I build Python services that run unattended: exchange pipelines that survive disconnects, async APIs that respect other people's limits, and LLM systems measured rather than assumed.
Most of the work is commercial and private — trading infrastructure, data platforms and internal AI services. Clients come back for a second contract. That is the metric I care about.
Market data. WebSocket collectors write into PostgreSQL / TimescaleDB. Order books are rebuilt from snapshot and deltas, and gaps are repaired on reconnect.
Exchanges. One execution interface behind 11 venues. Each venue gets its own client and error taxonomy, under shared retry and stream watchdogs.
Backend. FastAPI and async SQLAlchemy, SSE with a polling fallback, background workers, idempotent pipelines.
AI. RAG over internal knowledge bases, function calling with structured output, model routing with fallbacks, and cost accounting per user.
Billing that counted silence.
A platform charged per conversation, and its four-second window started the moment a message
was fetched — not when it was answered. So the obvious poll → llm → reply loop kept
donating paid conversations to the platform's own auto-responder.
I split polling from processing, and turned the uncovered time into a metric.
An API with no counter and no cursor.
Nothing to page by, and an aggressive 429 on top. I wrote a token-bucket client that backs
off honestly on 403, and reports live progress over SSE — falling back to polling wherever
SSE is blocked.
 |
 |
 |
 |
 |
 |
 |
 |
Open to backend and AI engineering work — contract or full-time.