Soumil
Baldota

Every packet, accounted for.

Per-cgroup network accounting for Cloudflare's edge — an eBPF datapath in the kernel that attributes every byte and packet to the service that caused it.

• Hot-path eBPF program with lock-free per-CPU counters
• Merged to production; rolling toward all Cloudflare traffic
• Split privileged/unprivileged design, shared wire-contract crate
• Migrated build to hermetic Bazel + remote cache, CI 8m → <1m

Forking a running container.

Software Systems Lab research on making container forking fast — cloning a running container's state cheaply inside a custom Linux kernel and Kata Containers.

• Multi-threaded fork syscall replicating memory + FDs via CoW
• Fork latency measured with custom eBPF hooks
• Extending forking to ARM CCA confidential-compute Realms

Keeping the plant running.

Backend and infrastructure for an industrial-plant SaaS — a BullMQ job-processing library running 1M+ jobs a week at 99.9% success, plus multi-tenancy and the latency work underneath it.

• Postgres row-level multi-tenancy — zero cross-tenant data leaks
• Query + index tuning: p99 API latency 450ms → 150ms
• GitLab CI/CD + ArgoCD; test coverage 40% → 90%

Intelligence, on the device.

Multi-modal document classification for low-powered mobile devices — real-time inference on-device, no server round-trip.

• LSTM / RNN / BERT / RoBERTa classifier in PyTorch
• Trained on scraped + BERT-augmented text across domains

Simulating collisions, in code.

A Google Summer of Code project with CERN's HEP Software Foundation, bringing Julia into high-energy-physics data pipelines under maintainer mentorship.

• podio → Julia code-generation for HEP data pipelines
• Jinja2 / YAML template-driven bindings
• Benchmarked Julia vs C++ for performance parity

Teaching a rover to see.

Leading SLAM for a competition Mars rover — the hard-real-world engineering that first pulled me toward systems work.

• Autonomous navigation + localization on rough terrain
• Real-time perception under sensor noise + tight compute
• Built for reliability where flaky code isn't an option