Skip to content

RESUME.md

Latest commit

 

History

History
41 lines (31 loc) · 2.22 KB

RESUME.md

File metadata and controls

41 lines (31 loc) · 2.22 KB
Raw

Resume bullets — realtime-md-eval

Copy-ready bullets for a systems / MLSys / infra SWE resume. All metrics are infrastructure facts (latency, cost, throughput, fault-tolerance) — no profit or return claims.

One-liner (project header):

Real-time market-data ingestion + offline strategy-evaluation system — async WebSocket ingestion, a ~$0/month self-healing deployment, and a paper-evaluation harness that reaches a validate-before-you-build verdict. Python. [github.com/…/realtime-md-eval]

Bullets (pick 4–6):

  • Built a real-time order-book ingestion service (async WebSocket, in-memory snapshot

    • delta reconstruction, watchdog reconnect) delivering ~90 ms update latency — ~100× fresher than the 10 s REST-polling baseline.

  • Ran it 7×24 at ~$0/month on a GCP free-tier e2-micro: IPv6-only networking to avoid the external-IPv4 charge, IAP-tunneled SSH (no public port), and systemd Restart=always self-healing with daily sessions that prune resolved markets and resume automatically across crashes and reboots.

  • Cut each polling round to a single batched POST /books request, keeping egress under the 1 GB/month free-tier cap while tracking multiple live markets.

  • Designed an offline strategy-evaluation harness that paper-simulates fills (modeling adverse selection), persists multi-day state, and attributes PnL by price-volatility regime, emitting a deterministic go/no-go verdict via a single pure aggregation function.

  • Used the evaluation to make a validate-before-deploy call: measured that the candidate strategy is net-negative under news-driven volatility (adverse selection) and net-positive in calm, low-competition conditions — capturing a live geopolitical news spike as the decisive case study — before committing to a low-latency executor.

  • Refactored a research prototype into a tested, importable Python package (I/O split from pure logic, single-source fee model + categorization, pure report aggregation) with a pytest suite covering the book state machine, quoting/inventory math, fee curve, depth-walk, and verdict logic.

Tech: Python, asyncio, websockets, systemd, GCP (free-tier e2-micro, IAP), pytest.