Savitur Swarup

nanoserve 2026 · source

An LLM inference server from scratch in Python/PyTorch. The interesting part is the scheduler, not kernels. I built the ladder from a naive loop to continuous batching to a paged KV cache with a Triton attention kernel and CUDA graphs, on one GPU:

                  tok/s
naive                24  ▌
static batching     126  ██▌
continuous          243  █████
paged KV + triton   252  █████▎
+ CUDA graphs       578  ████████████

p99 TTFT: 63 s (naive) → 0.57 s (final)

That's 23.7× naive and 86% of vLLM's throughput on the same hardware, at lower p99 time-to-first-token. I also ported the scheduler to C++ and proved the port equivalent by hashed replay of every scheduling decision, and sharded a 7B model that doesn't fit on one GPU across two at 7.8× naive.

Then the harness became an audit: every engine is checked token-exact against the baseline, and no speedup counts until it clears a measured noise floor. That rule killed five of my own first-pass results before it touched anyone else's. Chaos-testing cancellation under load turned up the same bug class in vLLM's V1 scheduler (#49684).

Vig 2025– · live terminal · data desk · pipeline

My market-research platform: a data pipeline that regenerates a seven-screen research terminal every day (model desk, screener, graded past calls, a live paper desk on prediction markets) with no server anywhere. Everything runs on GitHub Actions cron, and the repo is the database.

The data layer is the part I'd show an interviewer. A 20-hour rate-limited backfill is distributed across a 12-worker CI matrix, with the partitioning invariants pinned by tests and an idempotent merge: 74k records in 70 minutes, 12/12 workers green on the first run. When Python became the bottleneck I wrote a C++17 parallel stream-reconstruction engine: 2.0M messages/sec, 8× the Python reference, output parity-pinned against it on all 246 files. Cross-feed reconciliation caught my fill model overcounting by 1,735×; it was counting cancels as trades.

The research side exists to attack its own backtests: walk-forward validation, vol targeting, and a point-in-time S&P universe I rebuilt specifically to measure survivorship bias instead of guessing at it. Most of my first edge turned out to be a backtest artifact. The smaller one that survived is what runs on the terminal.

EEG microstate pipeline 2026 · source

From my research internship at Forschungszentrum Jülich: one preprocessing recipe and one global microstate model across 8 public resting-state EEG datasets: 504 people, 8 clinical conditions. Alzheimer's separates from controls at 0.89 AUC, and the Alzheimer's-trained model transfers at 0.92 to frontotemporal dementia it never saw.

The result I trust most is a negative one: a depression finding that looked real until the robustness battery (age regression, 5,000-shuffle permutation tests, cross-site validation, ComBat) exposed it as a spectral confound between recording sites.

Time Bank 2025 · source

A kids' app with its own economy: chores and routines earn minutes, parents approve requests and stock a reward store. React Native and Firebase, live on the App Store and Google Play.