# Thiago Goncalves

> Senior Engineer · AI Platform & Security · Builder-Operator. Based in Apollo Beach, FL.

Senior software engineer with 20+ years building and securing production systems. I work at the intersection of AI platforms, security, and developer tooling — designing the infrastructure around AI models (persistent memory, evaluation boundaries, guardrails, credential safety, auditability) so they hold up in production, under scrutiny, and against misuse. Founder-operator running the full vertical: bare-metal local inference on a self-hosted stack (Ollama on TrueNAS SCALE), Model Context Protocol middleware, IDE-resident extensions, marketing surface, payment integration, and the IP work underneath. Background in offensive security (CEH, TryHackMe rank #231 US) shapes how I design AI systems — adversarial conditions assumed, trust boundaries explicit, failure modes engineered for, not hoped against. Self-taught.

## Contact

- Website: https://thiagogoncalves.com
- Email: thiago@hackerware.com
- hackerware.com: https://hackerware.com
- linkedin.com/in/thiagogoncalvesofficial: https://www.linkedin.com/in/thiagogoncalvesofficial/
- Languages: English, Português

## Highlights

- **96.4%** (token reduction at n = 2,189): Continuity's bounded-retrieval middleware enforces a hard ceiling on per-turn decision injection (default top-K = 3, hard cap = 10, per-record character caps). Worst-case per-turn cost ≈17,500 tokens regardless of corpus size — against a static-embedding baseline that would require ~485,000 tokens at the same decision count. Local Sentence-BERT embeddings (Xenova/all-MiniLM-L6-v2 on ONNX Runtime), sub-15ms retrieval, zero telemetry.
- **Oct 2025** (shipped before the platforms): Published Continuity to the VS Code Marketplace in October 2025. Anthropic shipped a near-identical Session Memory architecture for Claude Code months later — independent validation of the design.
- **5 layers** (defense-in-depth credential safety): Designed and shipped a five-boundary credential-scrubbing architecture for memory-augmented AI assistants: redaction at the MCP input boundary, at the AI-generated output boundary, at the persistence write sink, again at the persistence read path, and at the MCP tool-result return. 27 provider-specific patterns plus a Shannon-entropy fallback empirically calibrated to 4.5 bits per character. Measured 100% recall on the positive corpus; 0% high-confidence false positives on ~1,200-string negative corpus. Idempotent scrub primitive verified across all 27 patterns in a 56-test suite.
- **RedArchives** (tamper-evident archive): Evidence preservation for human rights documentation. Cryptographic fingerprinting, blockchain-anchored provenance, explicit threat model covering insider risk, post-hoc tampering, chain-of-custody disputes, and decade-scale temporal attacks.

## Experience

### Founder & Technical Lead, Hackerware LLC (Sep 2022 - Present, Remote)

Security-focused applied AI lab. Builder-operator across architecture, implementation, deployment, and iteration — bare metal to paid Pro tier on the VS Code Marketplace.

**Continuity — Persistent memory and runtime governance for AI-assisted development.**

- Identified the core failure mode: AI assistants produce inconsistent output because architectural decisions, constraints, and rationale are lost between sessions. Treated context as infrastructure, not convenience.
- Designed a bounded-retrieval middleware that decouples per-turn token consumption from total stored knowledge. Hard-ceilinged top-K (default 3, max 10) with per-record character caps. At n = 2,189 decisions, worst-case per-turn injection is ≈17,500 tokens against a static-embedding baseline of ~485,000 — a ~96.4% reduction that holds as the corpus grows.
- Shipped a five-layer defense-in-depth credential-scrubbing architecture for the Memory Amplifier threat — credentials that enter AI context once and get re-injected into every future session through persisted memory. Five enforcement boundaries (input, AI output, persistence write, persistence read, tool result) consuming a shared idempotent scrub primitive. 27 provider-specific patterns plus a Shannon-entropy fallback calibrated to 4.5 bits/char (measured 100% recall, 0% high-confidence FP).
- Built the MCP security interception layer that validates and gates tool execution before LLM-initiated actions reach the file system or external services — enforcement boundary, not policy document. Intercept → Pause → Evaluate → Enforce pipeline operating under a Compromised Context threat model.
- Designed human-in-the-loop correction model: corrections stored as scoped annotations with timestamps, not global overrides. Prevents contamination of unrelated future outputs and resolves conflicts deterministically (narrower scope wins within applicability window).
- Drafted the full patent portfolio — Patent 1 (bounded retrieval + Governance Lock), Patent 1 CIP (multi-signal relationship enforcement with Markov decision-chain prediction), Patent 2 provisional (defense-in-depth credential scrubbing) — including §101 Alice arguments, §112 enablement, and OWASP LLM Top 10 (2025) controls mapping across nine categories.
- Security-first defaults: 100% local storage, zero telemetry, on-device vector index. Develops against a self-hosted inference stack (Ollama on TrueNAS SCALE) using the same MCP protocol it ships to users on Anthropic, Gemini, or OpenAI — backend is swappable, application code is identical. Shipped to VS Code Marketplace October 2025; active paid users in production.

