddy punching. Workers swiped badge cards for absent colleagues. The payroll team estimated they were paying for 15-20 ghost shifts per month. That is roughly ₹3-4 lakhs in annual losses from a single facility.

They tried fingerprint scanners. Workers with rough hands from the factory floor had a 30% failure rate. The system was abandoned in weeks.

We built them a facial recognition attendance system using on-device face detection and AI-powered face matching. No cloud dependency — the entire system runs on their local network. A worker walks past a tablet, their face is recognized in under a second, and attendance is logged.

The buddy punching stopped immediately. The system paid for itself within four months.

The important detail: this was not a technology project. It was a payroll accuracy project. The facility manager did not care about FaceNet embeddings or pgvector similarity search. He cared that his attendance numbers were finally correct.

Inventory management that actually predicts demand

A retail distributor in Raipur managing stock across multiple warehouses was doing demand forecasting in Excel. Their buyer would look at last year's numbers, adjust for "gut feeling," and place orders. Some months they had ₹12-15 lakhs of dead stock. Other months they ran out of their best-selling items.

The underlying data was already there — two years of sales history, seasonal patterns, supplier lead times. What was missing was the ability to process it systematically.

We did not build them a fancy machine learning pipeline. We built a relatively simple forecasting model that runs weekly, compares predicted demand against current stock levels, and generates reorder suggestions. The interface is a dashboard that shows three things: what to order, how much, and by when.

Dead stock dropped by 35% in the first quarter. Stockouts dropped by half. The model is not perfect — no forecast is — but it is systematically better than gut feeling.

Document search that understands meaning, not just keywords

A government-adjacent organization in Chhattisgarh had thousands of documents — circulars, policy updates, operational guidelines — stored across shared drives. Finding the right document meant either knowing exactly where it was or spending 30 minutes searching through folders.

We built them a document management system with semantic search. Instead of matching exact keywords, the system understands what you mean. Search for "leave policy for contract workers" and it finds the relevant circular even if it uses different terminology like "temporary staff absence guidelines."

The technology behind this — text embeddings and vector similarity search — is the same technology that powers tools like ChatGPT's retrieval features. But deployed on-premise, running against their specific documents, producing results in their context.

Staff who used to call colleagues to ask "do you remember which circular covered this?" now search for it in seconds.

What makes AI practical for small businesses now

Three things changed in the last few years that made AI accessible to businesses that are not technology companies:

Pre-trained models. You do not need to train a face recognition model from scratch. Models like FaceNet and ArcFace are freely available and work out of the box. The same applies to language models for text processing and forecasting frameworks for time-series data. The heavy research is done — the value is in applying it to your specific problem.

Cheaper compute. Running an AI model used to require expensive GPU servers. Many practical AI applications — face recognition, document search, demand forecasting — run comfortably on standard hardware. Our facial attendance system runs on a regular server, not a GPU cluster.

Better tools. PostgreSQL now has pgvector for similarity search. Python's ecosystem for ML has matured significantly. Deploying AI models in Docker containers is straightforward. The gap between "research prototype" and "production system" has shrunk dramatically.

How to think about AI for your business

If you are a business owner in Chhattisgarh considering AI, here is how to think about it:

Start with the problem, not the technology. The question is not "how can we use AI?" It is "what is costing us the most money or time?" AI might be the answer, or a simpler solution might work better.

Look for repetitive, data-heavy decisions. If someone in your organization spends hours every week making decisions based on data they already have — demand planning, quality inspection, document retrieval, attendance tracking — that is a candidate for automation.

Expect ROI within months, not years. If an AI project cannot demonstrate measurable value within 3-6 months, it is either solving the wrong problem or being built wrong. The projects we have described above paid for themselves within one to two quarters.

You do not need a data science team. You need engineers who understand both AI and production systems — who can build something that actually runs reliably in your environment, not just in a Jupyter notebook.

What we have learned

Building AI systems for businesses in Chhattisgarh has taught us that the technology is rarely the hard part. The hard part is understanding the business well enough to know which problem is worth solving, deploying the solution in an environment with real constraints (unreliable internet, non-technical users, harsh physical conditions), and making it reliable enough that people trust it.

That is production engineering, not just AI. And it is something that can be done from Raipur, for businesses in Raipur, without needing anyone or anything from Bangalore or Silicon Valley.