io_uring vs mmapNEGATIVE
hypothesis · Kernel-bypass-ish async reads should beat a mapped file for tape ingest.
measured · It didn't. mmap + madvise stayed ahead on this workload. Weeks of understanding the ring, honestly spent.
SIM CLOCK 09:30:00
MSGS 0
TAPE 0 msg/s
systems · low-latency · fullstack — Dehradun, IN
Pulkit Chauhan
Proof over promises.
I build the invisible machinery that has to be fast — and measure every claim. C++ where nanoseconds matter · distributed CI shipped for GSoC · AI agent fleets under verification harnesses · Candidate Master on Codeforces · fullstack when it needs to ship.
0x01algorithmic thinking· hover the headline
Competitive programming, ranked against the world.
why this matters
Codeforces is the global arena for algorithmic problem-solving — timed contests against tens of thousands of programmers. Candidate Master puts Pulkit in roughly the top 2% of all rated competitors worldwide, and every number here is pulled live from the official API. It can't be faked.
while (rating < target) solve(); // no base case yet
snapshot numbers — live feed reconnects on next visit
codeforces · pulk1t
1,980
candidate master
- peak
- 1,985
- rated contests
- 19
rating tape · hover any point
loading tape…
accepted work, not a streak badge
Submission signal across both arenas.
0
questions solved
0
active days
one heatmap, both arenas · hover any day
last 53 weeks
■ leetcode■ codeforces■ bothheat = submissions that day
the training ground — where this was forged
0x02open source · gsoc· hover the headline
Google Summer of Code — paid, reviewed, in production.
why this matters
Google Summer of Code accepts roughly 1,200 contributors a year from 40,000+ applicants — under 3%. It means an established open-source organization trusted Pulkit with production infrastructure, and Google paid for it. Every merge below survived real code review.
namespace gsoc::ccextractor { /* code with reviewers */ }
CCExtractor sample-platform · midterm passed · SonarQube clean on every merge
the season so far
before selection
security audit of the existing platform — found rate-limit bypasses and IP-spoofing vectors, reported with fixes
may
selected · one of the year's ~1,200 contributors worldwide
jun
REST API stack lands: auth, RBAC, rate limiting, OpenAPI docs
jul
midterm passed · artifact pipeline + platform-aware binaries merged
aug →
more merges landing — the constellation below grows on its own
the system I own pieces of — hover a stage
github webhook
→github webhook — a PR or sample upload lands; signature-verified, rate-limited at the edge
flask api
→flask api — auth + RBAC decide who may trigger what; OpenAPI-documented, SonarQube-clean
job queue
→job queue — work is fanned out per-platform; idempotent, so a dead worker costs a retry, not a corrupt run
gcp vm workers
→gcp vm workers — Linux and Windows runners execute the real decoder against the regression library
gcs artifacts
→gcs artifacts — outputs and diffs persist to object storage — every verdict is reproducible from its artifacts
status → pr
status → pr — results post back to the PR as checks; humans see a verdict, not a log dump
Benchmarks prove speed; production proves judgment. For CCExtractor's CI platform I shipped a REST API stack — auth, RBAC, rate limiting, OpenAPI docs, its CI pipeline — through real review. Before writing a line, I audited what was already there and reported what I found, with fixes.
- ▸19 merged PRs: a 9-PR REST API (auth, RBAC, OpenAPI, artifacts) and a 9-PR web console on top of it
- ▸SonarQube clean on every merged change
- ▸Qualification-phase security review: rate-limit bypasses and IP-spoofing vectors found and reported
merged upstream — every node is clickable, every line is a review survived
0x03shipped work· hover the headline
Projects that made it out the door.
why this matters
A portfolio of finished things beats a résumé of familiar words. Each card links to something real — deployed, documented, or runnable — across web, infrastructure, and tooling. Range matters: the same person who tunes nanoseconds also ships full products.
template <typename Anything> auto ship(Anything&&) -> production;
depth is the headline — breadth is the default setting
this site
Next.js 16 · React Three Fiber · edge API routes · scroll-driven canvas — the trajectory line you're following is drawn per-frame from your scrollbar
sample-platform (GSoC)
Flask REST with auth, RBAC, rate limiting, OpenAPI · CI runners on GCP VMs · artifact pipelines through GCS
ops
Kubernetes · Docker · GCP · CMake — one tree builds every PitchFork variant plus the WASM target
shipped under real constraints
0x04force multiplier· hover the headline
AI agents that have to prove their work.
