Problem J.time limit: one lifetime · memory: unbounded · sleep: segfaulted

JayPokale

Researcher at IIT Hyderabad, top 1% competitive programmer, powerlifter, and author of AI tooling people actually star. This page is the editorial. Modesty was cut in code review.

Input
an unreasonable problem.
Output
a shipped, verified system.
Scoring
partial credit not accepted.
Note
it is guaranteed that a solution exists.

The judges demanded proof. He brought receipts.

chapter one ↓

§1 · Observations

First, the evidence.

The section recruiters scroll to first. Hi. Screenshots are allowed.

Ratings prove he solves other people's problems. So he started shipping his own — in public, where strangers can file issues.

the story continues ↓

§2 · Solution, part one

Open source.

Code strangers can read. Some of them even starred it.

W1 · 2026 · open source

Chisle

★ 1,171

Live from GitHub. None purchased, all emotionally earned.

$ npx chisle # one command, thirteen agents

GitHub ↗Website ↗

Makes AI coding agents talk less, build less, read less and say more: answers 45% shorter, oversized tool output cut 48%, and 156 of 156 graded answers still correct. Thirteen agents, zero dependencies.

Like a senior dev who bills by the syllable.

The only tool in its class that publishes the benchmarks it loses. Terrible growth strategy. Weirdly effective.

follow the work

$ gh api -X PUT user/following/JayPokale

# 32 followers · new repos, merged PRs and the benchmarks Chisle loses, as they ship

Fewer followers than Chisle has stars. The code is more popular than me, which is the correct order.

Not everything gets a README and a star count. Some things just quietly run in production and never call home.

the story continues ↓

§3 · Solution, part two

Personal projects.

62 repos were audited. These survived. The snake game knows what it did.

  • GST Legal RAGin productionretrieval over Indian GST lawAbstains instead of inventing citations. Some professionals bill hourly for the opposite.
  • Case-law Knowledge Graphprivatetax judgments, linked by who cites whomLinkedIn for judgments, except the endorsements are binding.
  • Forgeprivatedurable task orchestration on Next.js and Cloudflare WorkersTasks survive crashes. I mostly do too.
  • Office Management4 platforms↗web, backend, Android and iOSOne developer. Four platforms. Zero mercy.
  • AuthorsLogarchivemulti-user blogging platformFrom the era when I said yes to everything.

Then one of his systems retrieved a poisoned answer. He took it personally. IIT Hyderabad is now funding the grudge — times three.

the story continues ↓

§4 · Solution, part three

Now, the lab.

Three theses, one sleep schedule. Anything titled 'X is all you need' is read with suspicion — mine included.

thesis 1 · GraphRAG security

“Corroboration Is All You Can Certify”

Poison a knowledge graph; I prove how much of the answer is still true.

Usually less than the answer thinks.

thesis 2 · semantic-cache integrity

“Cache Me If You Can”

Semantic caches can serve an attacker's answer to your question. Found the law behind it, then the loophole: attack success 63% → 0.

Condolences to the attackers.

thesis 3 · continual learning

“Some Memories Take Time”

Zipf's law already fixes how many learning timescales a model needs: about nine.

Every deployed system has two. Nobody asked Zipf.

Every villain has an origin story. His has a math page and twenty thousand accomplices.

the story continues ↓

§5 · Solution, part four

Where it all began.

The origin story. 20,000 witnesses, none of them paid.

Before the ratings and the retrieval pipelines, there was a math page. Dare2Solve grew into a community of more than 20,000 people solving problems they were told were too hard for them. I founded it, I still run it, and it is still the best proof I have that difficulty is mostly a rumor.

The classroom cuts both ways — five years of collecting tools and the scars that teach you how to hold them.

Off the keyboard, I'm a powerlifter. The only PRs I set without a code review.

toolbox — used in anger, sorted by eraProficiency claims audited by the same guy who wrote them.

languages — fluent in semicolons

  • C++
  • TypeScript
  • JavaScript
  • Python
  • Go
  • LaTeX

algorithms — the DSA years

  • data structures
  • graph theory
  • dynamic programming
  • number theory
  • computational geometry

AI / retrieval — the thesis arc

  • RAG pipelines
  • hybrid retrieval (BM25 ⊕ dense)
  • cross-encoder reranking
  • LLM evals & judges
  • multi-agent systems
  • MCP
  • PyTorch
  • NLP

web — the rent payers

  • React / Next
  • SolidJS
  • Node / NestJS
  • tRPC
  • GraphQL
  • Tailwind
  • three.js
  • GSAP / Motion

infra & security — trust issues, professionally applied

  • MongoDB
  • MySQL
  • Docker
  • GitHub Actions
  • Cloudflare Workers
  • Vercel
  • web security / CTFs
  • network security
  • Arch Linux (btw)

§6 · Appendix

In writing, too.

For when a commit message won't cover it.

all writing →

Which brings us to the verdict. Spoiler: it's green.

the story continues ↓