Portfolio · evidence reviewed August 2026
Operations-first systems builder
I identify fragile, repetitive, or invisible work inside businesses and teams and turn it into practical systems that people can actually use.
My work connects existing tools, custom workflows, automation, and increasingly AI assistance. AI accelerates my process, but it did not create it. My core skill is understanding how work actually happens and building the missing pieces that make it clearer, more reliable, and easier to manage.
My employment kept me close to real operations — customers, escalations, teams under daily pressure. The systems here are what I built alongside those roles; this portfolio is where the two converge.
ISO/IEC 27001 consulting background. I design automation with authority, data movement, credentials, review points, and failure behavior considered from the start.
Where this fits →Three proof points
Operations before AI, architecture at scale, then AI-era coordination
Working with AI, with a human owner
How I run a bounded AI assistant
An AI assistant is useful when it has good context, restricted tools, and a named human owner for every consequential action.
Across my systems the same separations recur: reading is separated from writing, drafting from sending, observation from modification, confident matches from ambiguous ones. The assistant prepares work and surfaces context; a person owns what happens next.
- Project Command Center — a Hermes coordinator with bounded project context and restricted interfaces. It can research, retrieve context, and manage defined work items, but cannot edit repositories or act as an unrestricted shell agent.
- Unified email review — AI assists classification and drafting; drafting and sending are separate stages; a person controls final content, recipient, and timing.
- WhatsApp → CRM reconciliation — the Kairos agent records sourced contact information only on sufficiently clear evidence; ambiguous matches are left for human attention rather than guessed.
This is the operating model I would bring to any AI-assisted role.
Supporting systems
Six operational problems and the systems built around them
Guest operations intake and CRM workflow
Check-in information, email authentication, and guest messages arrive
through different systems. One reviewed operational record, with each piece of
information keeping its source.
Guest operations intake and CRM workflow
Check-in information, email authentication, and guest messages arrive through different systems. One reviewed operational record, with each piece of information keeping its source.
- The problem
- Guest check-in information lived in TalkGuest and had to be reviewed or transferred into the operational CRM manually. Access could also require an email-delivered one-time code, adding another interruption to an already repetitive workflow. Separately, important operational updates arrived in a personal WhatsApp account and required repeated application checks.
- What I designed and built
- The information path from the source portal into Kairos, an operations workspace built on the open-source Twenty CRM. A Playwright worker signs in through the normal web interface, retrieves check-in information, normalizes it, and records processing state locally. It detects previously processed records before delivery, supporting repeatable runs while reducing the risk that retries create duplicate CRM information. A separate read-only WhatsApp listener subscribes to a deliberately limited set of chats, normalizes accepted events, and forwards them on independent paths to Telegram and to Hermes.
- My role
- Designed the read-only source integration, directed AI-assisted implementation, connected OTP retrieval through myemail, and developed the destination workflow and Twenty customizations. Twenty is credited as the upstream CRM; my contribution is the workflow design, custom data model, integration layer, and operational behavior built around it.
- Operational boundaries
- The source application is never modified — the boundary with TalkGuest is read-only. When authentication requires an email code, the worker uses a narrowly scoped myemail read interface to locate the relevant message; it does not receive general authority to manage mail or send responses. The WhatsApp listener does not send from the personal account or modify the source conversation. When the Kairos agent finds sufficiently clear evidence it records the contact with WhatsApp provenance and can set it as the preferred operational contact, without overwriting the fact that another source supplied the original booking. If the reservation match is ambiguous, the workflow leaves it unresolved for human attention rather than guessing.
- Evidence / scale August 2026
- The privately deployed listener had archived 711 messages from its permitted source. Delivery paths fail independently: a temporary problem with one destination does not erase the event or prevent the other path from being attempted, and retries are tracked to avoid duplicate alerts.
