I’m Brian. I run American Data Networks in Costa Rica. Here I share what we’re building, the problems we run into, and what I’m learning along the way.
“There’s nothing more satisfying than creating something that didn’t exist before—except creating it alongside my son, Mike.”
— Brian Green
Brian and his son Mike — creating together.Mike, Chief AI Engineer — a little ADN Brain humor.
TelecommunicationsPractical AIBusiness automationLearning by doing
The story behind the work
We decided to build it ourselves.
I spent two years waiting for the automation I wanted. Then I told my son we needed to do this ourselves. We had a lot to learn.
Our AI journey · A personal account
We had to learn. So we got started.
Mike’s question about installing Linux is a moment I want to preserve. So is José Alejandro’s willingness to take on anything—from hardware to software. This is their story too.
The starting point, as I remember it. I’m still putting the exact timeline together.
What needs to work better? 01
What do we already know? 02
Can we build something useful? 03
Did it actually work? 04
These questions keep us moving. We are still working through the answers.
Selected projects
The problems we’re working on.
Each of these started with something I wanted to work better in the company. Some pieces are working. Others are still being tested or put together.
01 / COMPANY KNOWLEDGE
American Data Brain
Built a foundation · Still improving it
We had information in emails, customer records, and work orders. I wanted us to be able to find it. Brain brings those sources together, but gaps and conflicting records still need attention.
I want a customer’s question to reach the right place. We built a router to help with that. Fresh tests showed us wording it still misunderstands, especially when people ask in unexpected ways.
Finding information is one thing. Turning it into a useful reply is another. We’re working on how to draft an answer, check it, and get it back to the right conversation.
I want to know who has taken responsibility and what happens next. We built a dashboard to bring conversations and commitments together. A closed ticket still isn’t proof of a solved problem.
Checking records · Keeping unclear cases in review
The service record and the bill should agree. We built an auditor to spot differences and make only narrowly approved corrections. When the match is unclear, it stays with a person.
We tried a local AI model on real email passages. It answered quickly, but some answers added facts that weren’t there. Finding that before staff relied on it gave us a clear next problem to solve.
From reviewed ADN project records only. Tests and private candidates are not production releases; this is not a complete account of all projects.
Conversation retrieval
A private search candidate was refined and replayed against historical support questions. Tests showed better retrieval of earlier discussions; this does not establish customer-facing answer accuracy.
Mail capture verification
Recent capture artifacts were integrity-checked and targeted recovery work was reconciled. Remaining source gaps and review gates prevent a claim of complete, production-ready knowledge.
Knowledge review workflow
The procedure-review workflow gained clearer ownership and review deadlines. Human technical acceptance and final document approval remain outstanding.
ADN progress · 31 October 2026
New developments in the ADN work
Progress from reviewed ADN project records. Offline tests and proposals are not production deployments; this does not represent all projects.
A safer contact-center test environment
A local simulator now exercises customer-identification and message-handling flows with 50 fictional profiles. Forty-one synthetic tests passed. It does not connect to live customers or send real messages.
Measuring the context needed to help a customer
The team measured selected network and billing data for bounded groups of 20 and 50 accounts to inform the next design. This is not a full data inventory, production cache, or capacity guarantee.
Continuing historical email reconciliation
Archive catch-up advanced through additional captured records, with remaining source gaps tracked separately. An offline search candidate is not yet a validated customer-facing knowledge service.
ADN progress · 1 November 2026
Progress from the ADN work
Progress from reviewed ADN project records only. Isolated tests and candidates are not production releases; this is not a complete account of all projects.
Customer identification with clearer circuit matching
A contact-center lookup candidate was refined to recognize circuit families while keeping different accounts separate. Isolated tests passed; hosted integration and customer release are still pending.
Visual guidance for network equipment
A reference library for five device models now includes reviewed indicator-state examples and short animations. The materials were successfully reused in a separate test project. Field-device validation and customer release remain pending.
Email archive reconciliation
Another archive integrity checkpoint was recorded, alongside a separate offline search candidate. Source gaps remain, and the candidate is not connected to customer-facing service.
ADN progress · 2 November 2026
Technical developments from the work
Progress from reviewed ADN project records. Internal tests and candidates are not customer-facing launches.
Cross-system reconciliation
An internal reconciliation index completed additional maintenance and verification runs while preserving historical records. Some source refreshes remain partial, so coverage is not yet complete.
Visual troubleshooting in a local prototype
A local troubleshooting chat now selects audited equipment examples from typed descriptions. Bilingual labels, example videos and a restorable checkpoint passed isolated checks; it is not deployed to customer channels.
Historical support search checks
A private search index built from archived WhatsApp support conversations passed integrity, reference-link and repeatability checks. It is a historical test candidate, not a live answer service or model training.
