An open-source, multi-agent orchestration framework (AGPLv3) designed to automate the cognitive load of engineering management and bridge the gap between high-level strategy and day-to-day execution.

• The Architecture: Engineered the "Hermes" orchestrator, leveraging the Model Context Protocol (MCP) and Mem0 to create a persistent, context-aware RAG brain over the repository's corpus. Secured the agentic network layer using NVIDIA OpenShell (NemoClaw) for strict, deny-by-default egress hardening, ensuring safe autonomous tool execution while routing executive alerts via Telegram.
• The Breakthrough: Designed autonomous, background-running loops for executive oversight. Fused a continuous technical debt auditor (GraphRAG + SonarQube) that routes context-rich tickets into Linear. These tickets feed directly into HumanLayer for human-in-the-loop (HITL) approval on brownfield maintenance, establishing the architectural groundwork for future OpenHands integrations to drive fully autonomous greenfield development.
• Stack: Python, GraphRAG, Mem0, MCP, NVIDIA NemoClaw, HumanLayer, OpenHands, Stripe API, Linear API, SonarQube, Telegram API.

A multi-agent LangGraph pipeline that turns unstructured threat-intel text into a MITRE ATT&CK-grounded attack graph and a structurally validated defensive configuration — engineered to demonstrate the full shape of a production multi-agent system, not a single prompt in a chat loop.

• The Architecture: Engineered a five-node guard_input → retrieve → extractor → graph_architect → defensive_guardrail state graph over a hybrid RAG retriever — BM25 + dense Chroma fused via weighted Reciprocal Rank Fusion, then cross-encoder reranked — grounding every extracted technique and mitigation in a live ~700-record MITRE ATT&CK corpus. Layered in hosted Mem0 long-term memory, a Guardrails AI schema-validated output gate, and a prompt-injection input classifier, all fail-open by design, with full Langfuse tracing across every node.
• The Breakthrough: Diagnosed and fixed a retrieval-diversity collapse — a single-query cross-encoder rerank was silently discarding multi-stage kill-chains down to one dominant technique; reordering the pipeline so RRF fusion owns recall/diversity and the cross-encoder only reranks a wide window restored full multi-tactic extraction. A Langfuse dataset + SDK-evaluator experiment then measured it: 1.00 mitigation faithfulness, 0.71 technique-id F1 against a hand-labeled eval set — shipped through three interchangeable UIs (Streamlit, React/Tailwind, Open WebUI) off one backend.
• Stack: Python, LangGraph, FastAPI, Chroma, BM25, Sentence-Transformers cross-encoder, Mem0, Guardrails AI, Langfuse, React, Vite, Tailwind CSS, Open WebUI.

An end-to-end open-source ML pipeline built on the ESA Anomaly Dataset (ESA-AD) that combines time-series anomaly detection with LLM-generated diagnostic reasoning.

• The Architecture: Engineered a stacked ensemble fusing classic LSTMs, fine-tuned text LLMs (Qwen3-8B via QLoRA), and vision transformers (Qwen3-VL-8B scanning rendered telemetry PNGs) to drive automated satellite health monitoring.
• The Breakthrough: Conducted a comprehensive 13-approach benchmark showing that while fine-tuned LLMs alone struggled with over-flagging calibration artifacts, a multi-modal fused stacker achieved a state-of-the-art $CEF_{0.5}$ score of 0.781—proving that ensemble detection coupled with LLM advisory logic yields the highest operational reliability.
• Stack: Python, PyTorch, QLoRA, GGUF/llama.cpp (Local Metal Inference), Scikit-Learn, Time-Series LSTMs.

A from-scratch Vision Transformer built to demonstrate raw mechanism engineering and Explainable AI (XAI), bypassing black-box weights to regress GOES soft-X-ray flux directly from 7-channel SDO/AIA solar EUV imagery.

• The Architecture: Engineered a mechanism-faithful reproduction of the FOXES model (Goodwin et al. 2026). Bypassed standard CLS tokens in favor of an intrinsic spatial-XAI head featuring an 8×8 patch embed and a 9×9 inverted non-local attention mask.
• The Breakthrough: Solved the vision black-box problem by ensuring every prediction serves as its own attribution map, where per-patch predictions sum exactly to the global flux. Achieved a Pearson correlation of $r = 0.943$ and an MAE of 0.368 dex (outperforming the baseline by ~47%), while proving extreme compute efficiency by training on a single local RTX 4090 in just 17 minutes.
• Stack: Python, PyTorch (Custom ViT Architecture), Explainable AI (XAI), SDO/AIA Imagery.

🏆 Top 25% @ Global South AI Safety Hackathon. A fork of LatticeFlow/ETH/INSAIT's COMPL-AI that turns an EU-AI-Act evaluation suite into the first compliance benchmark for a Global-South AI statute — Brazil's AI bill, PL 2338/2023 (Senate-approved Dec. 2024).

• The Architecture: Preserved all 30 original EU benchmarks on Inspect AI and layered on five Brazil-specific benchmarks mapped to the bill's Chapter II rights — AI disclosure (Art. 5, I), non-discrimination across IBGE racial/regional/intersectional categories (Art. 5, III), the full high-risk rights triad of explanation, contestation, and human review (Art. 6, I–III), and the Algorithmic Impact Assessment (Arts. 25–28) — scored with deterministic, multilingual (pt-BR/EN) rubric detectors (no LLM judge).
• The Breakthrough: Designed a same-model EU↔Brazil delta (two benchmarks reuse the exact same scorer, isolating language + legal framing from raw model strength). Across six models from five developers (8B → frontier), all six disclosed being an AI ~95–100% of the time in English but only ~50–55% in Portuguese under Brazilian law — a ~0.45 compliance gap invisible to any English benchmark — packaged as a per-article HTML scorecard that doubles as the Art. 28 "public conclusions" artifact.
• Stack: Python, Inspect AI, Anthropic API + Ollama (local, $0), deterministic rubric scorers, pandoc/XeLaTeX reporting.

A red-teaming sandbox demonstrating that a handful of dormant, benign-looking documents seeded into a RAG vector store act as an inference-time sleeper agent—invisible until a trigger query retrieves them and hijacks the model's output, with no weight access required. Built for the Global South AI Safety Hackathon (Latin America · Technical Safety sub-track).

• The Architecture: Authored a new NVIDIA garak rag_poisoning probe + paired detector driving a live PostgreSQL pgvector(384) + HNSW-cosine RAG victim over both in-process and WebSocket transports (reusing my merged garak PR #1379). Seeded with Latin-American poison derived structurally from the SESGO Spanish stereotype benchmark plus a Spanish–English code-switching evasion variant.
• The Breakthrough: Across six bias dimensions and two languages, 24 dormant poisons among 224 documents reached 100% attack success; adding an insertion-time BEFORE INSERT quarantine gate dropped it to 0% with no change to the attack—a two-sided attack-and-defense result instantiating "audit memory writes before they hit the retrieval hot path."
• Stack: Python, NVIDIA garak, PostgreSQL/pgvector, HNSW, sentence-transformers (all-MiniLM-L6-v2), Ollama (local LLM target), Matplotlib, pytest.

• NVIDIA/garak • PR #1379: Architected and implemented the core WebSocket generator module using the websockets library, expanding LLM vulnerability scanning capabilities to support real-time, bidirectional chat architectures with full authentication handling.

• JoshuaC215/agent-service-toolkit • PR #258: Resolved a critical asynchronous streaming exception (RuntimeError: generator didn't stop after athrow()) occurring when Claude models yield empty string tokens immediately prior to tool execution payloads, stabilizing production UI/Streamlit integrations.

Author / Lead Architect * Designed and documented the technical framework for an institutional valuation platform tracking complex transaction metrics, regulatory risk vectors, and market datasets within the healthcare sector.

• Outlined quantitative methodology for normalizing disparate healthcare operational matrices into deterministic financial evaluation paths.