Headline
We built an autonomous AI execution layer for an institutional fixed income trading platform. A trader defines one high-level instruction, and an AI engine powered by OpenAI GPT-6.1 runs the full RFQ (request for quote) lifecycle: negotiation, counter-offer evaluation, multi-venue order routing and real-time reporting. A guardrailed state machine enforces execution rules and compliance, and every AI decision is written to an audit trail.
At a glance
| Asset class | Fixed income (bonds by CUSIP) |
|---|---|
| Venues | Trumid, Tradeweb, ICE BondPoint |
| AI | OpenAI GPT-6.1, LangChain, LangGraph |
| Autonomy modes | AUTO_EXECUTE or CONFIRM_FIRST, per instruction |
| Evaluation cycle | 2 seconds |
| Hosting | AWS, Kubernetes, Docker |
The problem: Manual dealer negotiation across venues
Dealer negotiation across Trumid, Tradeweb and ICE BondPoint was manual. The engine replaces it.
What we built: A headless AI service for RFQ execution
Autonomous RFQ engine
The engine sits between the trader-facing web application and the execution venues as a headless AI microservice. It negotiates counter-offers using OpenAI GPT-6.1 with live market context.
AI advice, deterministic control
The LLM advises on pricing and venue splits. A guardrailed state machine enforces execution rules, audit trails, expiry and compliance, with full graceful degradation to rule-based pricing when the AI is unavailable.
Multi-venue execution router
A LangGraph-based RFQ hybrid workflow merges live quotes from initiated RFQs with open-market RFQs to identify the optimal execution candidate across venues. An LLM-based multi-venue execution router splits order quantity across ECNs based on market depth, venue characteristics and best-execution requirements.
Plain-English data queries
Traders and compliance officers query 26 trading data domains in plain English, with chart output and no SQL or BI tooling. The domains include trade blotters, P&L reports, TRACE data, reconciliation records and ECN trade history.
On-demand bond research
An AI research service acting as a Senior Fixed Income Research Analyst writes structured credit briefings for any CUSIP, ISIN or ticker. Each briefing covers ratings, coupon, maturity, TRACE guidance, issuance curve and curated regulatory links, and is cached in Redis for 3 hours for desk-wide reuse.
How it works: One instruction to an executed trade
The trader submits one instruction: the bond (CUSIP), side, quantity, target venues, price range and autonomy mode.
The engine fans out RFQs to multiple electronic communication networks (ECNs) at the same time.
It polls dealer quotes and evaluates counter-offers against live market data on a 2-second evaluation cycle.
In AUTO_EXECUTE mode it executes on the best qualifying quote within that cycle, without human intervention. In CONFIRM_FIRST mode it surfaces its recommendation and waits for the trader's approval before placing any order.
Every AI action (RFQs sent, quotes evaluated, counter-offers accepted or declined, pricing rationale) and every execution outcome is written to MySQL through stored procedures and broadcast in real time over Redis pub/sub to the trader's screen via WebSocket, providing complete auditability.
Technology: Python services on AWS
| Language and API | Python 3.12, FastAPI |
|---|---|
| AI | OpenAI GPT-6.1, LangChain, LangGraph |
| Data and messaging | MySQL (stored procedures), Redis (cache, pub/sub), RabbitMQ |
| Trader updates | WebSocket |
| Deployment | Cloud-native on AWS with Kubernetes and Docker; deployment guidance covers sticky-session ingress |
| Secrets | AWS Secrets Manager |
| Observability | Structured logging to OpenSearch for full operational observability |
