Senior Software Engineer

I contribute to production AI systems and enterprise platform capabilities across GPT-5, GPT-4.1, text-embedding-3, and OpenAI Frontier. I lead a seven-person team of full-stack, ML, platform, and QA engineers, working directly with enterprise customers from workflow scoping through pilot acceptance and production delivery.

I built enterprise investigation and workflow systems using React, Next.js, TypeScript, FastAPI, PostgreSQL, Temporal, REST APIs, and webhooks. These systems support multimodal inputs, evidence citations, version-bound approvals, durable execution, session recovery, transaction-backed action journals, outbox delivery, and downstream reconciliation.

I developed model adaptation and evaluation workflows using PyTorch, Transformers, LoRA, SFT, DPO, TRL, MLflow, vLLM, CUDA, and Triton. I also designed GPU-serving observability, telemetry, canary releases, and rollback gates on Kubernetes.

I built and optimized distributed serving infrastructure, including Python/FastAPI gateways, Go/gRPC interfaces, and a Rust/Tokio migration for a high-throughput service. The migration improved p99 latency from 180 ms to 105 ms and reduced memory usage from 4.0 GiB to 2.2 GiB at 5,000 requests per second. I also delivered governed inference telemetry using Kafka, Spark, Databricks/Delta Lake, Snowflake, and enterprise identity controls including SAML, OIDC, SCIM, ACLs, and audit contracts.

I optimized the C++ autonomy runtime for sensor freshness, bounded scheduling, real-time execution budgets, and memory use. This reduced component p99 latency from 32 ms to 23 ms and improved diagnosis of timing regressions through replay workflows.

I built reproducible ML evaluation infrastructure with Python, PyTorch, and TensorFlow across SIL and HIL scenarios. I versioned models, simulations, and inputs to help autonomy and field-validation teams turn road incidents into reviewable release evidence.

I delivered an autonomy evaluation dashboard using React, TypeScript, Node.js, Express, MongoDB, and Python services. The replay and comparison tools reduced median triage time from 35 to 22 minutes across 300 cases.

I also established sensor ingestion pipelines from four depots into Google Cloud Storage, supporting approximately 18 TB per day with resumable transfers, checksums, late-event metadata, and traceable replay references.

I built marketplace booking and availability experiences using React, Redux, Apollo GraphQL, Ruby on Rails, ActiveRecord, Java, and Thrift. I owned selected product changes from UI state and API design through event instrumentation, staged releases, and reporting reconciliation.

I optimized GraphQL resolvers and Java/Thrift fan-out with caching and deadline-aware execution, reducing endpoint p95 latency from 310 ms to 225 ms and reducing downstream RPCs from 12 to 5.

I migrated 18 Ruby on Rails endpoints into four Java/Thrift services on AWS EC2 and RDS. I used compatibility adapters, shadow tests, regression fixtures, and rollback checks, increasing eligible-request success from 99.50% to 99.90%.

I also owned Spark and Airflow data pipelines across 30 datasets processing approximately 8 TB per day on AWS S3 and EMR. I added schema contracts, deduplication, backfills, lineage, freshness checks, and validation gates for risk, experimentation, and financial reporting.

I worked on high-scale backend modernization and distributed ML systems. I tuned asynchronous Linux C++ and Python RPC paths using Folly Futures, tracing, and deadline management, reducing component mean latency from 45 ms to 30 ms and p99 latency from 240 ms to 120 ms.

I built three machine learning workflows and six operators for FBLearner Flow, which were adopted by ranking, ads, and Trust & Safety teams. The workflows standardized feature freshness, reproducible evaluation, and training-to-serving integration.

I developed reusable feature extraction, dataset validation, and offline evaluation components that connected versioned training inputs to serving workflows. I worked with partner teams to identify stale features, reproduce model comparisons, and reduce training-serving inconsistencies.

I benchmarked MyRocks and RocksDB, reducing workload-specific storage usage from 10 TB to 6.5 TB while maintaining QPS within 5% of the baseline. I documented compaction, write amplification, recovery, and operational trade-offs.

Master of Science in Electrical Engineering.

Bachelor of Engineering in Electronic Engineering.