why this matters
Anyone can prompt an AI. The rare skill is building the verification systems that make AI output trustworthy enough to ship. Here, agent fleets optimized a 161 GB production test library — and a harness Pulkit designed rejected every change it couldn't prove byte-identical. Zero regressions shipped.
for (auto& a : agents) trust += harness.verify(a.propose());
CCExtractor's 161 GB regression library, optimized by AI agents — under a harness that refuses anything unproven
Agents generate proposals fast and wrong. The engineering is the verification harness: every proposed optimization of the sample library re-runs the real decoder on two production-faithful builds, and ships only if the caption output is byte-identical on both. No diff, no merge — the agent's confidence is not an input.
−6.50 GB
reclaimed from the first verified migration cohort (7.24 → 0.74 GB)
−96%
single-sample record: strip video from a teletext sample, captions byte-identical
+39%
line coverage added back by coverage-selected replacement samples
1
real upstream bug found by the fleet and confirmed by hand — filed, not vibed
harness session · replayed from real logs
▮
0x05c++ · low latency· hover the headline
PitchFork — a market engine measured in nanoseconds.
why this matters
High-frequency trading firms compete on billionths of a second. PitchFork parses NASDAQ's real data feed at 8.91 million messages per second — an entire trading day replayed in 51.6 seconds, on a laptop. Every number is measured, and the failures are published next to the wins.
[[nodiscard]] auto pitchfork() -> order_book;
NASDAQ ITCH 5.0 · C++ · every number below is measured, -O3 +LTO, full-tape runs
Plain words first: an order bookis the live list of everyone waiting to buy or sell a stock, and exchanges broadcast every change to it as a firehose of tiny messages. PitchFork drinks NASDAQ's entire firehose and keeps that book provably correct — at speeds where the unit of failure is a billionth of a second. Hover any number for the human translation.
0.00M msg/s
parse throughput
in human terms
8.91 million exchange messages understood per second — the engine parses the market faster than the market can talk.
0ns
median dispatch
in human terms
153 nanoseconds to handle a typical message. For scale: light itself only travels about 46 meters in that time.
0.0M msgs
full tape, 51.6s
in human terms
One entire real NASDAQ trading day — 459.8 million events — replayed start to finish in under a minute, on a laptop.
0KB
WASM build
in human terms
The whole engine compiles to 46 kilobytes for the browser — smaller than this page's fonts — and outran the native build on a 22.8M-message slice.
measured wins — each delta from an A/B run, same tape
1.90×one order registry, not two
merged the order→price and order→qty maps into a single registry; one lookup, one cache line
1.40×tick-indexed ladder
price levels addressed by array index, not tree walk — O(1) top-of-book, always
1.20×dense hash map
ankerl::unordered_dense over std::unordered_map — fewer misses where it hurts
1.06×pre-sized order map
capacity reserved for the whole session; zero rehashes during peak volume
what didn't work — reported anyway · most portfolios only show the wins
profile-guided optimizationNULL
hypothesis · PGO on the hot parse loop should buy a free single-digit win.
measured · No measurable delta outside noise. The branch predictor had already learned everything PGO could teach it.
order-flow-imbalance signalTRUE BUT BROKE
hypothesis · OFI predicts short-horizon direction — and it does: 55.28% over 5.9M samples.
measured · Median net after spread: −6,199 bps. Real predictive power, still unprofitable. Both facts reported.
the lab — full write-ups, for readers who want the gory details
source + full benchmark report →incl. live replay, 6,307-symbol backtest, strategy lab — all in-browser
0x06the universe map· hover the headline
How deep does each discipline actually go?
why this matters
A skills list says 'C++'. This map shows what living inside a discipline looks like: the core ideas in bold are whiteboard-defensible, the outer orbit is the long tail of real experience. Click anything for a plain-language explanation and where it was used.
static_assert(depth_of(skill) > resume_bullet);
hover a discipline · click any topic for the plain-language version
size = how hard you can push me on it · click any topic
0x07 · now — the line is still drawing
Readyforasuspiciouslycompetentsummerintern?
Summer 2027. I bring the nanoseconds, the receipts, and an unreasonable amount of enthusiasm — you bring the hard problems. Benchmarks you think I got wrong also get replies, usually with a counter-benchmark.
for the curious — this site has instruments
~ opens a terminal (try sudo hire pulkit)ctrl+K command palettej / k scroll like vimtype latency anywhere↑↑↓↓←→←→BA for the CP archives