Skip to content
P01 / Project

129BOB

A trading research system that runs at three speeds: AI agents decide what to study, a C++ engine enforces the limits, and an FPGA pipeline keeps the order book in hardware. Paper only.

129BOB in presentation mode: a price chart of the recorded MSFT order book on Nasdaq BX, 30 December 2019, with the approved price band and each paper fill and rejection marked on a timeline.
Screenshot · Presentation mode, scene 4 of 6Recorded data, paper only
Status
Active
Scope
Paper only: recorded market data, no money, no broker
Board
TUL PYNQ-Z2 (Zynq XC7Z020)
Data
Nasdaq TotalView-ITCH 5.0, recorded 30 Dec 2019
Built with
SystemVerilog, C++20, Python
Research
TradingAgents-Jev ↗
Source
Private repository for now
  • FPGA
  • RTL verification
  • Low-latency C++
  • AI agents
  • Market data
Write-up

Three speeds

Trading decisions happen on very different clocks. 129BOB gives each one its own lane, and no language model sits in the order path.

LaneReacts inJob
AI research (TradingAgents-Jev)minutesWhat to trade, the risk budget, when to stand down
C++ engineµs to msMarket data, order books, strategies, risk checks, orders
FPGA pipelinensFeed decoding and the order book in hardware

From research to a paper trade

  1. TradingAgents-Jev, our fork of the open-source TradingAgents framework, has a team of AI agents write market, sentiment, news and fundamentals reports, then argue them out into a rating.
  2. 129BOB keeps the reports as evidence and the rating as context only. A versioned policy can shrink the operator's limits, never raise them.
  3. A mandate sets the order size, price band, time window and loss limit. A person approves that exact version.
  4. The C++ engine replays recorded Nasdaq data through a reference order book and runs 24 risk checks on every order. Paper orders fill only against shares that were really displayed.
  5. A stop switch blocks new actions at once, and every step is stored so it can be inspected and exported.
Architecture diagram. Top: TradingAgents-Jev's analyst, researcher, trader, risk and portfolio manager agents produce a rating and research reports. Middle: a proposed Research Bridge, not yet built. Bottom: the 129BOB workstation turns the evidence into a mandate, takes human approval, runs the C++ engine on recorded data and executes on paper, with a stop switch and a decision trace.
From multi-agent research to a controlled paper trade. The Research Bridge in the middle is proposed and not built yet.

What has been checked

  • Reference model. A C++ decoder for all 23 ITCH 5.0 message types and an order-by-order book. It reconciles two recorded trading days end to end: 29,156,757 messages on Nasdaq BX and 268,744,780 on Nasdaq, with zero decode errors.
  • FPGA in simulation. The SystemVerilog parser and order book match the reference on every recorded test frame, at about 4.7 clock cycles per message in Verilator.
  • FPGA on a board. First run on a PYNQ-Z2 on 21 October 2026: the parser returned 28,021 of 28,021 frames with zero field mismatches, and the order book returned 17,205 of 17,205 reference records.

The board run also found two problems: a counter that carried over between replays, and a board-agent binding that failed on PYNQ 3.0.1 (the book test used a validation adapter). Both are fixed in the source; a repeat run on the board is next.

129BOB's verification view: the C++ reference, the recorded RTL parser and the recorded RTL order book, each compared with the same golden model and each reporting zero mismatches.
The verification view holds the C++ engine and both FPGA cores to one reference, and labels recorded simulations as recorded.

What it is not

Paper only. Not a forecast, not calibrated, not profitable trading, not production high-frequency trading, and not connected to any exchange or broker.

Next

Repeat the board run with the fixes in place, then work toward paper trading on Indian market quotes, still without real orders.

All projects