NumiaAI

Introduction

Launched with dYdX, NumiaAI turns fragmented on-chain telemetry into natural-language answers for protocols and institutions. The objective was simple and hard: make institutional-grade blockchain analytics directly queryable, reliably fast, and cost-controlled. I led the data strategy behind this—pipelines, semantic modeling, and the AI controller that orchestrates model access—so the frontend stays thin and operations stay predictable. We standardized metric definitions across chains, ensured freshness where it mattered, and built guardrails that keep large-model variability in check. The result is a system that behaves like a product, not a demo: 99.9% uptime, p99 < 200 ms, and thousands of queries/day across datasets.

The Challenge

Bridging raw blockchain data with LLMs demanded hard choices: normalize datasets without losing nuance; balance freshness against cost; and deliver sub-200 ms responses while constraining model unpredictability. We needed deterministic, audit-friendly answers from probabilistic systems, clear semantic definitions, and scalable access patterns that wouldn't collapse under real usage.

Solution & Approach

We prioritized a batch-first, semantics-driven architecture: BigQuery + DBT pipelines to normalize and pre-aggregate; a semantic layer to enforce consistent definitions; and an AI controller that converts user intent into structured queries, routes models, and applies guardrails—so latency stays low, costs stay bounded, and results stay consistent.

Data & Pipeline Layer

• BigQuery as the warehouse for normalized telemetry
• DBT incremental models and materializations for fast, cost-efficient reads
• A semantic metrics layer standardizing token/liquidity/governance definitions
• Optimized SQL models that serve query-ready datasets to the controller
• Continuous ingestion with freshness SLAs on time-sensitive tables

AI Controller & Backend Endpoints

• Controller layer consuming DBT outputs and generating structured queries for LLMs
• Low-latency backend endpoints (p99 < 200 ms) returning deterministic, audit-friendly results
• Guardrails: prompt schemas, retries, fallbacks, and error handling to reduce variance
• Thin frontend abstraction so product velocity isn't blocked by backend complexity

Results & Impact

NumiaAI makes blockchain analytics accessible through a conversational interface backed by disciplined data and AI infrastructure. The platform maintains 99.9% uptime and p99 < 200 ms query latency, serves thousands of queries daily, and is adopted by institutional users and partner protocols.

Platform Achievements

• 99.9% uptime with reliable availability
• p99 < 200 ms on AI-backed query endpoints
• Thousands of queries/day across datasets
• Adoption by leading protocols and institutional partners

Technical Contribution

• Designed and optimized DBT pipelines and semantic models powering AI queries
• Built the AI controller and backend API that integrate GPT, Gemini and others
• Tuned BigQuery for sub-200 ms reads while controlling compute cost
• Established observability (monitoring/alerts) to keep reliability at 99.9%

NumiaAI has established a new category of AI-driven blockchain analytics, turning raw telemetry into decision-grade intelligence accessible via natural language. This foundation positions the platform for continued innovation and expansion in the blockchain data space.