#BuildInPublic
Started building a full algorithmic trading system for Indian markets. Doing it entirely with Claude Code and sharing the whole thing as it happens.
The system - one Python project, nine layers. Each layer is isolated and communicates only through a PostgreSQL database. Broker, data, features, strategy, risk, orders, backtest, and observability.
Some early architectural decisions that are currently shaping everything else -
Broker agnosticity - system layers never interact with broker APIs directly. A broker interface abstracts out the required functionality and ensures compatibility with one or multiple brokers. Selected Zerodha as the primary broker integration for now. Might consider Angel One as a secondary going forward. Swapping brokers does not affect a single line of code.
No machine learning in this system yet. Signal generation is mostly rules-based. Clean separation - get the system right first. Add machine learning later.
Development workflow - broken down into 9 phases, where each phase has a technical spec, MD - architecture, schema, contracts, and logic - and a Claude Code prompt that reads it and builds the corresponding layer. No manual coding for now.
Claude specifications - using a pro plan with Sonnet 4.6. I usually avoid using the Opus models, mostly to keep token usage under control. In my experience, Sonnet has been more than enough for most planning and coding tasks. That said, this might change as the project evolves.
The execution workflow at the moment is - use Claude to generate the phase specific MD and prompt files -> spend some time to manually review the files -> use Claude Code to implement the layer using those files -> final review and push to git.
I’d also want to track Claude usage and token consumption on a daily basis, just to get a realistic sense of what it actually takes to build and scale a system like this.
Let's see how this goes!
#SystematicTrading