Proof, not promises

School-list extraction and catalog matching

School-supplies retailer, Chile. Document intelligence: school-list extraction and catalog matching. In production since Dec 2025. 190,325 items extracted. 7,563 lists. More than 200 active schools. Peak of 316 lists in a day. 97.9% list-level extraction success. 95.7% resolved on first pass, 4.3% sent to human review. Model cost of about US$0.01 per list. We do not claim time savings: no baseline was instrumented. The client reports seasonal headcount held while the number of schools rose.

AI learning content inside the LMS

AI-generated learning content delivered inside the LMS, with grade passback. About 150 active activities. More than 20,000 students. More than 66,000 accumulated feedback responses. It runs on the client’s own cloud subscription. Releases go through their change-control committee. Full LTI 1.3 integration.

Controlled MCP layer over operational data

A controlled MCP layer over operational data, used through an LLM for analysis and demand planning. Nominated users in planning, advertising, logistics, IT, and management. 7 commercial accounts loaded. Two countries. Planning went from two-week test loads to a complete monthly plan. Efficiency gain reported by us, not independently measured. Per-user permissions, tenant isolation, separate prod and dev.

Marketplace listings, prices, and images

Automation of listings, prices, and images on marketplace channels. Operated by the client’s own team. A weekly production changelog is delivered to their team. Price adjustments are calculated against the live marketplace price. The production database was migrated without service loss.

Conversational AI on WhatsApp and web

Conversational AI on WhatsApp and web with governed escalation to human agents. It serves real end customers daily. Escalations route to the client’s own agents. Three escalation paths. Unique claim lock with automatic re-notification. Roles: admin, superadmin, agent.

Public MCP server for reservation search

A public MCP server that turns natural-language intent into validated reservation-search links. 17 countries served, verified by a direct call to the endpoint. JSON-RPC 2.0. Documented as compatible with the OpenAI Responses API, Claude Desktop, and Cursor.

Operational CRM for travel and visa filings

CRM and operational flow: applications, payments, receipts, messaging, and automatic replies with AI. 76 pages and more than 130 API routes. Tables with more than 1,000 records. Automatic code review on every pull request. A preview per PR. Runtime alerts to the delivery team. Row-level security.

WhatsApp activation with consent and NPS

Conversational WhatsApp activation with consent capture, FAQ, and an NPS survey. 12,583 messages. 1,273 unique users. 15 weeks. Typical response in 2 to 6 seconds. NPS 68.6 on 86 valid responses. The measured volume drop after the consent step was reported to the client, with a design correction.

Sales-rep training avatar

A conversational avatar to train sales representatives, in a controlled pilot. 52 sessions, 32 graded evaluations, and a grounding audit of 237 questions that found 192 unsupported answers, reported to the client.