← Selected work

Product engineering · Engineering systems · DevOps

Document AI

Engineering the product—and the system used to build it reliably.

Document AI is a multi-tenant document platform associated with My Talents Academy. My contribution covered the product surface and the engineering operating model used to keep a complex codebase coherent.

Organization
My Talents Academy
Role
Full Stack & DevOps Engineer / Technical Lead
Team
2 engineers · Nov 2025 — Sep 2026
Scope
Monorepo · 10 backend apps · 6-role dashboard · 586 test files

System overview

Architecture — simplified
HTTPHTTP1 · upload2 · extraction3 · review, if requiredfilesworkflowhuman decisiontask queuesactivityresults · review taskMulti-role interfaceNext.js · 6 rolesServer-side BFFSession · real-time eventsAPI gatewayAuth · tenant and role scopeDocument serviceUpload · versionsExtraction serviceAdmission · quotas · resultsReview serviceCorrect · validate · rejectAlfrescoSource files per tenantTemporalDurable orchestrationPostgreSQLState · results · decisionsAI extractionAdditional verificationWorkerActivities · retriesAudit trail · every stepBusiness eventsFrom every serviceOutboxPostgreSQL · retriesAudit serviceIdempotent processingProtected logAppend-onlyHTTPHTTP123filesworkflowtask queuesactivityhuman decisionMulti-role interfaceNext.js · 6 rolesServer-side BFFSession · real-time eventsAPI gatewayAuth · tenant and role scopeDocument serviceUpload · versionsAlfrescoSource files per tenantExtraction serviceAdmission · quotas · resultsTemporalDurable orchestrationWorkerActivities · retriesAI extractionAdditional verificationReview serviceCorrect · validate · rejectPostgreSQLState · results · decisionsAudit trail · every stepBusiness eventsFrom every serviceOutboxPostgreSQL · retriesAudit serviceIdempotent processingProtected logAppend-only

Human review is conditional. The audit trail covers every step. Ten backend applications in one Nx monorepo, built by a team of two.

A complex product needed an explicit way of being built.

The platform spans secure document intake, storage, AI-assisted extraction, OCR challenge, human review, notifications, audit, and governed data-handling workflows.

As the product grew, the risk was not only feature complexity. A large codebase, a small team, coding agents, and weak conventions can create architectural and quality drift across every layer.

Turn recurring decisions into repository-level systems.

I progressively formalized frontend and backend conventions, cross-stack contracts, testing expectations, security and documentation standards, and repository-specific guidance for developers and coding agents.

Recurring implementation failures became inputs to a learning loop: identify the pattern, write the rule, add validation, and reduce the chance that the same mistake must be rediscovered.

Validation across code, integration, delivery, and the real environment.

The quality model combined unit and integration coverage, Playwright end-to-end checks, CI gates, deployment readiness checks, and post-deployment validation using isolated synthetic resources where appropriate.

  • 586 test files, with end-to-end projects for nine of the ten backend apps and for the frontend
  • 12 versioned instruction files for coding agents (AGENTS.md and skills)
  • 5 CI/CD workflows, automated versioning, and 19 tagged releases
  • Append-only audit log fed by an outbox with idempotent retries
  • SHA-256 document integrity, AES-256-GCM secret encryption, HMAC pseudonymization, and retention holds
  • Reusable, accessible, bilingual, white-label frontend conventions

Make engineering knowledge useful beyond one feature or one person.

The result was not a single framework or checklist. It was a connected set of contracts, components, tests, documentation, release automation, and deployment practices that could continue guiding work as the product and team evolved.

Parts of that engineering model became useful references for additional repositories, which is the kind of leverage I aim for: solve the product problem and improve the system used to solve the next one.

+ Engineering principles

What the system is designed to preserve.

01

Contracts over ambiguity

Make cross-stack expectations explicit before drift reaches production.

02

Learning loops

Convert recurring implementation mistakes into durable repository rules.

03

Layered validation

Separate code checks, integration proof, deployment gates, and live validation.

04

Human review

Use coding agents as leverage inside clear standards and review boundaries.

Next case studyOCD · Bovann