Data and AI projects Semantix delivered for large companies: the challenge, what was done and the result. Clients are not named, for confidentiality.
case studies published
38
sectors, plus cross-sector cases
7
Financial servicesRisk and fraud
Fraud prevention in payments
Challenge
Manual review of transactions and of fraud committed by users, merchants and corporate clients.
What we did
Unsupervised models for each product, business rules and automatic collection of company registration data.
−70%samples sent to manual review
+25%fraud detected
Measured result
Financial servicesRisk and fraud
Fraud score in auto insurance
Challenge
Claims fraud, misuse of roadside assistance and underwriting rules without data behind them.
What we did
Supervised and unsupervised models, explainable decisions and monitoring of the model life cycle.
−20%loss ratio
Measured result
Financial servicesData governance
Data-driven credit policies
Challenge
A new cloud data platform the credit team barely used, and heavy dependence on IT.
What we did
Data consumer profiles, access policies, lineage, a data catalog and automated data quality.
+60%of credit policies based on data
+82%of the team using the platform within 30 days
−35%duplicate data
Measured result
Financial servicesData platform
Moving models to MLOps in the cloud
Challenge
On-premise architecture, SQL-centred development and no CI/CD flow for models.
What we did
A Google Cloud platform with a feature store, training pipelines, a model registry, separate environments and infrastructure as code.
22models running on the new platform
12feature sets migrated
Measured resultGoogle Cloud
Financial servicesProcess automation
Sales cancellation back office
Challenge
Cancellation requests arriving by email, manual checks of attachments and a risk of errors and duplicates.
What we did
Automation that reads the emails, checks the attachments, records the cancellation and notifies the merchant, with an analytical report.
60→2 minper request
168→1,000emails handled per month
−82%audit findings
Measured result
Financial servicesSales and customers
Cross-selling payment solutions
Challenge
A base of more than 100,000 business clients using few products in the portfolio.
What we did
Recommendations based on similarity between clients, with a business-rule layer and a propensity score.
50%of the base with a recommendation
41%of recommendations are cross-sell
Measured result
Retail and consumer goodsForecasting and planning
Demand forecasting in pharmacy retail
Challenge
Replenishment calculated on outdated assumptions, long lead times in some stores and high disposal costs.
What we did
AI demand forecasting for every SKU and a new replenishment calculation that includes external factors and lead time.
−42%stock-outs
−13%inventory
Pilot · MVP in 20 stores, 1 month
Retail and consumer goodsForecasting and planning
Stock-outs and losses in a supermarket chain
Challenge
Rising stock-outs, unavailable products that were hard to spot and losses from manual workarounds.
What we did
Demand forecasting by SKU and store, an automated pipeline to run the model and a monitoring dashboard.
−60%stock-outs
−12%losses
Pilot · MVP in 6 stores, 1 month
Retail and consumer goodsMaster data and customer 360
A single customer view in a fashion group
Challenge
Customers spread across many systems and brands, and a CRM to restructure in line with data protection law.
What we did
A dedicated data architecture, a customer master record, MDM, stewardship, AI segmentation and self-service BI.
100%of customers in MDM
Everybrand in the group segmented
Measured result
Retail and consumer goodsSales and customers
Clothing recommendations in e-commerce
Challenge
Generic recommendations that did not reflect each customer’s style.
What we did
Deep learning recommendations based on style, keywords and price history.
+22%sales over the previous recommendations
Measured result
Retail and consumer goodsForecasting and planning
Why products don’t sell at the point of sale
Challenge
SKUs selling below expectations, with no way to tell where or why.
What we did
Forecasts by SKU and point of sale, market intelligence collection and automatic detection of causes.
Delivery
Retail and consumer goodsNatural language
Consumer opinion on marketplaces
Challenge
Thousands of marketplace reviews with no structured reading.
What we did
Review collection, a sentiment model and an indicator dashboard by store and SKU.
Delivery
Manufacturing and automotiveForecasting and planning
Raw material demand forecasting
Challenge
High error in next-month forecasts, affecting purchasing and production.
What we did
AI models for one- to three-month horizons, with training, validation and testing.
11%→2%next-month error (MAPE)
Measured result
Manufacturing and automotiveOptimization
Inventory optimization across plants
Challenge
Excess stock alongside stock-outs across several plants.
What we did
Analysis of demand, lead time and criticality, optimal stock levels, purchase recommendations and a control tower.
−20%inventory levels
Measured result
Manufacturing and automotiveOptimization
Allocating people on the production line
Challenge
Unstructured allocation, absenteeism and overloaded teams.
What we did
An allocation optimization simulator, delivered as an end-to-end digital product.
+25%line speed
Measured result
Manufacturing and automotiveData governance
Data governance at a multinational chemicals company
Challenge
Duplicate data, unknown business rules and a data lake that other operations in the Americas were about to use.
What we did
Data curation, a catalog and glossary, review of every new pipeline, a data quality process and data owners.
−40%pipeline development time
−18%code refactoring
100%of new pipelines reviewed
Measured result
Manufacturing and automotiveProcess automation
Automatic approval of service orders
Challenge
Manual price checks on parts and labour in dealership service orders.
What we did
Data standardization, pricing models, external price bots and a containerized model integrated with the systems.
+5,000service orders automated
Measured result
Manufacturing and automotiveComputer vision
Safety and quality with computer vision
Challenge
PPE use, process quality and ergonomics checked by sampling.
What we did
Computer vision monitoring of PPE use and of each process step.
Delivery
Manufacturing and automotivePeople
Forecasting absenteeism in plants and stores
Challenge
Absences disrupting plants and points of sale.
What we did
Absence probability, a weekly forecast per unit for the next four weeks and alerts to managers.
