I build the layer underneath the product.I build the layerunderneaththe product.

Architected an event-driven microservice pipeline that decouples edge ingestion from neural-network compute through Kafka, so backpressure lands in a durable log instead of in the capture loop.

Implemented smart frame sampling — every 5th frame — cutting CPU load by 80% while keeping detection continuity for the tracked subject.

Deployed headless YOLOv8 inference nodes as Kafka consumers, letting throughput scale by adding consumers to the group rather than rewriting the pipeline.

Containerised the whole stack with Docker Compose and dependency healthchecks across Zookeeper, Kafka and PostgreSQL so the system comes up in the right order, every time.

Persisted detection events to PostgreSQL as the durable sink for downstream querying and replay.

Built a high-performance embedded key-value storage engine from scratch in C++17, implementing the Log-Structured Merge-Tree architecture that underpins LevelDB and RocksDB.

Engineered a thread-safe MemTable using std::shared_mutex read-write locks, so concurrent readers proceed in parallel while writers take exclusive access.

Implemented Write-Ahead Logging so that a crash between an acknowledged write and the SSTable flush is recoverable — zero data loss on crash.

Wrote a background compaction engine performing leveled SSTable merging, keeping read amplification bounded as the dataset grows.

Designed the write path so acknowledgement happens after the WAL append, not after the flush — durability without paying disk latency on every put.

Co-developed a visual simulation tool modelling memory-level hardware faults — cosmic-ray induced bit-flips and stuck-at gate faults — to bridge hardware–software interaction and fault-tolerance concepts.

Modelled the memory array with NumPy so fault injection and corruption propagation are computed over the whole address space at once.

Built an interactive Streamlit interface for real-time fault injection and live visualisation of memory corruption patterns.

Supported multiple fault classes — transient single-bit upsets versus permanent stuck-at-0 / stuck-at-1 gate faults — so their different failure signatures can be compared side by side.