- Status
- Completed — not deployed · TalkGuest → CRM Operated · WhatsApp listener
- Known limitations
- The TalkGuest-to-CRM integration was completed and validated but not operationally deployed at the August 2026 review; deployment still depended on production credentials and final destination configuration. Reduced copying, fewer transcription opportunities, and better CRM availability are therefore presented as intended operational value rather than measured production impact.
- Technologies
- Playwright browser automation · Twenty CRM (upstream) · Kairos workspace · Hermes · Telegram · myemail OTP interface
Level 3 — verification
TalkGuest integration: 25 automated tests passed, and a verification run demonstrated source authentication.
WhatsApp listener: 208 automated tests, the application build, and a dedicated read-only compliance audit all passed during verification. These checks demonstrate the implemented boundary, not a formal security certification.
Media is handled according to per-chat policy. A bounded query interface makes the permitted archive searchable without exposing unrestricted access to every personal conversation.
Institutional memory (KnowledgeForge)
Useful knowledge was becoming trapped in documents, repositories,
reports, conversations, and individual agent sessions. A durable memory and retrieval
layer for both people and AI-assisted workflows.
Institutional memory (KnowledgeForge)
Useful knowledge was becoming trapped in documents, repositories, reports, conversations, and individual agent sessions. A durable memory and retrieval layer for both people and AI-assisted workflows.
- The problem
- Knowledge was accumulating across repositories, documents, reports, conversations, and agent sessions. It was difficult to retrieve consistently or carry into later work without rediscovering the same context or depending on a particular chat history.
- What I designed and built
- A reusable institutional-memory layer. Documents, reports, code, selected conversations, and curated memory artifacts enter a searchable store with their origin and type preserved, so retrieved material can be interpreted in context. The service supports direct search as well as agent-oriented interfaces, letting later tasks recover decisions, implementation evidence, and reports without copying an entire historical conversation into every new session.
- My role
- Identified the memory problem, designed ingestion and retrieval workflows, directed AI-assisted implementation, integrated the sources, deployed the core privately, and continued improving its access model.
- Operational boundaries
- Retrieval provides relevant prior material; it does not turn every ingested statement into truth. People and agents still need to evaluate source, date, and authority. Remote access uses private Tailscale networking, and fixed-station onboarding uses scoped, revocable credentials with installation tooling for Linux and Windows stations.
- Evidence / scale August 2026
- Approximately 43,600 indexed chunks across documents, code, conversations, and curated memory artifacts.
- Status
- Operated · core Prototype · Explorer
- Known limitations
- The volume demonstrates breadth of ingestion, not the correctness, usefulness, or current relevance of every stored chunk. The access controls are concrete but are not proof of formal security: the system has not been independently audited, and the current embedding path uses an external model provider, so it is not an entirely local processing stack. KnowledgeForge Explorer is a working prototype rather than a deployed user-facing service.
- Technologies
- Vector search · Tailscale private networking · scoped revocable credentials · Linux and Windows station tooling
Level 3 — verification
79 automated tests passed during the portfolio review.
PageServer, which displays AI- and project-generated reports, is a reusable supporting utility rather than a KnowledgeForge component.
Monitoring and alerts
Systems should bring relevant changes to the person who needs them
rather than depending on someone repeatedly checking every source application.
Monitoring and alerts
Systems should bring relevant changes to the person who needs them rather than depending on someone repeatedly checking every source application.
- The problem
- Keeping track of feed activity required someone to check the source application repeatedly, and relevant changes were easy to miss.
- What I designed and built
- An operated, headless monitoring service that records observed feed activity and sends Telegram alerts, with a companion explorer for inspecting the accumulated records.
- Operational boundaries
- The claim is monitoring, retention, and notification — not ownership of the platform, control of it, or reliable unattended publishing.
- Evidence / scale August 2026
- Nearly 58,000 observed post records in the database at the snapshot.
- Status
- Operated · monitoring Inactive · posting API
- Known limitations
- A separate posting API existed but was not listening at the snapshot, so outbound posting is not presented as currently operated. No account or post content is disclosed in this portfolio.
