I build systems that stay correct when reality happens: retries, crashes, out-of-order delivery, and network partitions.
I design reliability-critical backend infrastructure and real-time processing pipelines—often at nationwide scale— where correctness must hold under retries, crashes, and network disorder.
My focus is deterministic processing: preventing silent data corruption, eliminating duplicate side-effects, and enabling crash-safe recovery and replay in production.
Re-architected sequential video analysis into a sharded parallel pipeline (receiver → analyzer → uploader) with per-channel ownership and file-level locking.
Designed a stateful streaming detection architecture for nationwide network telemetry (~4 Tbps across a 100-node cluster) with session tracking and a Kafka rule engine.
Designed temporal deduplication for high-frequency RF streams (8K pkt/s across 1000 channels), aggregating packets per frame and selecting the highest-confidence signal.
Replaced a volatile Redis queue with a durable Postgres-backed task ledger, adding transactional task claiming and semaphore-based rate limiting to prevent lost work on crashes.
Fixed real-time counter inconsistencies in high-traffic telecom monitoring by introducing queued aggregation with atomic batched writes.
Resolved a split-brain failure in an active-active control system by introducing external epoch-based leader arbitration to ensure a single authoritative writer under network partitions.
Exploring deterministic data processing where systems remain correct even when events are duplicated, ordering is not guaranteed, and workers restart at any moment.
Feed → Filter → Transform → Execute → Forward → Feedback
Add your GitHub link below after you publish the repo.
For opportunities related to IoT platforms, telemetry pipelines, streaming systems, and reliability-critical backend infrastructure: