02 / AI-ASSISTED ENGINEERING / RESEARCH / 2026
PROJECT I built an AI-assisted research workspace that brings market-data collection, trend-strategy backtesting, paper trading and monitoring together in a web dashboard.
SKILLS Python · FastAPI · React · PostgreSQL · AI-assisted development
HIGHLIGHT One workspace for strategy research, simulation and monitoring.
SYSTEM ARCHITECTURE
The implemented collection, research, simulation and exchange integration paths.
The scheduler collects and stores market data, then requests cache refresh and subsequent processing.
The backend connects research and paper trading to stored data; live exchange operations are conditional on configuration and safeguards.
AI-assisted development Claude Code and Codex support implementation, experiments and review.
APPLICATION SCREENS
Select an image to enlarge it.
OVERVIEW
A workspace for researching trend strategies using market data. It brings together strategy backtests, paper trading and operational monitoring in a web application developed with AI assistance.
MY ROLE
I built a FastAPI backend and React dashboard to collect market data, run strategy experiments, and inspect research results. I connected historical simulation, paper-trading workflows and monitoring views so that research and system operation can be reviewed in one place. I used AI coding tools to support implementation and iterative review while preserving the distinction between research results, simulated trades and actual trading.
DEVELOPMENT JOURNEY
I began with a moving-average crossover analyzer: collect market candles, measure price cycles and compare strategy conditions. With AI coding agents assisting implementation and review, I built a FastAPI backend, database and React interface, then added date-range controls, saved experiments and paper-trading workflows. The goal was to make each hypothesis repeatable rather than leave it as an interesting chart.
As research expanded, collecting and maintaining comparable history became a challenge. Listing dates, the period available from each source and market sessions can differ, so identical-looking tests need not use identical inputs. I used AI to help inspect the data and comparison code, then checked coverage and execution timing against the underlying records instead of treating every discrepancy as a strategy improvement.
To work around collection coverage and compare across markets, I extended the research to TradingView. I used Aside to operate the browser and AI assistance to express ideas as Pine strategies, run tests and document results. Local data and Python research continued alongside it; TradingView provided another price source and execution environment for checking whether an idea survived the same conditions.
I used AI to help formulate hypotheses, implement Pine and Python changes, organize runs and review mismatches. I fixed the market, period, capital and cost assumptions for a comparison and changed one condition at a time; I also checked signal timing, fees, slippage, drawdown and later periods. The AI accelerated iteration, while I set the evaluation rules and examined the evidence.
The workspace grew to include stored experiments, simulated portfolios, paper trading and operational monitoring. Research records also preserve unsuccessful targets and unresolved differences. A historical simulation is useful for the next question, but it is not evidence of future performance or actual fills; that distinction shaped both the dashboard and the development process.