- Technologies
- Headless monitoring service · Telegram alerting · companion record explorer
Unified email review (myemail)
Important messages were spread across Gmail, Hotmail, and Migadu. One
review workspace: sort, draft, review, then send.
Unified email review (myemail)
Important messages were spread across Gmail, Hotmail, and Migadu. One review workspace: sort, draft, review, then send.
- The problem
- Important messages were distributed across Gmail, Hotmail, and Migadu accounts. Repeatedly switching inboxes made it harder to maintain one consistent process for sorting, drafting, reviewing, and responding.
- What I designed and built
- A unified review workspace rather than another autonomous mail sender. Messages from the connected mailboxes synchronize into one central interface, which can organize them for review and turn selected messages into follow-up tasks. The operating sequence is deliberately simple: sort, draft, review, then send.
- My role
- Identified the fragmented-inbox problem, designed the workflow, directed AI-assisted development, integrated the providers, deployed the service, and refined it through ongoing personal use.
- Operational boundaries
- AI assists with classification and draft preparation, but drafting and sending are separate stages, so generated text does not automatically become an external communication. A person retains authority over final content, recipient, timing, and the decision to communicate. The system exposes narrow read interfaces for dependent automations — the TalkGuest worker can retrieve a relevant one-time code without receiving broad mailbox access or send authority. Audit entries record important processing actions.
- Evidence / scale August 2026
- Five connected mailboxes across three provider families, deployed and operating in daily use, with a unified message and task workflow and hundreds of audit entries.
- Status
- Operated · with known validation debt
- Known limitations
- Operation does not mean the current development checkout is defect-free. Verification found 52 passing tests and three failures, together with three TypeScript errors. I therefore do not describe myemail as production-ready, formally audited, or fully reliable. The honest status is an operated personal system with demonstrated value and known validation debt. No formal time study was performed, so the case supports qualitative claims about consolidation, review consistency, auditability, and reusable access boundaries rather than productivity figures.
- Technologies
- Gmail · Hotmail · Migadu · unified sync layer · AI classification and drafting · audit records · scoped read interfaces
Level 3 — verification
52 tests passing, three failing. Three TypeScript errors outstanding at the August 2026 review. Recorded here rather than in the narrative, so the limitation is stated once and stays checkable.
Messaging operations (Telegram)
Governance of a messaging ecosystem: discovering what exists, organizing
it, and separating account authorization from credential registry.
Messaging operations (Telegram)
Governance of a messaging ecosystem: discovering what exists, organizing it, and separating account authorization from credential registry.
- The problem
- Channels, accounts, and bot credentials had accumulated across multiple personal accounts with no inventory of what existed or which code referenced which credential.
- What I designed and built
- Personal Telegram is a deployed web application for organizing channels across multiple personal accounts with account-aware synchronization and tagging. TGsecrets is a separately deployed registry that inventories Telegram bot credentials and their code references.
- Operational boundaries
- TGsecrets inventories credentials and their code references without presenting secret values. Account authorization and bot-registry concerns are kept separate.
- Evidence / scale August 2026
- Three connected accounts, 283 channels, eight tags, and 74 bots. These counts are aggregate evidence; identities and channel names are excluded.
- Status
- Operated · August 2026
- Technologies
- Telegram client integration · account-aware synchronization · tagging · credential registry
Private remote operations (Invidtiv)
A central control surface for selected private-network tools and
services over Tailscale, replacing the need to remember individual addresses and
runtimes.
Private remote operations (Invidtiv)
A central control surface for selected private-network tools and services over Tailscale, replacing the need to remember individual addresses and runtimes.
- The problem
- Private-network tools and services each required remembering individual addresses and runtimes.
- What I designed and built
- An operated command-center dashboard linking private tailnet tools and selected control functions, providing a central entrance to services that would otherwise be reached individually.
- Evidence / scale
- Operated at the August 2026 snapshot. Private network and host details are omitted from this portfolio.
- Status
- Operated · Invidtiv Inactive · SSH Bridge
- Known limitations
- The related SSH Bridge — a FastAPI service, web dashboard, and OpenSSH server for managing tunnel access on the private network — was deployed previously and is currently inactive. Deployment artifacts and persistent state remain, but no containers were running at the snapshot. The combined status is therefore component-specific.
- Technologies
- Tailscale private networking · dashboard control plane · FastAPI · OpenSSH
Additional working tools support the systems above: a report-viewing utility (PageServer, a Flask-based library for viewing agent- and project-generated HTML, Markdown, and image artifacts through a consistent browser interface), private text-to-speech interfaces (web operated; Android a buildable prototype), and a previously deployed SSH tunnel manager (inactive at the August 2026 snapshot). Details available on request.
Additional projects
Prototypes, prospective-client work, and experiments
Two-way spoken interaction over a private network
(Hermes Android Voice)
A Hermes gateway plugin and Android companion enabling two-way spoken
interaction over Tailscale, without using the telephone network.
Two-way spoken interaction over a private network (Hermes Android Voice)
A Hermes gateway plugin and Android companion enabling two-way spoken interaction over Tailscale, without using the telephone network.
- What it does
- Connects a Hermes gateway plugin to an Android companion for spoken interaction over the private network. Hermes is credited as the upstream agent runtime.
- Evidence
- 85 automated tests passed, with active private-network voice and control listeners at the snapshot.
- Status
- Prototype · August 2026
- Known limitations
- It remains a working prototype because operation used a development launcher, and no completed phone pilot or soak test was evidenced.
Operating systems for two independent service businesses
Two anonymized prospective-client projects translating owner-operated
business requirements into coherent customer and administrator workflows.
Operating systems for two independent service businesses
Two anonymized prospective-client projects translating owner-operated business requirements into coherent customer and administrator workflows.
- The problem
- Two independent service professionals needed more coherent digital operations. Customer discovery, bookings, follow-up, content creation, publishing, and administration were split across manual processes and disconnected tools, with the same person responsible for delivering the service, attracting customers, handling bookings, and running daily administration.
- What I designed and built
- The Client Operations Platform combined customer discovery, booking, client records, scheduling, reminders, communications, and privacy workflows — including consent handling, data export, deletion workflows, access controls, and audit records around sensitive operations. The Content and Booking Platform combined a public conversion site with an administrator workspace for retreats, inquiries, campaigns, content, scheduling, media, and analytics.
- My role
- Translated each prospect's needs into an operating model, designed customer and administrator workflows, directed AI-assisted implementation, integrated the major components, reviewed and hardened the resulting systems, and documented their deployment requirements and limitations. Clients contributed discovery input and domain needs; open-source frameworks and external services supplied foundations.
- Operational boundaries
- AI could assist with content drafting, summarization, image generation, and other preparatory work, but content passed through approval before scheduling or publication. Business owners retained authority over bookings, generated content, external-service configuration, and anything sent or published.
- Status
- Completed — not deployed
- Known limitations
- Both systems reached completed prototype or MVP stage but were not commercially deployed, because the engagements did not proceed for non-technical reasons. Telegram was the only implemented live publishing path documented in the archive; LinkedIn, Instagram, and Facebook remained manual or mocked pending real integrations and external approvals. Neither platform received a formal security audit or completed real-world production validation. Their safeguards are implementation evidence, not a claim of compliance or operational reliability.
Level 3 — verification
The Client Operations Platform currently passes 34 frontend tests and 24 backend tests, and its production build completes successfully.
The Content and Booking Platform archive contains 40 backend, 26 administrator-frontend, and one public-frontend test files, together with Docker deployment assets and a local-usage completion audit. I did not rerun that archived suite during this review, so archive-reported completion is distinguished from independently verified current results.
Testing the limits of agent autonomy
Clawteam Agent Squad and the QuietSystems/Paperclip Incubator explored
agent-run development organizations with explicit work-control planes and Kanban
sources of truth.
Testing the limits of agent autonomy
Clawteam Agent Squad and the QuietSystems/Paperclip Incubator explored agent-run development organizations with explicit work-control planes and Kanban sources of truth.
- What they produced
- Meaningful orchestration designs, role models, and operating documentation.
- What they exposed
- Unresolved runtime and governance gaps, including enforcement, retry, QA, and synchronization concerns.
- Status
- Experiment · learning artifact
- Known limitations
- Presented as partial experiments and learning artifacts, not autonomous production organizations. Their value lies in tested boundaries and lessons, not claims of unattended delivery. Included here as evidence that I test the limits of agent autonomy before recommending it.
Security and governance
Authority, data movement, and failure behavior, considered from the start
I think explicitly about authority, data movement, credentials, review points, and failure behavior when designing automation.
Security is treated as concrete boundaries rather than a label. The recurring practical controls across the newer projects:
- Role-based access
- Read-only integrations
- Scoped credentials
- Private Tailscale networking
- Signature verification
- Provenance
- Audit records
- Automatic offboarding
- Restricted agent interfaces
- Human approval for consequential actions
- Fail-closed matching
The portfolio also records limitations directly: inactive components, external model use, missing deployment validation, failed checks, mocked publishers, and experiments that did not reach production.
Several hundred automated tests pass across the portfolio's operated systems at the August 2026 snapshot. Per-project counts appear in each case's Level 3 evidence, alongside recorded failures. Several systems have not undergone independent security audits; that limitation stays visible in the status fields where material.
Career
Keeping real businesses running since 1997
- 2023 — present Independent systems builder Personal portfolio. Carried out alongside employment until April 2026, and full-time since.
- Feb 2024 — Apr 2026 Call Center Team Leader Teleperformance, Lisboa. Led a team of 28 customer service agents: daily operations coordination, performance and service-quality monitoring, coaching and feedback, and support in resolving escalated customer cases.
- Feb 2023 — Feb 2024 Call Center Agent Teleperformance, Lisboa.
- Nov 2017 — Sep 2022 Systems Engineer and Administrator Private venture. Planned, engineered, and ran a 50-GPU mining operation: physical build-out, Linux administration, MySQL, and API-based data collection for monitoring.
- Dec 2017 — Jul 2018 Information Security Consultant (ISO/IEC 27001:2013) Global Score Lda., management consultancy.
- Mar 2012 — May 2017 Systems Administrator / Systems Engineer Ative Lda., industrial health and safety, 34 people. Windows Server, SQL Server, SSRS, Primavera ERP, networking, helpdesk.
- Jun 2007 — present Systems Administrator (ongoing support) Cópia Igual Lda., printing shop.
- Jan 2006 — present IT Freelancer Systems and server upgrades, networking, general IT.
- Jan 1997 — Sep 2011 Customer Service and Print Production CIAU Lda., printing shop. AutoCAD and Adobe professional tools.
My employment history and my systems work ran in parallel. The roles paid the bills and kept me close to real operations — customers, escalations, team performance, daily service pressure. The systems in this portfolio are what I built alongside and after those roles, and they are the demonstration that the capability outgrew the jobs. Since April 2026 I do this work full time, as a freelancer under Quiet Systems and, for the right team, in a role where the two converge.
Internal tools, data, and processes at Teleperformance are not described here or in the portfolio, in line with confidentiality obligations.
Training — ISO/IEC 27001:2013 Information Security (IPAC, 2017) · Microsoft MCSE/MCSA and Cisco CCNA (Galileu, 2006) · Audiovisual Technician (2001). Languages — Portuguese native · English C2 reading and listening, C1 writing · French C2 reading and listening, B2 writing. CEFR self-assessment.
Where this fits
I learn workflows quickly and build around what a business already relies on
I learn workflows quickly, find where the real operational problems are, and build careful, useful improvements around the tools a business already relies on. If you run a business and recognise this kind of work, Quiet Systems is how I take it on: a scoped audit first, then the system. If you are hiring for a team where this converges — operations, internal tools, and an AI toolkit that needs a human owner — that is the skill that transfers directly.