Active across 50+ public developer, AI, open source, and research surfaces
_{Code, datasets, package registries, preprints, reproducible demos, agent workflows, and OSS contribution traces.}

Open Source  ·  Publications  ·  Recently Shipped  ·  Packages  ·  Projects  ·  Impact  ·  Experience  ·  Stats

  8+ years building production systems at Fortune 100 scale
  Former SDE at Amazon Web Services  •  Currently at Southwest Airlines
  Deep expertise in ML systems, distributed architectures, and full-stack engineering

The profile README is partly managed by scheduled automation. Pull requests run profile checks that compile the Python refresh scripts and verify the managed README markers plus the cached stats files before changes merge.

🌐 Live at mukundakatta.github.io/agent-stack - single landing page for the whole 119-package ecosystem (npm + PyPI + MCP Registry + GitHub Marketplace).

🤗 Try it live on the HuggingFace Space · jailbreak fixtures on the HF Dataset.

Five small, focused npm packages that fix the boring problems every long-running agent eventually hits. Pure ESM JavaScript, zero runtime deps, TypeScript types in the box. Designed to compose into a pipeline: fit → guard → snap → vet → cast.

npm i @mukundakatta/agentfit @mukundakatta/agentguard @mukundakatta/agentsnap @mukundakatta/agentvet @mukundakatta/agentcast

Each one also ships as an MCP server so Claude Desktop, Cursor, Cline, Windsurf, and Zed can call them directly mid-conversation:

npx -y @mukundakatta/agentfit-mcp     # fit a chat history into a budget
npx -y @mukundakatta/agentguard-mcp   # check URLs against an egress policy
npx -y @mukundakatta/agentsnap-mcp    # diff tool-call traces
npx -y @mukundakatta/agentvet-mcp     # validate tool args + generate retry hints
npx -y @mukundakatta/agentcast-mcp    # extract / validate JSON from LLM text

_{Sibling libraries that share a design philosophy: small, focused, zero-dep, BYO-LLM. Each one solves a single concrete reliability problem so you can pick the ones you need without dragging in a framework. Previous drop, streamparse (streaming JSON parser, npm + Homebrew + MCP Registry), is still in active use.}

I contribute practical fixes to AI SDKs, MCP tooling, eval frameworks, agent infrastructure, structured outputs, and developer experience.

My lane is finding the sharp edges that slow builders down: unclear contracts, brittle tool calls, docs that almost answer the question, eval gaps where regressions hide, and AI tooling that needs better failure signals. I like small, reviewable patches with clear intent, and compact packages that turn repeated manual checks into reusable workflows.

Recent contribution areas (merged upstream):

• Microsoft - security and architecture docs for internal AI-engineering toolchains (hve-core, physical-ai-toolchain)
• Pydantic - pydantic-ai integration with the Vercel AI SDK
• Hugging Face ecosystem - safetensors Python bindings, sentence-transformers trainer migration docs
• Meilisearch - heed multi-target docs.rs infrastructure
• Vercel - next.js documentation
• Apache Software Foundation - doc / comment fixes across iceberg, pulsar, skywalking, ozone, iotdb

I keep a public log of selected OSS work in oss-contributions.

_{Across ~425 unique external repos (excluding my own). Source: GitHub search API, author:MukundaKatta is:pr -user:MukundaKatta. Last refreshed 2026-05-14.}

Top external repos by merged PRs

Top external repos by total PR volume (merged + open + closed)

Distribution pattern. Each flagship ships as a complete unit, not a single npm package:

library  →  Python port  →  CLI binary  →  GitHub Action  →  Homebrew formula  →  MCP server
   npm           PyPI                           Marketplace        brew tap            npm

So the same problem (mcpcheck, skillint, streamparse) is solvable from any environment a developer or AI assistant happens to be in: a TypeScript app, a Python script, a CI workflow, a terminal, or directly inside Claude / Cursor / Cline / Windsurf / Zed.

• openai/openai-node #1831 — improved fallback handling for non-standard JSON error bodies
• openai/tiktoken #529 — added PyInstaller hooks for dynamic encoding plugins
• googleapis/python-genai #2298 — clarified response_schema vs response_json_schema
• microsoft/playwright-mcp #1562 — clarified extension connection and tab-selection flow
• anthropics/anthropic-sdk-python #1412 — fixed async memory tool example docs
• stanford-crfm/helm #4210 — fixed later-page deep links for run instances

Last refreshed 2026-05-13 from npm, PyPI, and the GitHub API.

