Systems engineer — 6 years building high-throughput backends, ML pipelines, and real-time multiplayer architectures. Currently designing server-authoritative VR and NLP-driven analytics at Northeastern.
Read More
Engineering principles
Ambiguous ownership causes ambiguous bugs. In the VR sandbox, the server owned world state — clients sent inputs, never mutations. In queue-based pipelines, the queue owns ordering. The authority model is the first architectural decision, not the last.
Dead-letter queues, circuit breakers, retry with backoff, idempotent operations — these aren't edge-case handling. They're the architecture. At Geode, replacing a fragile third-party automation with BullMQ meant failure modes became observable and recoverable by design.
Accuracy on imbalanced clinical data is a vanity metric — a false negative on a Critical class has patient cost. p99 latency matters more than mean throughput when tail latency drives user abandonment. The metric should reflect the real consequence of failure in that domain.
I build systems that handle complexity at the infrastructure layer so the product layer can be simple. BullMQ with dead-letter recovery, Netcode authority models, sub-50ms vector retrieval — invisible when they work, load-bearing when they don't.
Highlight projects
AI-era puzzle competition platform. Server-validated submissions, semantic warmth feedback, abuse protection that held under 10k+ submissions in 10 hours.
Vocabulary builder for ADHD learners. Mnemonics carry the load — AI-generated plus a community submission and voting layer. Per-user progress, saves, and notes.
Talk to your coursework. A RAG layer over Canvas LMS with bidirectional voice — Whisper in, TTS out, GPT OSS in the middle. RAGAS-evaluated retrieval.
Classify free-text EHR notes into Stable, Deteriorating, or Critical. Class imbalance under 8% — F1 and recall on Critical drive the design, not accuracy.
01 — Backend
API architectures serving hundreds of thousands of requests. BullMQ pipelines that compressed 30-minute ETL jobs into 2-minute runs. Production infra on bare-metal VPS when managed services weren't an option.
02 — ML & Real-Time
At Northeastern: NLP classifiers for clinical urgency on imbalanced EHR data. Server-authoritative VR with frame-level terrain sync across Meta Quest headsets. A cross-chain analytics platform indexing 17,000+ dApps through concurrent Redis-backed pipelines.
03 — Thread
BullMQ with dead-letter recovery. Netcode authority preventing state drift. Vector similarity returning semantic results in under 50ms. Invisible when it works, load-bearing when it doesn't.
Let’s build
Open to roles in applied ML, distributed systems, and real-time backend engineering. Internships, full-time, or research collaborations.