Páez Aduanas
AI agent system + avatar · 2026
Páez Aduanas, a customs brokerage in Almería founded in 1928, wanted to talk on LinkedIn about what really moves its industry, without putting its reputation on the line over a regulation gotten wrong. We built six AI agents that read the official regulations, decide what’s worth talking about, and leave three scored, fact-checked drafts on a dashboard. The system never publishes. For its EXPORT ASIA event, we also created Celia, an AI-generated host.
- Client
- Páez Aduanas
- Year
- 2026
- Our role
- System design, development, and AI video
- Platform
- Agent system + Notion dashboard
- Status
- System validated in July 2026 · Celia at EXPORT ASIA in September
- 6
- chained agents
- 329 s
- from 23 items across 5 sources to 3 scored posts
- 14
- 4K clips of Celia for the event screen
Context
01
From mule carts to Brexit.
Páez was born in 1928 with mule carts hauling grapes from Berja to the port of Almería; today it clears the tomatoes, peppers, and cucumbers from Almería’s greenhouses for Germany, France, and the UK.
The brief wasn’t a website: it was an agentic content system for LinkedIn, with a review dashboard.
The challenge
02
Rigor over volume.
In customs, a regulatory fact gotten wrong is a reputation problem, and it can read as advice. Every claim has to be backed by its source.
- 01A person always makes the call: the system proposes, publishing is human.
- 02Never a blank day, even in August and over long weekends.
- 03Sources the client can edit without touching code.
- 04Celia: a believable avatar on a giant screen, from a single image, in European Spanish.
How we worked
03
The process, step by step
- 01
Research
History, generations, AEO certification, markets, and the growing season, flagging anything unverified instead of filling in the blanks.
- 02
Sources
Vetted one by one: Spain’s tax agency (AEAT) through its official RSS feed, industry newsletters, a filtered BOE (Spain’s official gazette), and the European Commission. Three sources ruled out, with the reason in writing.
- 03
Architecture
Six agents chained through files, each one runnable and debuggable on its own, with prompts versioned outside the code and the brand in a single file.
- 04
First run
23 items from 5 sources, 3 scored posts in 329 seconds, and 5 claims flagged for fact-checking.
- 05
Four reviews
Product quality, code robustness, parity between the engine and n8n, and legal and GDPR risk. Anything urgent got fixed the same day.
- 06
Demo
Set up to run from the client’s own computer, with nothing but a browser.
- 07
Celia
Technical research on LED screens, models, image, and voice; a pilot clip before the full batch; and several rounds of iteration on expressiveness and pronunciation.
Problems & fixes
04
Problems we ran into and how we fixed them
Fact-checking had to block, not decorate
- Problem
- On the first run, all three posts carried a serious warning, but the dashboard ranked them by score as if they were ready to publish.
- Fix
- A serious warning drops the post into Notion as “Blocked — verify,” with its risk in plain sight and its score capped; the editor agent gets the source’s actual text, and brand rules are checked in code.
- Result
- Fact-checking now has teeth. The editor agent even caught that the AEAT said “in 2026” with no specific date, while the draft treated the requirement as already in force.
Three posts, not three takes on the same one
- Problem
- Two of the three angles came from the same news story, and the drafts kept repeating the same tic: seven “it’s not X, it’s Y” lines in three posts.
- Fix
- Each angle declares its topic and can’t repeat a source or data; three is now a maximum. Each writer agent knows its siblings and has to stand apart from them.
- Result
- Less filler, and a tone that doesn’t give the AI away.
Never a blank day, never a stowaway
- Problem
- One Monday, the 72-hour window left out everything from the AEAT; in the BOE, “dua” matched inside “individual” and let university contracts slip in.
- Fix
- A window that widens step by step and says so, and a filter with word boundaries and normalized accents.
- Result
- There’s always material, and the system tells you when it’s less fresh.
Solid when nobody’s watching
- Problem
- A source that returned HTML counted as “0 updates,” and the chain kept going with no real material.
- Fix
- A feed that can’t be parsed is an error, explicit contracts between agents, classified retries, and traces for every agent from day one.
- Result
- The system fails loudly instead of flashing green while producing nothing.
Images only for the approved post
- Problem
- Illustrating all three drafts means throwing away two out of three, and images are the most expensive part of the system.
- Fix
- The art director agent only kicks in after a human approves.
- Result
- Around 66% less image generation, with no drop in quality.
Celia, believable on the big screen
- Problem
- In the pilot, the avatar kept closing its eyes and bringing its hands into frame; the voice read “Export Asia” the English way, and an emotion tag pushed up the pitch of the opening.
- Fix
- An engine with a fixed gaze and no hands, pitch matched by measuring the fundamental frequency before generating, silences built into the audio, and a local transcription to check that the voice doesn’t change a single word.
- Result
- 14 clips in 4K, lips and audio less than 25 ms apart, and 56 frames checked with no hands in sight.
What we built
05
Pieces of the project
- 01
Celia
The EXPORT ASIA host, generated with AI from a single image.
02On the big screen
Fourteen segments for the event screen: welcomes, speakers, a panel discussion, and the closing.
Results
06
What we pulled off
329 s
from 23 items to 3 posts
First real run: 5 sources, 3 scored posts, and 5 claims flagged for fact-checking.
5 + 3
sources, active and ruled out
Five active and verified; three ruled out, with the reason in writing.
≈ 66%
less image generation
Images only get made for the post a person approves.
28
banned phrases
Checked in code, not in the prompt.
10:11
minutes of Celia on screen
14 clips in 4K at 25 fps, mastered to −16 LUFS.
≤ 25 ms
offset between lips and voice
Checked frame by frame.
Services
Built with