Latest releases

• 2026-05-11 · @mukundakatta/lorem-mcp v0.1.0 · npm
• 2026-05-11 · @mukundakatta/color-mcp v0.1.0 · npm
• 2026-05-11 · @mukundakatta/mime-mcp v0.1.0 · npm

Recently merged PRs

• 2026-05-12 · freeCodeCamp/freeCodeCamp #67330 — fix(curriculum): clarify DOM element node wording
• 2026-05-04 · PrefectHQ/fastmcp #4069 — Fix #4056: keep blank query values, add token bucket regression test
• 2026-05-04 · PrefectHQ/fastmcp #4076 — fix(openapi): keep blank values in parse_qs (refs #4056)
• 2026-05-04 · PrefectHQ/fastmcp #4070 — docs(integrations): add Pydantic AI FastMCP toolset guide
• 2026-05-07 · elastic/beats #50281 — docs: fix 'ElasticSearch' casing and 'a SSL' -> 'an SSL' across reference docs

npm (scope @mukundakatta):

Flagship packages:

Package	Why it matters	Install
@mukundakatta/streamparse partial JSON for LLM streams	Streaming JSON parser that yields partial valid trees as tokens arrive. Render LLM tool calls mid-stream, recover dropped responses, parse messy ` ```json ` blocks. Zero deps, 64 tests. Also published as an MCP server in the official MCP Registry.	npm i @mukundakatta/streamparse
@mukundakatta/streamparse-mcp MCP: parse partial JSON	MCP server that lets Claude / Cursor / Cline / Windsurf / Zed parse partial, truncated, or messy JSON on demand. Three tools: parse_partial_json, extract_json_from_text, validate_json.	npx -y @mukundakatta/streamparse-mcp
@mukundakatta/mcpcheck MCP config quality gate	Lint MCP config files for Claude Desktop, Cursor, Cline, Windsurf, and Zed. CLI, GitHub Action, and SARIF for code scanning.	npm i -g @mukundakatta/mcpcheck
@mukundakatta/designlint frontend quality checks	HTML/CSS accessibility and design linter for contrast, touch targets, headings, form labels, and leaked secrets.	npm i -g @mukundakatta/designlint
@mukundakatta/skillint AI skill validation	Lint Claude Code SKILL.md files for frontmatter, required fields, descriptions, and hardcoded secrets.	npm i -g @mukundakatta/skillint
@mukundakatta/ai-eval-forge eval harness	Zero-dependency eval harness for comparing model, prompt, and agent behavior. CLI plus programmatic API; also on PyPI.	npm i @mukundakatta/ai-eval-forge
@mukundakatta/codex-skill-kit Codex skill tooling	Scaffold and validate Codex skills from the command line. Published for npm and PyPI workflows.	npm i -g @mukundakatta/codex-skill-kit
@mukundakatta/kavach AI-app threat signals	Small, inspectable threat-scoring library for AI-app security monitoring: signals to weighted score to tier and playbook.	npm i @mukundakatta/kavach

PyPI:

Package	Purpose	Install
claude-skill-check	Lint Claude Code SKILL.md files for YAML frontmatter, required fields, description quality, and secret patterns.	pip install claude-skill-check
mcp-config-check	Validate MCP configs across Claude Desktop, Cursor, Cline, Windsurf, and Zed; catches auth, transport, duplicate, and placeholder issues.	pip install mcp-config-check
claude-hooks-check	Audit Claude Code hooks for malformed matchers, dangerous commands, invalid events, and hardcoded secrets.	pip install claude-hooks-check
claude-commands-check	Validate Claude Code slash-command files for naming, frontmatter, model values, allowed-tools shape, and secret leakage.	pip install claude-commands-check
llm-usage-report	Parse raw LLM API response logs and generate token and cost reports by provider, model, day, project, or user.	pip install llm-usage-report
codex-skill-kit	Scaffold and validate Codex skills from Python environments; mirrors the npm CLI workflow.	pip install codex-skill-kit
ai-eval-forge	Zero-dependency LLM and agent eval harness with exact, regex, token-F1, JSON, and citation-coverage checks.	pip install ai-eval-forge
agent-run-diff	Compare baseline and current agent runs across success, errors, tools, output drift, steps, latency, and cost.	pip install agent-run-diff

