Evidence is scattered across providers
Price, fundamentals, news, macro and sentiment live in different systems, so a single reasoning step never sees the whole picture.
FIELD NOTES / FINTECH & CAPITAL MARKETS
A desk does not trade
on one person's view.
A multi-agent trading platform that mirrors an institutional desk's structure — separate research, debate, risk and approval stages feeding a deterministic approval gate, so a trade proposal is never the output of a single model's opinion.
AI SYSTEMS CASE STUDY
TradingAgents
Adversarial research.
Deterministic approval.
TradingAgents / Decision architectureSOLUTION ARCHITECTURE
THE DESK STRUCTUREResearch, challenged before it trades.
Analyst team, research debate, order ticket and risk committee — each a distinct stage rather than one opinion.
Follow the pipeline ↗ResearchDebateTicketReviewRelease
DOMAIN
Investment research & execution
WHAT WE BUILT
Multi-agent decision platform
OUR ROLE
Architecture & AI systems engineering
01 / THE REAL ENGINEERING PROBLEM
One model's opinion
is not a decision process.
Financial decision-making depends on far more than price movement. Before responsibly taking a position, a team may need to evaluate historical price behaviour, technical indicators, fundamentals, news, macroeconomic conditions, sentiment, event probabilities, portfolio risk, position sizing, and previous decisions and their outcomes.
That information normally sits across different platforms and data providers.
More importantly, professional investment teams rarely rely on one person's view. Research is challenged, risk teams scrutinise the proposed exposure, and portfolio managers decide whether the trade should actually proceed. Traditional workflows achieve that through people and meetings.
Unchallenged research is just an opinion
A desk works by argument — a bull case and a bear case, bounded and compared — not by accepting the first thesis produced.
A model should not release an order
Approval has to be a deterministic gate that a proposal passes or fails, rather than a judgement the model makes about itself.
02 / INSIDE THE ARCHITECTURE
Every stage, separate.
Every decision, traceable.
The delivered decision architecture.
From market evidence to an approved, machine-readable order.
An agentic pipeline shaped like a desk.
An analyst team gathers market, news, fundamentals and sentiment; a bounded bull-versus-bear debate follows; an order ticket is typed; a risk committee reviews it; and only an approved rating releases an order.
THE DECISION PIPELINE
Analyst team→Research debate→Order ticket→Risk committee→Order release
03 / GROUNDING AND APPROVAL
Abstain beats estimate.
Every time.
Financial data is verified deterministically rather than accepted from a model. A no-data sentinel makes the system abstain instead of estimating, because an invented figure is worse than a missing one.
Past outcomes are injected into the next approval, so the desk's memory improves the decision rather than sitting in a report nobody reads.
Look under the hood ↘01
Gather and verify
Market data, news, fundamentals and sentiment are collected, then checked against deterministic identity and staleness guards before any agent reasons over them.
Proceeds to reasoning
Single-symbol authority, identity checks and look-ahead guards mean the agents reason over data that is actually current and correct.
Abstains explicitly
The no-data sentinel stops the pipeline rather than allowing an estimate to stand in for a fact.
02
Debate, bounded
Bull and bear positions are argued to a bounded conclusion, producing an investment plan rather than a single unchallenged thesis.
03
Review, then release
A risk committee reviews the typed ticket in conservative-to-aggressive rotation, and only an approved rating releases an order to the configured venue.
04 / THE TECHNOLOGY & THE REASON
Agents where judgement helps.
Code where it must not.
The stack separates the parts that reason from the parts that verify and authorize, so the reproducible, auditable half of the system never depends on a model behaving well.
LG
AGENT ORCHESTRATION
LangGraph + LangChain
Runs the multi-stage workflow: analyst team, bounded research debate, order ticket and risk committee.
Why a graph?
A desk is a sequence of distinct responsibilities; a graph makes each stage explicit rather than one long prompt.
PY
PLATFORM
Python + Pydantic
Types the artifacts that move between stages, including the executable trade ticket itself.
Why typed tickets?
An order that leaves the system has to be machine-readable and validated, not free text a model happened to produce.
LLM
REASONING
OpenAI, Anthropic & Bedrock
Provide the reasoning behind analysis and debate, with the provider a configuration choice rather than an architectural one.
Why multiple providers?
Reasoning quality varies by task and changes over time; the pipeline should not be rebuilt to change model.
MEM
INSTITUTIONAL MEMORY
SQLite decision ledger
Holds the decision ledger, realised return and alpha, benchmark comparison and written reflection.
Why remember outcomes?
Past outcomes are injected into the next approval, so the desk learns from what actually happened.
RD
STATE
Redis
Carries fast-moving state across the pipeline stages during a decision cycle.
Why separate from the ledger?
Working state and the permanent record have different lifetimes and different guarantees.
DK
DEPLOYMENT
Docker
Packages the pipeline for reproducible runs, whether supervised by a desk operator or on scheduled coverage.
Why reproducible?
A decision process that cannot be re-run identically cannot really be audited.
The architecture diagram and narrative reproduce the supplied project documentation. This project's source material contains no product interface screens; the architecture diagram is shown in their place.
Explore the documented platform stack
- Agent orchestration
- LangGraph · LangChain · analyst, debate, ticket and risk stages
- Platform & contracts
- Python · Pydantic · typed executable trade tickets
- Reasoning
- OpenAI · Anthropic · AWS Bedrock
- Memory & state
- SQLite decision ledger · realised return and alpha · Redis working state
- Verification
- Deterministic identity checks · staleness guards · look-ahead guards · no-data sentinel
- Deployment
- Docker
Technology names identify the documented implementation, not partnerships or endorsements.
A CONVERSATION ABOUT YOUR DECISION WORKFLOW
Who challenges the thesis
before it becomes a trade?
Bring one decision your team makes repeatedly with AI assistance. In 20 minutes, we will map which stages need argument, which need verification, and where approval has to be deterministic.
20 minutes · Your process, data and constraints