Backend / distributed systems engineer at BNY Mellon, working on high-throughput event pipelines, financial data systems, and reliable workflow-driven services.

I build around three themes:

• LLM reliability and agent guardrails - evals, calibration, logprobs, structured outputs, MCP-style workflows
• Distributed systems and backend infrastructure - Kafka, gRPC, Redis, queues, caches, concurrency, observability
• Developer tools - small, practical tools that help engineers validate, debug, or ship faster

Currently exploring: LLM evaluation, hallucination risk scoring, agent preflight checks, MCP integrations, and production-grade distributed systems.

Open-source Python toolkit that scores LLM output risk using calibrated token entropy. It implements the Calibrated Entropy Score (CES) method from arXiv:2605.28264 and is designed for developers building RAG, agent, and batch QA systems.

• CLI + Python API for scoring entropy sequences and provider outputs
• Calibration workflow, thresholding, evaluation harness, and provider logprob smoke tests
• Agent/RAG guardrail examples with clear limitations: this is a risk signal, not a truth oracle

Static + dynamic analysis tool for finding goroutine leaks, deadlocks, and lock bugs in Go programs using execution traces.

• Detects goroutine leaks and deadlock patterns without application instrumentation
• Goroutine provenance tree, Tarjan SCC deadlock detection, go/ssa static analysis, CI baseline flags
• Built for debugging real concurrency failures, not just toy examples

Distributed caching system with etcd service discovery, gRPC communication, consistent hashing, singleflight deduplication, and Prometheus monitoring.

• Thread-safe LRU cache with concurrent access
• Consistent hashing and hot data replication across nodes
• Benchmarked locally at 23K+ ops/sec with sub-1ms P50 latency

Task queue built with Redis and Go, designed around reliability and at-least-once delivery semantics.

• Concurrent worker pool with lease-based ownership and automatic recovery
• Priority queues, retries with exponential backoff, graceful shutdown
• Web dashboard for queue and worker monitoring

Voice-controlled AI diagramming app that turns spoken architecture descriptions into system diagrams.

• Django + React + Groq + Deepgram
• Structured output validation with Pydantic / Instructor
• Built around real-time AI workflow UX, not just chat

Open-source tool for improving Facebook Marketplace listings from raw item descriptions.

• Generates clearer titles, descriptions, and pricing suggestions
• Practical AI utility for sellers and resellers
• Small product-shaped project focused on immediate usefulness

Languages: Go · Java · Python · C++ · TypeScript

Systems & infra: Kafka · gRPC · Redis · etcd · Temporal · Postgres · Elasticsearch · Docker

AI & LLM: OpenAI · Claude · Groq · Deepgram · Instructor · Pydantic · MCP · evals · logprobs

Patterns: Consistent hashing · Singleflight · At-least-once delivery · Lease-based ownership · Exponential backoff · Goroutines · Worker pools · Structured outputs

Building LLM reliability tools, AI agents, or real-time backend systems? Let's talk.