Algorithmic Trading Platform

Three Asset Classes, Real Orders, and the Machinery That Makes That Survivable

Section 1 of 5

Three independent trading systems I built and run, one per asset class: NIFTY index options, crypto perpetuals across BTC, ETH, SOL and PAXG, and Indian equity momentum. Each is a standalone Python project with its own package, test suite, systemd units and dedicated server. Behind them sit research labs that never deploy anywhere — their only job is to try to kill a signal offline before it is allowed near the live code.

Since August 2026 these systems place real orders. That single fact changes the engineering completely. A paper system that loses a fill has a reporting bug; a live one has an untracked position on an exchange and real money behind it. Most of what is described in these sections — the reconciliation paths, the layered kill switches, the deliberately distinct error type for “I do not know what happened” — exists because of that difference.

The repositories are private, since they contain live strategy logic. These pages cover the architecture, the safety design and the operational history instead. No position sizes, capital figures or strategy parameters appear anywhere here.

3

asset classes trading independently

~225

merged pull requests

705

commits

1,900+

automated tests

79

numbered research experiments

System Architecture

Nothing reaches a live system without first surviving its research lab. The labs are offline by design and never deploy — the arrow between the rows is the only path a signal has.

Equity Research Labmomentum, walk-forwardoffline onlyCrypto Research Labregime, sizing, patternoffline onlyOptions Researchpremium, theta, open interestoffline onlyEquity MomentumNSE session loopin-memory + Parquet snapshotCrypto PerpetualsBTC / ETH / SOL / PAXG24/7, no market-hours gateNIFTY Optionswebhooks + autonomous pollersKite primary, Groww failoverShared Operational SubstrateSQLite state · systemd services · Telegram bot per project · Excel reporting · backup pipelineOne dedicated server per live systemseparate projects, separate environments, separate blast radiusvalidated signals promoted down — the labs themselves never deploy

How It Is Organised, and Why

The tempting design is a single trading platform with pluggable broker and asset adapters. I went the other way: three standalone projects, each with its own package, dependencies, virtual environment, tests, deploy units and server.

The reason is blast radius. These systems place real orders, and a shared core means a refactor made for options can break crypto at three in the morning on a market that never closes. Keeping them separate means a change is scoped to one asset class by construction, and an outage in one cannot take down the others.

What they share is shape rather than code. Each has a signal layer, a service enforcing risk limits as a single choke point, a repository over SQLite, a Telegram notifier and a backup path. Knowing one makes the next legible, without coupling their release cycles. The one genuine divergence is the equity system, which keeps state in memory with a Parquet snapshot rather than SQLite, because its session model is different enough that forcing it into the same persistence pattern would have been pretending.

There is exactly one real coupling across the boundary, and it is documented rather than accidental: the equity system reads its broker credentials from the options project rather than keeping a second copy, which means a credential rotation must restart both. Writing that down is the difference between a known constraint and a surprise.