Vinicius Morais

Principal Engineer & AI Systems Architect

20+ years building regulated financial and federal software — ERISA retirement platforms, a 60-application USDA modernization portfolio — now building agentic AI: hierarchical LLM instruction architecture, evaluation gates and guardrails, and tiered-model orchestration.

Nashville, TN · M.S. Artificial Intelligence, CU Boulder · English / Portuguese

Home office with curved monitor, plants, and articulated lamp on a wooden standing desk

About

Most of my career has been spent in systems where being wrong carries a cost: qualified retirement plans governed by ERISA, federal agricultural platforms, and a real-time wagering platform built from zero. That work, much of it as a consultant, is about reading an ambiguous problem, choosing the framework that fits, and naming the tradeoffs before they become incidents.

I bring the same discipline to agentic AI. At NetImpact I designed an AI-assisted development governance framework (tiered model routing, auditable prompt-session logging) and presented it to VP-level engineering leadership so a security-cleared federal team could adopt AI in a controlled way. On my own time I build systems that put those ideas into practice: an on-device NLP pipeline with validation gates on every generated output, and a multi-runtime agent orchestrator whose memory is designed privacy-first.

I also make complex AI work understandable to executives: why a guardrail exists, what a tiered model buys you, and where a human still needs to make the call.

Experience

NetImpact Strategies — Principal Software Engineer

Federal (USDA/NRCS) · Jul 2025 – Present · GenAI/LLM, C#/.NET, AWS

  • Built AI-assisted .NET migration assessment tooling — a risk-scoring repository scanner (dotnetupdateutility) integrated into Jenkins, Playwright, and SonarQube pipelines — replacing manual, app-by-app evaluation with repeatable, portfolio-level risk classification across a 60-application legacy .NET portfolio targeting AWS via Terraform-based infrastructure patterns.
  • Designed an AI-assisted development governance framework and presented it to senior engineering leadership (VP and Senior Manager level), establishing tiered model routing and auditable prompt-session logging to enable controlled AI adoption in a security-cleared federal engagement.

Ascensus — Principal Software Engineer

Financial services · Oct 2021 – May 2025 · C#/.NET Core

  • Hired, mentored and led a seven-engineer team (four nearshore LATAM, two US-based), growing engineers into end-to-end owners of production features across EIA and PAS and setting shared design-review and delivery standards across both locations.
  • Architected and led Employer-Initiated Amendments (EIA), replacing a fax/mail-based qualified-plan amendment process with a digital self-service workflow — eliminating manual re-keying for online submissions and improving data accuracy and traceability against a published two-week processing window.
  • Led the technical assessment and modernization strategy for Unified Plan Document Fulfillment (UPDF), defining a .NET 8 / CoreWCF containerized architecture on Kubernetes that preserved existing service contracts and established the foundation for new external integrations.
  • Redesigned part of the Plan Adoption System's document-generation model with reusable dictionary components and Excel-based configuration shared across business units — enabling new document types and regulatory changes (including SECURE 2.0 and CARES Act) without modifying core application code.

Fanvest — Co-Founder & Chief Engineer

Fintech · Apr 2019 – Dec 2022 (concurrent venture)

  • Sole Chief Engineer for a real-time, WebSocket-based wagering platform, owning engineering end-to-end from architecture through two beta NFL product launches.
  • Grew to ~6,500 user signups against a ~7K registration ceiling while the company raised ~$375K; the company has since wound down.

Core10 / Insight — Lead & Senior Software Engineer

Fintech / enterprise (consulting) · Oct 2017 – Oct 2021

  • Client-facing delivery across fintech and enterprise customers; rapid prototyping and ramp-up across a broad set of stacks. For TechData, optimized build and deploy on a large monolith (MSBuild, Azure Pipelines).

Selected AI / ML Work

Sticky Insights

Browser-native applied ML / GenAI · TypeScript, Transformers.js, WebAssembly · 2026

A fully client-side NLP pipeline that turns a wall of sticky notes into named themes — zero server calls, no API keys, no backend.

Pipeline

MiniLM sentence embeddings → dual k-means / agglomerative clustering, selected by silhouette score → multi-signal keyphrase extraction → quantized small-LM (LaMini-Flan-T5) theme labeling.

Guardrails

A three-gate validation layer — degeneracy, cosine relevance, and vocabulary — screens every generated label before display and falls back to the top-scoring keyphrase on failure. Roughly 50–80% of generations pass on the first attempt.

Why it matters

Privacy by architecture: no text ever leaves the browser tab, and model weights cache locally for fully offline reuse. Built for UX research and regulated-data contexts — healthcare, finance, legal — where research data can't cross a network boundary.

View live →Source →

FINN: Portable Agent Orchestration Framework

Markdown-native agents, MCP, multi-runtime · 2026

An hierarchical orchestration framework in which a triage agent routes each request to one of eight domain specialists. 13 of its 14 skills declare their dependencies as MCP resources and tools.

Architecture

Thin adapter files keep all runtime-specific details out of the markdown core, so the same core runs unchanged in Claude Code, Claude Cowork and Cursor. Bindings for ChatGPT and a prompt assembler for local models (Ollama, LM Studio) are also documented.

Infrastructure

Designed context infrastructure where plain markdown/TSV files are the source of truth and SQLite (WAL mode, continuous Litestream backup) is a rebuildable index. Retrieval combines FTS5 full-text search, vector embeddings and a knowledge graph. Agent context stays human-readable, reviewable in git, and can be regenerated from the markdown at any time.

Privacy

Enforced a hard boundary between a public open-source framework repo and a private instance repo. Personal memory, databases, secrets and runtime state never enter version control. Model API keys are resolved from 1Password or Keychain at runtime, never stored in files. Outbound model calls can be held for manual approval and are recorded in a tamper-evident egress log. Pre-commit hooks block commits that fail the error-handling test harness or a byte-level healthy-path check.

Source →

Writing — Recursive Copilot Instructions

Substack · Feb 2026

A practical way to make AI coding assistants consistent across a team: structure one shared copilot-instructions.md as four layers (global invariants → architectural principles → domain context map → named pattern anchors), where each layer refines the one above and never repeats it.

Where it comes from

Grew out of applying MITs Recursive Language Models research to GitHub Copilot instruction design on an enterprise modernization engagement. I first shared it with engineers in an internal Copilot practice session, then wrote it up for teams adopting AI-assisted development.

Read the essay →

Education & Certifications

Get in touch

Open to principal and architect roles and advisory engagements in agentic AI, especially in regulated environments.

[email protected]Resume (PDF)LinkedInGitHubSubstack