Stack: TypeScript · Node.js · MCP · SQLite · Sentence-BERT (Xenova/all-MiniLM-L6-v2 on ONNX Runtime) · Markov decision-chain · keytar (OS keychain) · multi-LLM (Anthropic / Gemini / Ollama) · VS Code Extension API · CLI

**RedArchives — Tamper-evident digital evidence platform.**

- Designed for environments where evidence must survive adversarial challenge, insider threats, and legal scrutiny over decades. Mission: preserve documentation of war crimes and human rights violations with cryptographic integrity that doesn't depend on trusting the platform operator.
- Built layered integrity architecture (Artifact Ingestion → Provenance Ledger → Metadata Store → Verification Layer → Presentation Layer) with explicit separation of duties between storage, verification, and presentation.
- Implemented blockchain-anchored cryptographic fingerprinting at ingestion; fingerprints cover normalized metadata, not just content, preventing semantic rewriting without detection. Designed for algorithm agility and re-verification paths to address long-term cryptographic decay.
- Modeled explicit adversarial threats: post-hoc tampering, chain-of-custody disputes, insider risk, temporal attacks, selective disclosure, platform trust collapse. Same integrity discipline informs how I design trust-critical AI systems (training data provenance, evaluation artifacts, feedback integrity).

Stack: Next.js · TypeScript · IPFS / libp2p (Helia) · Hyperledger Fabric · Anthropic Claude · Google Vision · Prisma · AWS S3 · FFmpeg · Kubernetes

### Penetration Tester (Independent), Quantum IT (Nov 2021 - Present, Remote)

- Full-scope penetration tests across networks, web applications, cloud infrastructure, and endpoints for SMB/enterprise clients. Tooling: Burp Suite, Nessus, Nmap, Metasploit, SQLMap, Kali Linux, custom scripts.
- Identified OWASP Top 10 vulnerabilities and high-risk findings (SQLi, XSS, auth bypass, broken authorization, API flaws, lateral movement, privilege escalation). Delivered executive risk reports with CVSS scoring, remediation guidance, and compliance alignment (NIST, ISO, HIPAA).

### IT Specialist, Independent Practice (2003 - Present, Remote)

Foundation for the operational instincts that shape my current security and AI platform work.

- Windows/Linux server administration, Active Directory, Group Policy, virtualization, hybrid cloud (AWS/Azure).
- Network architecture (VLANs, firewalls, VPNs), system hardening, patch management.
- Mentorship of junior staff.

## Core Competencies

- **AI Platform & Security**: persistent context and bounded retrieval; semantic retrieval (Sentence-BERT embeddings, ONNX Runtime, cosine similarity); Markov-chain modeling over decision-tag sequences; MCP tool interception and runtime governance; defense-in-depth credential safety (5-layer scrubbing pipeline); Memory Amplifier threat model for memory-augmented AI; OWASP LLM Top 10 (2025) controls mapping; self-hosted inference (Ollama on TrueNAS SCALE); multi-LLM orchestration (Anthropic, Gemini, OpenAI, local); human-in-the-loop correction without RLHF-style contamination; prompt injection / jailbreak / unsafe tool-use testing
- **Security Engineering**: threat modeling (Compromised Context, Memory Amplifier, adversarial input); offensive security (web/API/cloud, OWASP Top 10, MITRE ATT&CK); incident response & DFIR; EDR/SIEM (Microsoft Defender); identity (Entra ID / Azure AD); tamper-evident system design; provenance, auditability, chain-of-custody; cryptographic fingerprinting and blockchain-anchored integrity
- **Software & Infrastructure**: TypeScript, Node.js, Python; Next.js 15 (App Router), Radix UI, Tailwind, ShadCN; VS Code extension architecture, MCP middleware, language servers; Flutter / Dart (cross-platform mobile), BLE protocols, AR wearable SDKs; Prisma, PostgreSQL, Redis, SQLite; IPFS / libp2p, Hyperledger Fabric; AWS, Azure, DigitalOcean, Cloudflare, Docker, Kubernetes, CI/CD; Linux & Windows administration, virtualization, network segmentation; TrueNAS SCALE, self-hosted infrastructure
- **Communication & Ownership**: architecture decision records (dogfooded — Continuity logs its own); patent specification drafting (claim language, §101 Alice analysis, §112 enablement); technical writing (security threat models, white papers); runbooks, post-mortems, executive-ready risk reporting; founder-level scoping, vendor management, end-to-end delivery

## Credentials

- Certified Ethical Hacker (CEH) (EC-Council)
- Microsoft Cybersecurity Analyst Specialization (Entra ID · Defender · threat vectors · SC-900 prep)
- ISC2 Cybersecurity Specialization (2024)
- TryHackMe Leaderboard (Rank #231 United States · top 2,000 worldwide)
- Captain — Children of Exu (Huntress CTF · 2023)
- Patent Portfolio (in preparation) (USPTO · 3 applications: bounded retrieval, relationship enforcement (CIP), defense-in-depth credential scrubbing (provisional))