Delivery
Manufacturing and automotiveOptimization
Logistics network for a food manufacturer
Challenge
Production, storage, distribution and taxes decided separately.
What we did
Network optimization with the interstate tax matrix and price equalization.
Delivery
Health and pharmaForecasting and planning
Demand forecasting in the pharmaceutical industry
Challenge
Next-month forecasts for hundreds of SKUs, with a goal of cutting error by 20%.
What we did
Several models tested for each SKU, with the best one chosen automatically.
−24%average error (MAPE) across 329 SKUs
−35%error on the top 20 SKUs
Measured result
Health and pharmaRisk and fraud
Abuse at a health insurer
Challenge
Manual pre-audit of procedures and fraud in high-cost procedures.
What we did
A denial probability model, automatic retraining, scoring every 2 hours and automatic decisions on low-value requests.
R$ 1.1Min authorizations that became automatic
R$ 1.3Min procedures flagged for denial
Measured result
Energy and utilitiesData platform
Forecasting regulatory penalties at a power distributor
Challenge
Fines for long outages, fragmented low-quality data and indicators calculated by hand.
What we did
A lakehouse with a medallion architecture (Bronze, Silver and Gold) on Databricks in Microsoft Azure, machine learning forecasts, a penalty simulator and dashboards.
36%more accurate than the business team’s forecast
R$ 3.4Min avoidable fines
Expected result · Forecast measured in the pilot; the avoided fine is an expected resultDatabricks · Microsoft Azure
Energy and utilitiesAI agents
AI agents that recover tax credits
Challenge
More than 112,000 entries per quarter and extensive tax law, all reviewed by hand.
What we did
An orchestrator agent and nine agents specialized by type of service, identifying PIS and COFINS tax credits.
90%precision in identifying credits
+R$ 20Min credits identified per month
Measured result
Energy and utilitiesForecasting and planning
Consumption forecasting in the free energy market
Challenge
Consumption and billing for thousands of clients that were hard to predict, with a risk of penalties.
What we did
Hourly forecasts per client and per market using external data, energy balance reports and scenario dashboards.
+2,000consumer units forecast hour by hour
Measured resultAWS
Energy and utilitiesRisk and fraud
Anomaly detection in power plants
Challenge
Operating limits that were hard to set and outages that reduce sellable energy.
What we did
Unsupervised models on sensor time series, model life-cycle management and an anomaly dashboard.
Month 1first significant anomaly detected
Measured resultAWS
Energy and utilitiesOptimization
Risk and return in an energy portfolio
Challenge
A concentrated portfolio of futures contracts.
What we did
Portfolio optimization with Monte Carlo simulation, backtesting and decision dashboards.
Delivery
AgribusinessForecasting and planning
Agricultural commodity price forecasting
Challenge
Volatile soy and corn prices, limited capacity and expensive spot freight.
What we did
An ensemble of models, news monitoring and forecasts for 1, 5 and 22 business days.
+24AI models
+5,000variables monitored
Measured resultGoogle Cloud
AgribusinessOptimization
Allocating mobile balers for collection
Challenge
Limited vehicle and supplier capacity for collection.
What we did
Supplier clustering and route and allocation optimization.
+2,000 t/daycollected above the previous level
Measured result
LogisticsSales and customers
Flight and offer recommendations at an airline
Challenge
Flights with low load factors and the same offer for every passenger.
What we did
Clustering, association rules and passenger similarity to recommend flights and services.
+50%revenue from email marketing
+400%revenue from social media
Measured result
LogisticsOptimization
Dynamic pricing for bus tickets
Challenge
Prices set manually on every route.
What we did
Data ingestion, demand forecasting, price optimization and API integration with the ticketing system.
4,000prices calculated per day
200prices published per day
Pilot · Test version
LogisticsOptimization
Scheduling rail equipment
Challenge
Wagon and equipment scheduling done without optimization.
What we did
Constraint programming and integer linear programming.
Delivery
LogisticsForecasting and planning
Road freight price forecasting
Challenge
Freight costs that were hard to predict on reference routes.
What we did
A forecasting accelerator, news monitoring and a geographic scenario simulator.
Delivery
Cross-sectorAI agents
Procurement assistant integrated with SAP Ariba
Challenge
The procurement team handling first-level support, and duplicate entries caused by human error.
What we did
A virtual assistant that answers questions and checks order status in SAP and Ariba, 24/7.
51%of requests handled by the assistant
~R$ 5Mper half-year in freed-up capacity
Expected result · Projected savingsSAP Ariba
Cross-sectorNatural language
Video interview analysis
Challenge
Interviews analysed by hand by teams of five to seven people.
What we did
Transcription, speaker separation, topic modelling and sentiment in an end-to-end pipeline.
−65%analysis time
Measured result
Cross-sectorComputer vision
Trademark similarity in registration filings
Challenge
Logo conflicts identified by hand.
What we did
A deep learning model that compares images and flags similar marks.
13.5%error, against 27.3% for the previous method
Measured result
Cross-sectorSales and customers
Customer referrals in real estate
Challenge
High acquisition cost and few sales coming from referrals.
What we did
Propensity models for referring and converting, segmentation and a simulator.
15%→45%conversion in the prioritized audience
Measured result
Cross-sectorPeople
Labour claim risk
Challenge
Labour lawsuits with no predictability.
What we did
The probability of a claim after each dismissal, with a dashboard for legal and HR.
Delivery
No case matches this sector and theme. Choose another filter.
“Pilot” means an MVP result in a limited scope; “Expected result” means a projection calculated with the client. All other figures were measured in the project.
Want a result like this in your operation?
An assessment shows which data and processes bring a return first.