Marcus AcostaSoftware Engineer
About
I'm a software engineer who has worked across the stack, from product UI and APIs to infrastructure and data systems. I specialize in AI engineering: shipping agents, probability models, and retrieval pipelines that turn live data into reliable product signals.
Outside of engineering, I recently picked up golf, and you'll usually find me in the gym, watching sports, or hunting for vintage pieces and following fashion. I'm also always looking for the next place to travel.
Experience
- BETTORCACo-Founder & EngineerBettorca.com
Bettorca is a bankroll manager for sports bettors, built on a data aggregation layer that unifies books into one place. Multi-book sync lets users track every slip live across their books, with high-signal analytics on their performance, recommended props from our in-house machine learning models, and a monitor that suggests betting more conservatively or aggressively based on how they're doing.
As co-founder and engineer, I own the product end to end across the stack, from architecture and deployments to new features, product improvements, and system plus AI observability. I also collaborate with outside engineers on debugging and technical decisions, and work directly with the CEO on tradeoffs and product choices that circle back to the business.
Projects
Helped LiteLLM's proxy get smarter about which MCP tool servers it trusts. Added trust scoring so third-party servers get filtered and ranked before model calls go out, with caching per URL, fail-open defaults, and a solid test suite covering both HTTP and stdio servers.
Fixed a quiet billing bug where Flutterwave payments would verify successfully but never update the matching invoice. Added a fallback so the webhook handler can still resolve the right payment reference when the webhook and verified refs don't line up.
Built a semantic caching layer that cut LLM response times from a few seconds down to under 100ms on hits, and dropped token spend to zero when the cache already had the answer. OpenAI-compatible proxy with prompt normalization, deterministic keys, and simple hit/miss cost tracking.