Christopher Stassinos
Labs / Private Project Brief
Private Repository

Public MarketLab

A split trading research lab built around two machines with separate roles: a Mac mini running Hermes Agent for data collection, orchestration, research workflows, and paper-trading oversight, and a 5-card GPU rig for model training, ranking runs, and other compute-heavy analysis. The stack uses the Alpaca API for algorithmic paper trading while storing market data in SQLite so research, screening, and testing stay repeatable across both systems.

  • The Mac mini acts as the always-on research node, collecting stock, options, and crypto data, managing SQLite state, driving Hermes Agent workflows, and supervising Alpaca paper-trading activity.
  • Database snapshots can be copied to the GPU machine so heavier training and ranking work runs on the hardware built for it, while Alpaca positions and strategy state remain visible in the lab tooling.
  • The setup separates orchestration from compute so data collection stays stable while experiments and ranking runs scale on the GPU rig.
Back to portfolio Download CV
Architecture

How the two-machine workflow is structured

The architecture is split by job. The Mac mini continuously collects public market data, maintains the local SQLite dataset, runs Hermes Agent research and reporting workflows, and coordinates Alpaca API paper-trading activity. When there is enough history for training or broader ranking work, a consistent snapshot of the database is moved to the 5-card GPU rig so heavier experiments can run without interrupting collection. One machine stays stable and always on; the other is reserved for larger compute jobs.

01 / Research Node

Mac mini handles collection and orchestration

Public stocks, delayed options chains, and crypto ticker data are gathered continuously while Hermes Agent coordinates reporting, research flow, repeatable operator tasks, and the paper-trading side of the Alpaca workflow.

02 / Handoff

SQLite snapshot transfer preserves consistency

Instead of sharing a live mutable database between machines, the project uses safe local snapshots so the GPU side can train against a coherent dataset copy.

03 / Compute Node

5-GPU rig handles the expensive passes

Higher-throughput model training, ranking experiments, and GPU-heavy analysis jobs are pushed onto the rig designed for them rather than the orchestration node.

Functionality

What the system actually does

Collection + Storage

Local dataset over time

  • Collects market snapshots into SQLite as a durable local research dataset.
  • Keeps the time-ordered history needed for screening, simulation, and training readiness.
Screening + Risk

Deterministic filters plus Alpaca paper execution

  • Screens candidates through transparent momentum, liquidity, spread, delta, and premium constraints before promoting ideas into the paper-trading workflow.
  • Uses the Alpaca API for paper positions while tracking invested capital, current value, and unrealized P&L through the lab ledger and dashboard.
Telemetry + Ops

Operational visibility built in

  • Flask dashboard surfaces candidates, P&L, database health, training readiness, recent runs, and GPU telemetry.
  • Local reporting makes it easy to review positions, system health, and experiment readiness in one place.