Agentic EMS

Demo

About

Institutional-grade DeFi lacks a unified intelligence and execution layer. While traditional fixed-income traders rely on Bloomberg to monitor macro liquidity, assess underlying credit risks, and execute structured instruments, DeFi equivalents are forced to manually cross-reference fragmented data silos—DefiLlama for TVL, L2Beat for roll-up security, governance forums, and disparate on-chain analytics. This operational fragmentation makes sizing large allocations across liquidity pools, money markets, and yield protocols inefficient and structurally risky and time consuming. The Solution: welcome the Agentic Execution Management System (EMS) This project rebuilds the Bloomberg terminal kind of experience for fixed-income and yield-focused DeFi allocators, powered by an autonomous, dual-agent architecture accessed through a natural language interface: Macro & Market Ingestion: An automated ETL engine aggregates real-time macro indicators across DeFi sectors ( using theGraph, defillama and other data providers): scraping TVL flows, fee yields, protocol governance shifts, and underlying L1/L2 systemic risk vectors. - Predictive Intelligence: By pairing transformer-based language reasoning with time-series forecasting models, the system evaluates protocol counterparty risk and models multi-chain yield projections. Institutional-Grade Execution The agent synthesizes trader intent into an optimized execution strategy, deconstructing large capital allocations into multi-leg routing transactions using bridging solutions like CCTP, lifi, layerzero, OneInch SwapVM to minimize slippage and MEV exposure and can convert positions to stablecoin in less time risk. Continuous Portfolio Management: Once on-chain legs settle, the platform tracks positions autonomously, surfacing dynamic hedging recommendations and automated rebalancing alerts.

How it's made

The platform is organized within a modular TypeScript-first Turborepo monorepo comprising five distinct applications: - The Trading Desk UI: A high-performance dashboard that guides allocators through intent capture, visual portfolio payoffs, quantitative risk simulation parameters (vega, alpha, beta, gamma which are generated by the AI agent pipeline) and live execution tracking across categories of DeFi protocols (LP, lending, prediction markets, etc…). Serverless Inference Microservice: A dedicated engine hosted on Google Cloud product running multi-agent workflows that stream typed, step-by-step reasoning events directly to the interface. - Predictive Risk & Indexing Service: An analytical back-end integrating temporal data retrieval, automated market indexing, and quantitative yield projections. - Execution & Settlement Engine: A venue-agnostic transaction compiler that turns strategic goals into concrete on-chain routing instructions across standard bridging protocol (layerZero), standard bridge aggregators (Lifi), and CCTP. - Chain Validation Harness: An automated environment that tests and verifies execution scenarios against live network state forks to ensure deterministic gas fees and possibility of other challenges like MeV sandwich attacks, etc. Partner Integrations: * Data & Risk Aggregation: Instead of building ad-hoc indexers, we leveraged The Graph alongside Messari’s standardized subgraphs across ~200 deployments to normalize liquidity, lending, and derivative protocols. Macro risks and protocol vulnerabilities are sourced directly from DeFiLlama, L2Beat, and governance hubs. * Sovereign Custody & Approval: We combined Privy (embedded wallets), Ledger (hardware clear-signing), and Safe (multi-sig smart accounts) so institutions retain total sovereignty: agents propose allocations, but devices approve them. - Cross-Chain Settlement & Liquidity: Circle Arc and CCTP v2 provide native cross-chain settlement with zero bridge friction, while 1inch and Uniswap v4 serve as our programmable routing layer for deep liquidity and custom pool hooks. Inference layer: it runs on Google Vertex AI based Gemini-3.1 / 2.5 models for multi-step reasoning, orchestrated with LangGraph (as the multi-agent system doing custom analysis across the DeFi data analysis pipeline and order execution setup based on the risk parameters), traced via LangSmith (in order to provide the observability), and paired with TimescaleDB and Google’s TimesFM3 open source for continuous time-series yield forecasting. Resolutions to the challenges I faced during developing the app: - Bypassing Ingestion Bottlenecks: Institutional governance forums and analytics dashboards aggressively rate-limit scrapers. We engineered a stealth Python worker running specialized browser virtualization to scrape critical governance signals, backed by cross-language schema tests to prevent data drift. - Robust error / hallucination handling: That these AI / RAG agents are famously bad news when it comes to hallucination and lack of attribution of data. We have been very strict in terms of detecting all those generated hallucinated patterns, and we have defined very strict guardrails. Either the AI agent will generate the correct response (that is, the predictions for investment into a strategy) or generate the ER EIP-712 transaction typed data structure correctly, or it will give null as an error. We handle the AI agent orchestration and its reliability very seriously. - Simulation of the APY’s of yields: To provide more comprehensive analysis for the fixing traders to define their backstops and other rebalancing strategies. The Times FM3 model provides them with the possibility to predict the potential API/APY yield for the future. It is a state-of-the-art model that can start predicting any time series data, provided it is sufficiently fine-tuned, starting from 10 time series data points of lending protocols, it will be able to do a simulation perfectly.