Skip to content
View parthdagia05's full-sized avatar
😶
😶

Organizations

@sugarlabs

Block or report parthdagia05

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
parthdagia05/README.md

Parth Dagia

I build the layer underneath an agent: the part that has to be fast, cheap and correct when nobody is watching.

CNCF · Hyperledger · Sugar Labs  |  Go · C++17 · Java · Python · TypeScript  |  BITS Pilani '27 · Bengaluru



$ parth --bench

I don't have a portfolio. I have a changelog of things that got measurably better.

BENCHMARK                              BEFORE              AFTER                  DELTA
────────────────────────────────────────────────────────────────────────────────────────
besu · uint256 mul (EVM hot path)      ████████████████    ████                   3.9x faster
besu · div / addMod / mulMod           ████████████████    ████████               28-78% faster
musicblocks · total blocking time      ████████████████    █████                  66% lower
musicblocks · bundle size              ████████████████    ███████████            3.8 MB lighter
musicblocks · startup delay            █████               ·                      5.0s -> 0s
mandate · agent attack success         ████████████████    ·                      80% -> 0%
deception-gym · detection accuracy     █████               ████████████████       32.7% -> 96.7%
deception-gym · sycophancy             ████████████████    █                      17x lower
────────────────────────────────────────────────────────────────────────────────────────
                                       all numbers reproducible from the linked repos

$ parth --upstream

Production codebases, no onboarding, every change fenced by a test that fails on master first.

Project What it is Lang Merged
Music Blocks + v4 Sugar Labs · nonprofit education software · maintainer, GSoC '26 TypeScript 86
Kubescape CNCF · Kubernetes security & admission control Go 39
WasmEdge CNCF · WebAssembly runtime, loader/validator/executor C++17 9
Besu Hyperledger · Ethereum execution client & EVM Java 6
Kubescape · 39 PRs in three weeks, mostly by deleting things  Go
  • Found a scan that passed by scanning nothing. Resource plurals were guessed by appending s, so a policy scoped to sandboxes never matched a Sandbox and the run reported a clean pass over zero resources. (#3355, #3177)
  • Per-resource scoping and multi-action support for generated ValidatingAdmissionPolicy bindings, which previously matched every resource the bound policy did and could not target a CRD at all. (#3283, #3295)
  • Speed came from caching the right things and deleting the wrong ones: cached per-file work in the SARIF writer, indexed a scope's resources once instead of once per rule, replaced a full directory walk with an ancestor lookup.
  • Hardened live-cluster discovery: kind filtering before collection, dedup when several API versions serve one resource, correct handling of an undecided access review. (#3411, #3297, #3446)
Hyperledger Besu · made the EVM stop allocating  Java · JMH
  • 256-bit opcodes on a flat long[] operand stack, zero heap allocation. ADD with carry propagation across four 64-bit limbs, SUB with borrow, plus MUL, DIV and bitwise ops, ~3.9x faster than the boxed path. This is the arithmetic under every EVM opcode.
  • Power-of-two fast paths for UInt256.div, addMod, mulMod: Knuth Algorithm D replaced with a shift or a mask, 28-78% faster, verified under JMH with property-based differential tests against BigInteger.
  • Enabled NullAway null-safety analysis across datatypes and ethereum:rlp.
WasmEdge · post-quantum crypto and the Component Model  C++17 · CMake
  • ML-KEM post-quantum key encapsulation merged into the wasi_crypto plugin with FIPS 203 known-answer tests. Plus ECDSA and EdDSA public-key verification.
  • Component Model value-section loader, value-linearity enforcement, and core global/table type checks at instantiation, rejecting malformed components before they execute.
  • Enabled three official spec suites in CI: naming, invalid, memory64.
Music Blocks · maintainer, and the reason it loads now  TypeScript · React
  • Total Blocking Time down ~66% and a five-second startup delay deleted, by lazy-loading 17 modules (3.8 MB off the bundle), held in place with Lighthouse CI budgets.
  • Designed the brick-rendering engine: stroke-width-aware SVG outline paths with rounded corners and connector notches, so bricks snap together at any scale. Path geometry is unit-tested.
  • Rebuilt the Brick Palette as a declarative system where one schema object determines what renders and how it groups.
  • Review and merge community pull requests as a maintainer.

$ parth --own-work

A deterministic enforcement kernel for agent actions Python SQLite

Keeps the model off the safety path, so the check costs zero tokens and returns the same answer every time. Nine deterministic checks, fail-closed on any store failure, and a CI test that fails the build if the kernel ever imports the model client.

Two-phase idempotency with a recovery scan, so a crash mid-capture and a duplicate webhook each leave exactly one debit.

Content-addressed cache over every model call, so a whole eval run replays byte-identically with no API key.

Attack success 80% → 0% across seven attack classes, scored by programmatic oracles, no judge model. The 12% false-block rate is published beside it, not buried.

An OpenEnv-spec RL environment for deception detection Python GRPO

Trained a model to stop believing confident liars, on a single A10G.

Detection 32.7% → 96.7%, sycophancy 17x lower, held on an unseen harder distribution.

A JIT compiler for integer expressions C++17 LLVM ORC JIT

lexer → recursive-descent parser → AST → IR gen → O2 pass pipeline → ORC JIT. Variables resolve to array slots up front, giving one i64 @expr(ptr) signature.

A PostgreSQL schema migration runner Go PostgreSQL Docker

Per-migration transactional apply with rollback, advisory locking so concurrent deploys cannot race, dry-run and plan previews, one static binary per platform.


$ parth --stack

performance    hot-path profiling, allocation-free data paths, caching strategy,
               cold-start and bundle reduction, JMH / Lighthouse CI
ai systems     Anthropic API, structured output, response caching, deterministic
               replay, RL fine-tuning (GRPO), programmatic eval oracles
cloud native   Kubernetes API discovery, CRDs, RBAC, ValidatingAdmissionPolicy,
               CEL, OPA, client-go
compilers      LLVM IR generation, pass pipelines, ORC JIT, AOT/JIT, WebAssembly

note: the contribution graph below is generated by GitHub and cannot be disabled. It counts days I pushed to a non-fork repository, a strictly smaller set than days I did work, which is a strictly smaller set than days I suffered. The table above is the honest one.

Pinned Loading

  1. BigO BigO Public

    Coding practice platform built with Java 21, Spring Boot, PostgreSQL, Redis & RabbitMQ with a Docker-sandboxed code judge

    Java

  2. WasmEdge/WasmEdge WasmEdge/WasmEdge Public

    WasmEdge is a lightweight, high-performance, and extensible WebAssembly runtime for cloud native, edge, and decentralized applications. It powers serverless apps, embedded functions, microservices,…

    C++ 10.8k 1.2k

  3. sugarlabs/musicblocks sugarlabs/musicblocks Public

    Music Blocks -- A musical microworld

    JavaScript 889 1.7k

  4. sugarlabs/musicblocks-v4 sugarlabs/musicblocks-v4 Public

    A complete overhaul of Music Blocks

    TypeScript 113 240