I build systems at the intersection of AI agents, cryptography, and confidential computing.

My background spans cryptographic hardware (PhD), SGX-based key management in production (Ava Labs, $2B+ in assets), and agentic AI systems built with LangGraph, CrewAI, and the Anthropic and OpenAI SDKs.

A LangGraph agent that calls Claude from inside an Intel SGX enclave. The API key never touches the host. Every response is cryptographically signed and independently verifiable via DCAP remote attestation.

What makes it real, not conceptual:

• ECDSA-P256 output signing: sign(SHA256(prompt ‖ result ‖ timestamp ‖ MRENCLAVE))
• DCAP quote binds signing key to enclave measurement via Intel's PKI
• Standalone verification CLI — validates the full trust chain without a running enclave
• Sealed storage using _sgx_mrenclave hardware key — no external wrap key after bootstrap

An autonomous CI test failure analyst built with LangGraph. Classifies failures, performs root cause analysis via tool-augmented reasoning, and generates structured reports. Features a reflection loop where a GPT-4o evaluator independently reviews every Claude triage classification before routing proceeds.

What makes it genuinely agentic:

• The reflection loop uses two different model families (Claude + GPT-4o) to reduce shared systematic bias — a single-model loop would approve its own wrong answers
• For flaky failures, the LLM decides whether to call get_test_run_history — querying pass/fail rates across 50 recent CI runs at the test level, not the workflow level
• Conditional routing means each failure type receives a specialized diagnostic prompt, not a generic one

Intel SGX · Gramine · LangGraph · Anthropic SDK · OpenAI SDK · DCAP attestation · Python · Azure DCsv3

LinkedIn · marcelokaihara.com