How We Built a Zero-Hallucination MCP Rule-Verdict Agent in 6 Hours Without Writing a Single Line of Code

The Story: Zero Code, Maximum Engineering

When you participate in hackathons (and especially when you like to wrap them up in 48 hours max), the bottleneck is usually typing boilerplate, debugging weird syntax, or wrestling with unpredictable LLMs.

For this project—AERO-KIT, a deterministic MCP (Model Context Protocol) agent designed to solve complex rule contradictions using Sanity structured content—I didn't write a single line of code myself.

Instead, I acted entirely as the Chief Architect and Director. The real workforce? A meticulously orchestrated AI stack:

1. Gemini Chat as the core strategic brain (designing the architecture, orchestrating the execution steps, and maintaining strict technical guardrails).
2. OpenCode powered by Big Pickle as the local rapid-execution nano-engine, pushing out files, tests, and configurations at lightning speed.

How It Works (The Method)

Most AI agents fail at authority arbitration because they rely on fuzzy vector RAG (Retrieval-Augmented Generation). If a rulebook says 75°C and an errata says 60°C, a naive keyword search treats them as a tie and hallucinates a blended or wrong answer.

AERO-KIT solves this through structured data and rigid execution guardrails:

• Sanity Content Lake as Knowledge Base: Linked types (policyOrRule, errataOrAmendment, resolutionCase) where amendments explicitly point to the rules they override via targetRuleRef and a higher precedenceLevel.
• Deterministic MCP Integration: The agent connects to a Sanity Context MCP endpoint in GROQ mode.
• The Guardrails / Skills Methodology: Borrowing from internal engineering standards (testComplet, fixe-de-bug, and strict anti-hallucination gates), the agent is programmed never to guess. If a contradiction isn't formally arbitrated by a higher-precedence rule or a human record, it refuses to answer and states why.

The Proof

We didn't just hope it worked. OpenCode and Gemini built a rigorous test suite (node --test) resulting in a 9/9 green test bench:

• TEST 1: Univocal queries return precise normative values with Sanity ID citations.
• TEST 2: Contradictions are formally resolved using precedence levels (3 > 1).
• TEST 3 (Anti-RAG): Proves that text-only search fails while structured graph traversal wins.
• TEST 4 (Guard-rails): Guarantees zero hallucinations and correct handling of out-of-bounds or ambiguous questions.

Every execution creates a verifiable CLI session transcript (TRANSCRIPT.txt) proving that tools like initial_context, schema_explorer, and groq_query were actually utilized.

What We Learned

1. Architecture beats parameter size: You don't need a massive, expensive cloud model to build sophisticated software. A fast, lightweight local nano-engine guided by precise procedural skills (testComplet) outperforms an unguided frontier model any day.
2. Prompt-driven development is real engineering: Directing an agent like an engineering team leader—giving it strict error boundaries, mandatory test-driven validation, and zero-tolerance policies for hallucinations—turns AI into a true force multiplier.

---

Built with Gemini Chat, OpenCode (Big Pickle), and Sanity. Because great engineering is about orchestrating systems, not fighting syntax.