I’ve been fortunate to design, lead, and implement the entire Infrastructure as Code powering the Deductive stack.

• Infrastructure Provisioning – Built a one-click system (kernel IaC stack) to provision AWS, Kubernetes, ArgoCD, secrets, persistent volumes, SOC2-compliant load balancers, auto-scaling groups, ECR, WAF, TLS, Route53, and more.

• CI/CD & Release – Replaced ArgoCD’s default image updater with a custom solution (complete rewrote within 1 week, ~2k loc, without crappy AI), enabling ~1,000 releases in <2 years and fast iteration.

• Config Management System – Designed a Google/Meta-inspired DSL system (Jsonnet + Python) that continues to drive Deductive’s infra and workflows.

• Scaling & Reliability – Implemented elastic auto-scaling for inference on spot instances, built safe container management for AI agents, designed DB backup/restore systems, and scaled ingestion pipelines with batching + vector embeddings.

• Developer Velocity – Optimized Docker builds (parallelization, caching), introduced inference profilers & load generators (Locust), enabled hot-reload for Python inference, and built a local Kubernetes environment (Minikube, NGINX, Ingress) mirroring production.

• Application: created slack app integration and Microsoft team bot integration for alerts (learned Azure within a few days)

## About Deductive AI

Deductive AI is disrupting a $60+ billion market at the intersection of Databases and AI. We are a well-funded startup assembling a founding team of world-class experts, passionate about the 0→1 journey and committed to making history.

If you are exceptional at what you do and want to help shape the future of infrastructure, AI, and data systems—drop me a note. Let’s build the next era together.

• Data Insights

• Creating application with Bard (LLM) + fine tune

• Next generation & Native (C++) execution engine: https://github.com/facebookincubator/velox (Spruce: SHJ, SMJ, Lambda Expression, RDDScanExec, ...)

• Efficiency engineering saved the spark runtime millions dollars per year:

* Lead the Spark on Intel Turbo Boost across 7 teams - 16% faster across 65k fleets, saved $20M per year based on the performance per watt

* Leverage GraalVM to replace OpenJDK yields 12% faster spark runtime and $38M saving per year: https://medium.com/graalvm/graalvm-at-facebook-af09338ac519

* NUMA binding for spark executor - saved millions dollar per year and reduced the remote memory access about 20%

* 50%+ speed up of spark-wrapper: startup time reduced from 10 secs to less than 5 secs (multi-threads, python-debugger/pdb)

* Replaced the spark runtime to use jemalloc instead of glibc malloc memory allocator

• Revamped and modernized the spark infrastructure fundamentally and completely

* Wrote a new SparkSQL parser (3K LOC within one week) that brought up 50% test coverage (from 50% to 100% queries) and closed one of the biggest feature gap in A/B test (ANTLR4)

* Migrated to new build system (83K LOC within one month) - this revived the test framework and paved the road for the spark test/release infra fundamentally - resolved the 7 yrs old tech debt (maven to buck/blaze/bazel, JDK8->JDK11)

* Scaled the test throughput 10x to Metastore API from bash pipe (2K LOC)

* Automated spark namespace move/cutover (auto cell migration & data placement & disaster recovery)

• Data Warehouse / Data Infra

• Distributed Systems (Spark, Chronos/Scheduler)

Performance history: (MA = Meet expectations, EE = Exceeds expectations, GEE = Greatly exceeds expectations, RE = Redefine expectations)

* 2019 H1: MA

* 2019 H2: GEE

* 2020 H1: Suspend due to COVID for all people

* 2020 H2: RE top 1% employees got this rating, not the inflation 10% these days

* 2021 H1: RE top 1% employees got this rating, not the inflation 10% these days

* 2021 H2: GEE

Teaching Assistance in Operating System course (CS450)