Sushanth K S

Concurrent Event Processing Engine

High-throughput, multithreaded backend for real-time data ingestion.

▸Engineering Challenges

• Designed a non-blocking event loop architecture to handle 10K+ events/sec.
• Optimized thread contention using lock-free data structures (disruptor pattern).
• Implemented batched writes to Redis to maximize pipelining efficiency.
• Conducted extensive profiling with JProfiler to identify and eliminate GC pauses.

Polyglot Microservices System

gRPC-based distributed services with schema evolution guarantees.

▸Engineering Challenges

• Migrated monolithic logic into 4 discrete microservices using gRPC for inter-service communication.
• Enforced strict backward compatibility using Protobuf schema versioning.
• Orchestrated deployment on Kubernetes with Helm charts for varying environments.
• Benchmarked serialization performance, achieving ~7x speedup over JSON/REST.

Fraud Detection using Graph Neural Networks

Graph-based ML system for detecting anomalous transaction patterns.

▸Engineering Challenges

• Modeled transaction data as a heterogeneous graph to capture relational dependencies.
• Tackled extreme class imbalance (1:1000) using focal loss and undersampling.
• Built a streaming inference pipeline to score transactions in near real-time.
• Achieved 91% recall at a strict 3% false positive rate constraint.

Ephemeral Vault System

Secure, time-bound secret storage system with automatic expiration.

▸Engineering Challenges

• Architected a 'shared-nothing' storage model to ensure strict isolation of secrets.
• Implemented a cleanup daemon to deterministically purge expired records.
• Designed a threat model focusing on data-at-rest encryption and access audit logging.
• Prioritized security correctness over feature bloat, minimizing the attack surface.

Isolog — Structured Logging System

Centralized log ingestion, indexing, and querying platform.

▸Engineering Challenges

• Built a custom log ingestion pipeline capable of parsing structured JSON logs.
• Implemented correlation ID injection for end-to-end request tracing.
• Designed an efficient indexing strategy for rapid time-range queries.
• Focused on backend service observability and debuggability.

Qumail — Distributed Email Pipeline

Reliable, asynchronous email delivery system with retry logic.

▸Engineering Challenges

• Decoupled email dispatch from user actions using persistent job queues.
• Implemented exponential backoff and jitter for resilient retry logic.
• Designed idempotency keys to prevent duplicate emails during network partitions.
• Integrated rate limiting to respect downstream SMTP provider quotas.

Low-Latency Market Data Pipeline

High-frequency market data processing with minimal allocation.

▸Engineering Challenges

• Processed ~1M events/min using a zero-copy buffer design.
• Optimized hot paths to minimize heap allocations and GC pressure.
• Designed cache-friendly data structures to improve CPU instruction cache hit rates.
• Instrumented precise histograms for p50/p95 latency monitoring.

Deterministic Exchange Simulator

Event-driven matching engine enforcing market microstructure.

▸Engineering Challenges

• Implemented a price-time priority matching algorithm.
• Ensured deterministic execution by sequencing all inputs through a single event log.
• Stress-tested the engine under simulated liquidity collapse scenarios.
• Verified correctness of order book state transitions.

Mini CI/CD Testing Pipeline

Custom pre-merge pipeline to enforce code quality and coverage.

▸Engineering Challenges

• Automated unit and integration test execution on every commit.
• Integrated coverage reports to block merges falling below thresholds.
• Reduced main branch regressions by gating incomplete features.
• Cut down CI flakiness by isolating test environments.

NLP Customer Support Automation

Production-ready NLP pipeline for automated ticket tagging.

▸Engineering Challenges

• Built a scalable NLP pipeline to classify customer support tickets.
• Implemented an active learning loop to continuously improve model performance.
• Focused on model deployment, versioning, and monitoring (MLOps).
• Prioritized practical utility and latency over pure academic accuracy benchmarks.