Computer vision and demand forecasting for operations teams
By Vishal · published 25 October 2026
Classic machine learning still wins on tabular and visual problems: forecasting demand, extracting fields from documents, and spotting defects in images cost a fraction of a language model to run and are easier to audit. VAI Five builds these for Indian manufacturers, retailers, logistics firms and clinics, measured against the forecast or inspection you do today.
Demand and time-series forecasting
Indian demand has shapes a generic model misses: festival calendars that move each year, regional holidays, monsoon effects, salary-day spikes, and promotions that cannibalise the following week. A forecast that knows them beats a spreadsheet by enough to change stock, staffing and cash decisions.
- What we forecast: SKU-level demand, store footfall, call and ticket volume, energy and raw-material use, cash collections.
- How: gradient-boosted models with calendar and promotion features for most retail and logistics problems; probabilistic models where you need a range, not a point; hierarchical reconciliation so SKU forecasts add up to category forecasts.
- Output: a forecast table your planners already understand, with uncertainty bands and a weekly accuracy report against actuals.
Document extraction
Invoices, purchase orders, KYC documents, prescriptions, lab reports, shipping bills — the paperwork of Indian business arrives as scans, photos and PDFs in several languages and layouts. We combine vision-language models with validation rules (totals must add up, GSTIN must be valid, dates must be plausible) so the output goes straight into your ERP with exceptions routed to a person. Typical results move a team from typing every field to reviewing a minority of flagged ones.
Visual inspection
Surface defects on fabric, tiles, castings and printed packaging; fill levels and label placement on lines; shelf compliance in stores; counting in warehouses. Models are trained on your images — augmented with synthetic defect images when real defects are rare — and run on an industrial camera's edge device or a GPU server, with a reviewer interface for low-confidence cases.
How we measure success
| Problem | Metric you will see | Baseline it must beat |
|---|---|---|
| Forecasting | Weighted error at SKU and category level; stock-outs and over-stock avoided | Your current planner or spreadsheet forecast |
| Extraction | Field-level accuracy; share of documents needing human review | Manual entry time and error rate |
| Inspection | Defect recall and false-alarm rate; escapes to customer | Current manual inspection sampling |
How a project runs
- Feasibility note (free) — a call, a sample of your history or images, and a one-page note on the achievable accuracy and the pilot price.
- Two-week pilot (fixed price) — a model on one product line, document type or camera, scored against your baseline.
- Build and handover — integration with ERP, WMS or the production line, monitoring, documentation, and training for the people who will use it.
Questions people ask
Do we need special cameras or hardware?
Often not. Many inspection problems work with an industrial or even a good phone camera and consistent lighting; we assess this in the feasibility call and specify hardware only when the accuracy target needs it.
How much history do you need for forecasting?
Two years of weekly data captures annual seasonality and festival shifts; one year works with more uncertainty. For new products we use similar-item and category patterns.
Can this run without internet at the factory?
Yes. Inspection and extraction models can run on an edge device or a local server and sync results when a connection is available.
Tell us the problem, not the model.
A 30-minute call and a one-page feasibility note, free. If a two-week pilot makes sense, we scope it with a fixed price. If it does not, we say so.
Related