Aniruddh Chandratre

Data platform & GPU cluster efficiency · ~250K H100-equivalent GPUs (A100 · H100 · H200 · GB300 · MI325)

• Own end to end the internal data processing platform running 10B+ jobs and 3,000+ years of compute per month across 10 large GPU clusters — all data operations (acquisition, curation, mining, auto-labeling, training, evaluation) for the FSD, Optimus, and Digital Optimus tracks.
• Proposed and executed a ground-up re-architecture — lower-level core, distributed worker coordination in place of centralized services, rebuilt memory model. Throughput 940M → 15B+ data items/month (17×); maximum dataset per run 1M → 1B+ clips, three orders of magnitude, making dataset sizes that previously could not complete routine.
• Built the cluster scheduling and live defragmentation layer that compacts allocations back into contiguous topological blocks while clusters stay hot, across a heterogeneous fleet normalized to H100 equivalents. Usable compute capacity rose 24% → 88% and effective utilization now averages 98% fleet-wide.
• Compute commanded by the platform grew ~$5M → ~$200M per month at public cloud list price — hundreds of millions of dollars of AI compute unlocked through software rather than hardware — plus an estimated $15M/year in recovered engineering productivity. Tesla's VP of AI Software credited the redesign with pushing Tesla's AI infrastructure to a new level.
• Built lineage tracking across artifacts, datasets, training jobs, and models, giving multi-stage pipelines end-to-end provenance for the first time.

Edge-to-cloud data infrastructure for the gigafactories

• Designed and shipped an any-to-any edge-to-cloud data broker moving 10B+ manufacturing data points/day — MQTT into InfluxDB, Kafka into ClickHouse — in production across the Austin, Berlin, and Fremont gigafactories.
• Wrote its low-level primitives from scratch when available options could not sustain the required throughput on this storage profile: an mmap-backed durable queue sustaining 5M writes/s on NFS-backed Kubernetes volumes roughly 100× slower than an SSD, and a distributed cache at 10M entries/s that deduplicated across 50 deployments and cut storage ~80%.
• Built and operated 20 high-availability services for factory insights across global sites, owning the full deployment lifecycle.

• DARPA ARCOS (Automated Rapid Certification of Software) with Lockheed Martin. Co-authored PSY-TaLiRo, a falsification toolbox for cyber-physical systems still in use in Lockheed's CertGATE pipeline. Ran the bare-metal Kubernetes cluster behind high-volume formal-verification simulations.

• Built the image-ranking model still at the product's core six years later. AfterShoot now serves ~250K photographers, has processed 9B+ images, and is valued at $55M.