ADN progress · 3 November 2026
Technical developments from the work
Progress from reviewed ADN project records. Internal reviews and limited tests are not customer-facing launches.
Data dictionary consistency checks
An internal review compared field definitions with an earlier catalog and confirmed matching names and types in the reviewed set. The historical catalog was preserved; this was not a production change.
Read-only integration verification
An integration review documented additional bounded availability and access checks for existing data and AI services. Results distinguish confirmed scoped reads from limited or pending checks; no new features were deployed.
We changed the guide Gemma uses to route a message. Its reference library stayed the same.
ADN progress · 5 November 2026
Verified developments from ADN work
Progress from reviewed ADN project records. Isolated tests and pilots are not production launches.
RAG source and recovery audit
A read-only review reconciled existing knowledge sources and documented remaining freshness, permission and restore-readiness checks. No database migration or customer-facing knowledge release was performed.
Offline message classification comparisons
A bounded comparison evaluated three request-classification approaches on 1,000 test cases and documented disagreements and retry evidence. These were offline evaluations, not a deployed routing change.
Private customer monitoring pilot
A limited monitoring pilot began observing a fixed group of circuits using existing measurements and a local operator view. The signals require human diagnosis; no automatic repair or general customer rollout was established.
ADN progress · 6 November 2026
Verified developments from ADN work
Progress from reviewed ADN project records. Isolated tests and pilots are not production launches.
Independent recovery-readiness monitoring
A recovery-readiness monitor was observed continuing after the initiating session ended. Bounded synthetic validation passed, but this does not establish a completed recovery or customer-serving readiness.
Bilingual routing boundary checks
An offline request-classification candidate passed targeted checks for English and Spanish preferences about not receiving invoices by email. Authorization checks remain necessary, and no change was deployed to customer conversations.
IPTV startup guidance for review
A bilingual television startup guide and supporting procedure were indexed for technical review. Real-device compatibility testing and formal SOP approval remain pending.
ADN progress · 8 November 2026
Verified developments from ADN work
Progress from reviewed ADN project records. Isolated tests and evaluations are not production launches.
IPTV documentation review
An internal reference candidate collected updated vendor middleware and API documentation and reviewed documented playback, program-guide and time-shift functions. Compatibility with installed versions and live service testing remain pending; no production API operation was performed.
RAG recovery-readiness review
A read-only inventory reconciled source and recovery evidence across authorized systems and documented outstanding verification gates. It does not establish a completed restore, customer-facing release or production migration.
Spark AI evidence registry
A versioned register consolidated verified DGX Spark hardware and observed local-model information for a proposed network-operations evaluation. Controlled tests and approval remain necessary before wider deployment.
ADN progress · November 9, 2026
Verified ADN project progress
Progress from reviewed ADN project records. Plans and isolated tests are not production deployments.
Development source history preserved
A reviewed source-control consolidation preserved earlier development histories and verified a clean canonical checkout. This is a development baseline; application tests and production release remain outstanding.
Knowledge-index recovery reviews
Read-only checks and a versioned recovery plan documented outstanding freshness, access and restoration checks across the knowledge indexes. The work does not certify a completed recovery or a customer-facing service.
Isolated database migration pilot
A controlled pilot restored the actual database structure on an ARM-based test system without customer records and verified the tested index structures. Security and cutover approvals remain pending; no production migration occurred.
ADN progress · November 10, 2026
Verified ADN project progress
Progress from reviewed ADN project records. Plans and isolated tests are not production deployments.
Guarded billing requests tested in Spanish
An isolated bilingual test checked more natural phrasing for requests using already-supported billing evidence. The reviewed synthetic cases retained evidence safeguards and blocked unsupported answers. This is not a live customer rollout or a test of unrestricted model responses.
Interactive AI sales prototype tested
A phone-style prototype completed a bounded trial with fictional sales conversations and preserved test transcripts. The review identified wording and workflow gaps to address before wider use. No customer messages, orders or production routing occurred.
ADN progress · November 12, 2026
Verified ADN project progress
Progress from reviewed ADN project records. Plans and isolated tests are not production deployments.
Safer handoffs evaluated in isolated support tests
An experimental support handoff passed 30 local component and integration-bridge checks, including recovery from simulated duplicate requests. The operational integration remains pending. No live customer assignment or message was made.
Bilingual support navigation validated with synthetic cases
A held support-template candidate passed 4,518 isolated checks of bilingual actions and category navigation. These results are from synthetic tests, not a customer rollout or proof of live service improvements.
A little about me
I want technology to do useful work.
I’m the CEO of American Data Networks. A lot of my questions start with everyday work: why does this take so long? Why is someone doing it by hand? What would it take to make it work better?
This site is where I want to share what happens when we try to answer those questions. The AI journey is a big part of that. I want to show the people, the progress, and the things we haven’t figured out yet.
