Building internal developer tooling on the Performance Infrastructure team at Meta — turning fragmented, manual workflows into self-service platform abstractions.

• Consolidated multiple fragmented benchmarking tools into a unified, self-service platform that abstracts away the underlying tooling complexity. Developers now specify the metrics they care about (latency, throughput, resource utilization), and the platform handles tool selection, orchestration, and execution — eliminating the need to babysit long-running benchmarks or hunt through messy logs for results

• Built persistent execution tracking across all underlying benchmarking tools — storing commands, configs, run metadata, and resulting metrics — making benchmarks fully reproducible and enabling teams to compare runs over time without manual log archaeology

• Designed the platform so the underlying benchmarking tools are transparent to users: a single interface in, clean metrics out, with full traceability of every run

Technologies: Python, distributed systems, internal developer platforms, developer tooling, developer productivity

• Reduced model serving memory footprint by 600MB per instance across

1,000+ production models by replacing TorchServe with a decoupled serving

architecture that separated the serving application from the serving

framework, improving modularity and eliminating unnecessary memory

overhead.

• Designed and launched a distributed embeddings service for a RAG-based

retrieval system by building a high-throughput, batched and asynchronous

inference pipeline that transforms chunked documents into embeddings,

enabling scalable semantic retrieval in production.

• Reduced infrastructure provisioning time from 10 days to 8 minutes by

re-architecting a shared, PR-driven Terraform pipeline into a self-service CLI

backed by a serverless control plane, enabling safe, on-demand infrastructure

provisioning with guardrails and eliminating cross-team deployment

bottlenecks.

• Improved ML development velocity and platform scalability by transitioning

infrastructure ownership from research teams to a centralized platform and

introducing standardized APIs and observability practices.

• Designed and implemented isolated, GDPR-compliant experimentation environments for ML engineers and data scientists, enabling safe and rapid ML experimentation on production-grade data without impacting production systems or violating data privacy requirements.

Technologies: Python, AWS, Terraform, Docker, ECS, Kubernetes, CI/CD, PyTorch, RAG, embeddings, inference serving, Infrastructure as Code

Built and owned the containerized compute platform at Plusgrade (travel industry fintech) that service teams used to deploy and run their applications. Designed the platform end-to-end — from the ECS-based container orchestration layer to the out-of-the-box observability stack — so that product teams could ship independently without managing infrastructure.

• Increased system availability from 99.5% to 99.99% by architecting the containerized compute platform on ECS with gated CI/CD pipelines, replacing monolithic EC2 deployments with a self-service, component-level deployment model

• Built the platform's observability layer from scratch: centralized logging with the EFK stack (Elasticsearch, Fluentd, Kibana), monitoring and alerting with Prometheus, Grafana, and AlertManager — provided as turnkey capabilities for all service teams

• Pioneered large-scale infrastructure modernization, migrating services from EC2 Classic to VPC and containerizing workloads on ECS

• Decomposed monolithic systems into microservices and introduced component-level CI/CD pipelines, enabling independent deployments and faster release cycles

Technologies: AWS (ECS, EC2, VPC), Docker, CI/CD, microservices, EFK Stack, Prometheus, Grafana, AlertManager, observability, distributed systems