13 cases, all measured

What was built, what was measured, what the client did not have to do. Clients anonymized by default.

AI in Education

AI-powered pronunciation feedback for an online English school

Five teachers correct students' pronunciation every day using a system that listens to the audio, compares it to the lesson's target, and returns feedback — at a per-audio cost 89% lower than the first version.

R$ 0.26
per 100 audios (was R$ 2.29)
41,651
audios screened in a single day, zero errors
AI Customer Support

AI-powered WhatsApp support, measured in seconds

A school's business WhatsApp began responding with a median time of 96 seconds, complete with a monitoring dashboard, imported history, and two response modes — without changing the number or the phone.

96 s
median response time (before: hours)
4 tabs
dashboard: conversations, queue, metrics, pairing
Integrations

Online store and ERP talking to the CRM — without duplicating a single order

Orders from Nuvemshop and Bling now automatically become deals in the CRM. The duplicate race that showed up during every sales spike was eliminated with a serial queue and a safety sweep.

0
duplicates after the serial queue
77
automated tests
Transcription and audio at scale

+38% capacity across a transcription fleet, without buying hardware

A benchmark of 789 real audio files revealed that the speech recognizer was resampling up to six times per file. Fixing one parameter cut −27% off processing time per audio file, confirmed in production the same day.

+38%
fleet throughput (3.56 → 4.90 audio files/s)
−33%
time per audio file on the tightest machine
Voice and agents

Voice agents: from HUD to a voice-driven terminal

Three voice systems in production: a WhatsApp assistant that talks, a JARVIS-style panel with camera and tools, and a bridge that lets you give a code agent voice commands and get a spoken summary back.

30
voices in the gallery, cached
72
tests on the voice → terminal bridge
Infrastructure and continuity

Infrastructure that fits in $5/month and comes back on any machine

Everything runs on Cloudflare (Workers, Pages, R2, KV, D1) with a cost ceiling enforced in code. Encrypted backup to two destinations, and a recovery kit that restores the entire environment on a clean machine — tested, not just promised.

< US$ 5
per month, limit enforced in code
30/30
items restored in the clean-HOME test
RAG and knowledge bases

25 YouTube channels became 26 queryable mentors, with timestamps

Thousands of videos transcribed and indexed per second. Every answer cites the video and the exact moment. On top of that, a course app with exercises graded by AI.

25
channels indexed
8,232
videos with clickable timestamps
AI Content and Imagery

Three illustrated textbooks with ~550 consistent images

Volumes 3, 4, and 5 of an English-language collection for children, with characters that stay consistent across hundreds of images, QA at scale, and a client approval portal.

3
volumes produced and assembled
≈550
images generated
Google data and integrations

Google Ads, AdSense, GA4, and GAM data integrations and auditing

An engine that read ad revenue, calculated the real value of each conversion, and fed it back to Google Ads to optimize for return — with cross-API auditing because dashboards lie.

5 APIs
Ads, AdSense, GA4, GAM (REST+SOAP), BigQuery
10 accounts
under one MCC
Agent orchestration

Idle AI quota turned into assets — 16 deliverables in one night

I measured that 42% of AI usage windows were sitting empty. A dispatcher started spending the leftover capacity on verifiable tasks, with numeric criteria and a reviewer that rewards refutation. First night: 17 tasks, 16 approved.

42%
of the 5h windows were empty (21 days measured)
16/17
tasks approved on the first night
AI-Powered Sales

AI Prospecting: The Agent Writes, the Human Sends

A system that builds the list, researches each business from public sources, writes a one-page brief with evidence, and prepares the message cadence. The first message never sounds like a bot — because a bot in the first message kills the conversation.

5
messages per prospect, with a 24h/+2d/+4d cadence
1 page
brief with a source for every claim
Assessments and diagnostics

Technical assessments that decide purchases, architecture, and incidents

One-to-ten-page documents that answer a business question with measurement: which server to buy, why the platform is slow, whether it's worth building a new product. Published on our own infrastructure, shareable with the client.

4
assessments delivered in 2026
789
audio files on the capacity assessment's test bench
Portals and files

Delivery and approval portal for clients, on your own infrastructure

Large file uploads, temporary links, and a PIN-protected approval portal where the client reviews and approves deliverables — no Google Drive, no WeTransfer, no third-party file hosting.

4 h
drop link validity
multipart
large file upload

Yours could be the 14th

Start with one page and the math done.