GitHub Marketplace (7 Actions):

Composite GitHub Actions, discoverable on the GitHub Marketplace:

Linters:

• claude-skill-check
• mcp-config-check
• claude-hooks-check
• claude-commands-check

Agent-stack CI gates:

• agentvet-action - fail PRs on bad LLM tool definitions
• agentsnap-action - fail PRs on tool-call trace drift
• mcp-stack-validate-action - one CI gate that runs all 5 agent-stack tools

Homebrew tap - mukundakatta/tools:

brew tap mukundakatta/tools
brew install claude-skill-check mcp-config-check claude-hooks-check claude-commands-check

Each ships a CLI, a programmatic API, and (for the linters) a composite GitHub Action you can drop into any workflow in 3 lines.

🦀 crates.io (Rust) - MukundaKatta - small focused crates for the LLM / RAG / agent niche:

Crate	Purpose	Install
ragdrift RAG drift detection	Five-dimensional drift detection for production RAG: data, embedding, response, confidence, query mix. MMD + sliced Wasserstein + KS + PSI in pure Rust. Sibling Python wheel ships as ragdrift-py.	cargo add ragdrift
ragdrift-core RAG drift core	Pure-Rust core of ragdrift: KS, PSI, MMD (RBF kernel), 1D and sliced Wasserstein, k-means. No BLAS dependency. Used as the backbone of the Python ragdrift-py wheel via PyO3.	cargo add ragdrift-core
embedrank vector top-k retrieval	Batched cosine / dot / L2 distance for f32 embeddings + a heap-based top-k selector. No BLAS, no allocator surprises. Designed for the hot path of small-to-medium RAG retrieval.	cargo add embedrank
promptbudget token-budget truncation	Token-budget-aware text truncation with multiple strategies (head, tail, head+tail, smart-cut with marker). Bring-your-own tokenizer; no hard tiktoken dependency.	cargo add promptbudget
stopstream streaming stop-sequence detector	Streaming-safe stop-sequence detector for LLM token streams. Holds back exactly the suffix-prefix overlap so partial matches at chunk boundaries never leak downstream. UTF-8 boundary safe.	cargo add stopstream
citecite RAG citation markers	Citation-marker [1] [2] injector + parser for RAG outputs. Round-trippable: inject markers tied to source ids, parse them back when post-processing, or strip them entirely.	cargo add citecite
ragmetric RAG retrieval IR metrics	IR metrics for RAG retrieval evaluation: Recall@k, Hit@k, MRR, NDCG@k. Pure data ops, no model dependencies. Sibling-in-spirit to ragdrift.	cargo add ragmetric

The five rust-llm-stack crates are also available together as one workspace: MukundaKatta/rust-llm-stack. Each crate is independently versioned and published.

Total: 85 published crates under MukundaKatta on crates.io.

🤗 HuggingFace - mukunda1729 - 14 Spaces · 13 Datasets:

🚀 Live Gradio playgrounds (6):

📖 Static reference & explainer pages (8):

📊 Datasets (16) - all MIT, all datasets.load_dataset("mukunda1729/<name>") ready:

 ML Systems           Fault prediction, embedding pipelines, model evaluation, cost-optimized inference
 Agentic AI           RAG pipelines, LangGraph workflows, query routing, hallucination detection
 Cloud Infrastructure AWS (Bedrock, SageMaker, ECS, OpenSearch), GCP, Azure, Kubernetes, Terraform
 Full-Stack           React/TypeScript + Java/Python backend APIs, CI/CD, zero-downtime deployments

If you follow my work here, you’ll mostly see:

• open-source contributions to AI SDKs and agent tooling
• MCP, eval, and developer-experience improvements
• practical full-stack and infrastructure-heavy AI projects
• systems thinking around memory, retrieval, orchestration, and production reliability

University of Central Missouri - M.S. in Big Data Analytics and Information Technology (Jan 2021 - May 2022)

SRM University - B.Tech in Mechanical Engineering (2012 - 2016)