← All work

AI research system

TradeLab

An eight-role research workflow that coordinates evidence gathering, critique, synthesis, and review without handing an agent trading authority.

Role
System design · agent orchestration · evidence boundaries
Timeframe
Iterative research platform
Status
Ongoing
Publication
Sanitized public case study
PythonFastAPIReactPostgreSQLTemporal

Eight bounded roles

Separated planning, evidence collection, challenge, synthesis, and presentation responsibilities.

Evidence-first output

Designed a ledger that keeps research artifacts attributable and reviewable.

Fail-closed boundary

Kept the public and research experience read-only with no execution capability.

The challenge

Make complex work decision-ready.

Multi-agent research can produce volume without trust. The design problem was to make delegation legible, preserve provenance, and prevent unsupported conclusions from silently becoming recommendations.

The workflow assigns distinct responsibilities to eight roles, records evidence as structured artifacts, and makes review states visible. The replay is deterministic and represents a prerecorded synthetic session.

Prerecorded synthetic session

Inspect the research system

Fictional subject. No account or brokerage connection.

Research graph

Eight bounded roles

Read-only
Lead
analyst
Planner

Breaks the question into bounded research assignments.

Produces: Research plan

Evidence ledger

Reviewable provenance

Synthetic evidence ledger
SourceTypeOwnerStatus
EV-101Synthetic quarterly filingDocumentData CollectorVerified
EV-102Fictional industry outlookReportDomain ExpertVerified
EV-103Generated competitor tableDatasetFact CheckerVerified
EV-104Counter-thesis memoAnalysisSkepticUnder review
EV-105Assumption registerLedgerArchivistVerified

Workflow timeline

Planning

Planner

Question decomposed and source boundary approved.

Behind the build

System decisions that shaped the outcome

Sanitized architecture and workflow artifacts from the TradeLab development workspace.

01

Orchestration with accountability

Each role receives a bounded assignment and produces a typed contribution. The lead analyst integrates only evidence that has passed the defined review path.

02

A ledger, not a black box

Sources, owners, artifact types, and verification states are surfaced together so a reviewer can trace how the narrative was assembled.

03

Safety boundary

TradeLab is presented here as a research system. The demo contains no account data, live market feed, broker action, or execution path.

Continue exploring

Browse every case study