Resources · Case Studies

What happened when real companies bet on private AI

Anonymized implementations across Brazilian mid-market companies — retail, finance, healthcare, industrial, services. What changed in 90 days: ROI, operational metrics, architecture designed and LGPD compliance.

Retail · 1,200 stores Revenue BRL 4.8B/year

From USD 18k/month in GPT-4 to USD 2.4k/month with private SLM

Challenge: Customer service assistant handling 17k SKUs changing prices daily. Monthly fine-tuning of Llama 70B was expensive and the agent still quoted last month's prices.

−87%
Monthly Cost
99.8%
Price Accuracy
90d
To Production
QLoRARAG · QdrantPresidiovLLM · FP8
Talk to architect
Fintech · B2B SaaS 300 products · +340% BI

Conversational CFO ingesting the entire Data Lake

Challenge: Accounting team spending days consolidating NPS, engagement, and financial data spreadsheets to generate reports for the board.

+340%
BI Productivity
+35%
Cross-sell
−60%
CAC
MedallionAirflow · dbtBigQueryMDA LLM 256k
Talk to architect
SaaS B2B · Mid-Market Team of 12 SDRs

3× more meetings booked without new headcount

Challenge: High acquisition cost, SDRs spending 60% of their time finding data instead of talking to leads. Email outbound with 20% open rate.

Meetings/month
−60%
CAC
70%+
WA Open Rate
LeadBot · 45 fieldsWhatsApp BusinessHubSpotSuperGrow
Talk to architect
E-commerce · Marketplace ~ 40k tickets/month

L1 support eliminated, team focused on critical cases

Challenge: 70% of tickets were repetitive questions (order status, returns, invoices). Human team spending 4h+ in queue, CSAT below 70 and costs scaling linearly with volume.

−70%
L1 Tickets
23s
Resolution
+35%
CSAT
WhatsApp · ZendeskMCP · Shopify · ERPAporiaHuman handoff
Talk to architect
Healthcare · Health Plan 2.3M members

AI for call center approved by legal in 6 weeks

Challenge: Health plan operator wanted to automate scheduling and FAQs over WhatsApp, but legal blocked it due to PII exposure and regulator required full traceability of all automated decisions.

100%
PII Masked
6wks
DPO Approval
VPC
São Paulo · BR
Presidio · LakeraImmutable logsSOC 2 · ISO 27001VPC sa-east-1
Talk to architect
Industrial · 100-500 DataOps + Vibe coding

Legacy codebase deciphered in 1 sprint with 256k context

Challenge: Internal pricing system with 47k lines of undocumented Python, only senior dev retired. Intermittent production bug and no one understood the complete data flow.

256k
Single Context
1sprint
For the fix
100%
Doc Generated
MDA LLMChain-of-ThoughtRAG over codeAporia
Talk to architect