This case study is password protected

October 2026

TraderUK: Evidence-Gated AI Research

typed research agents, deterministic risk controls and a multi-agent development workflow

Agentic AI

Market Intelligence

Role

AI Engineer

Timeline

October 2026

team

Saman Salari + AI coding agents

platform

Python, FastAPI, PostgreSQL, LangGraph

Diagram of a supervisor agent, a doer agent, a pull request, a PR-checker agent and automated QA gates

i developed TraderUK as an early-stage market research platform, with typed AI agents assessing frozen evidence snapshots. deterministic code owned trade eligibility, risk limits and order handling. the research runtime had no broker access.

my focus was the boundary between a model's judgement and an allowed action... including the cases where the system had to stop or admit it didn't know.

Constraints

i kept live automated trading disabled. model outputs couldn't set prices, position sizes or risk limits, and missing evidence had to stay unknown rather than become a pass.

research jobs were limited to frozen, point-in-time snapshots, without web, shell or database access. i kept synthetic fixtures separate from forward paper evidence, so passing tests couldn't become a profitability or calibration claim.

Architecture

i separated ingestion, research, policy and execution into Python modules, with typed Pydantic contracts between services. persistent job leases and immutable accepted results made the research workflow auditable.

Jev adapters handled typed semantic judgements in the research layer. budget checks, retries and abstention rules surrounded those calls. a deterministic verification gate could downgrade a result, never upgrade it.

AI-assisted development

i used a LangGraph development loop with separate implementation and review roles. DeepSeek handled implementation in isolated worktrees. Claude handled planning and specification review, while Codex independently reviewed the same commit.

Ruff, strict mypy, import-boundary checks and automated tests were code-level gates. protected changes still needed owner approval... a green check didn't let the agents approve their own exceptions.

Stack & operational controls

the backend used Python 3.11+, FastAPI, Pydantic and PostgreSQL, with httpx for service calls. Docker defined the local services, and GitHub Actions ran lint, type checks and tests.

i added explicit readiness states and audit events. the execution gate checked the kill switch and readiness before broker I/O, and unknown position state blocked new orders.

Status & limits

this remained an early-stage implementation. the equity research loop was fixture-only, and broker integrations and parts of the dashboard were still scaffolded or degraded until connected.

live automated trading stayed disabled. profitability, predictive accuracy and calibration remained unproven without measured forward paper evidence. local contract tests weren't proof that upstream services were live or production-ready.

Let's Talk

i build generative AI pipelines and the apps around them. if you have a workflow that works on one machine and needs to work for a whole team... tell me about it. i'm looking for work where generative models have to run in production, not just in a demo. if that's your problem, send me the details.

Comment

Saman

Open to contract work, full-time roles, and interesting conversations about hard Agentic problems.

1