Available for internships
based in  Hyderabad · India

Karunya Muddana · CS (AI & ML) @ GITAM

Builds things that ship.

Document review for a healthcare certifier in Germany. A data assistant for a chemical plant in Hyderabad. Fixes merged into IPython and git-cola.

See the work My story
now: AI for a healthcare certifier in Germany
that’s me
Karunya Muddana
Hyderabad · '26
maker of things ↓
Second-year CS (AI & ML) GITAM Builder · OSS contributor

A short version,
in case you don’t
have time for the long one.

↳ ten seconds, promise.

I picked up Python the year my school exams ended — mostly out of curiosity, partly because nobody was telling me what to do next.

A year later, the small projects had turned into things real people use: a CBSE result analyzer running in two schools, and an AI assistant a Hyderabad chemical manufacturer’s staff used every day to ask questions of their Excel data. I later open-sourced its engine as ExcelMCP.

Now I’m building a document-review system for a healthcare certifier in Germany, and sending fixes to libraries I use. Two have been merged so far, into IPython and git-cola.

Read the longer story

By the numbers, so far.

~ small wins ~
9.15/10
CGPA
B.Tech CS (AI & ML) · GITAM
2
Fixes merged upstream
IPython · git-cola
4
Open-source projects
2 merged, 2 in review
33
Test cases
some of them forged on purpose
600+
Students helped
across 2 schools
2
Countries with clients
India · Germany

Where the code runs.

  1. HEAD → now Oct 2026 – present

    AI Engineer contract · remote

    Healthcare certification project · Germany

    • Building an AI system that reviews certification applications by checking the uploaded documents against what applicants typed into their forms.
    • Caught a bug where valid certificates scanned with phone apps were read as expired. Good applications would have been sent back.
    • Changed a rule that rejected a whole application over one missing file. Now that file is requested instead.
    • Added a second, independent reading of every document that flags disagreements for a human, and built a 33-case test set that includes tampered files.
  2. upstream Oct 2026 – present

    Open-source contributor

    IPython · git-cola · icalendar · scikit-learn

    • Got fixes merged into IPython (image display) and git-cola (error reporting) within two weeks.
    • Further fixes in review at icalendar and scikit-learn, two widely used Python libraries. see the PRs ↓
  3. freelance 2024 – present

    Freelance AI Engineer

    Chemical manufacturer · Hyderabad

    • Designed the spreadsheet system the operations team ran their day-to-day work on.
    • Built an AI assistant that let staff ask questions about their Excel data in plain English and see the exact cell behind each answer. Used daily until the company moved to an ERP.
    • Open-sourced the engine behind it as ExcelMCP, an MCP server for live Excel access. Later built the company’s website and its chatbot.
  4. init expected 2029

    B.Tech, Computer Science (AI & ML)

    GITAM University, Hyderabad · CGPA 9.15 / 10

Things that shipped.

View all
Agent workbench · Smart India Hackathon 2026 · Apache 2.0

Native

An agent you run yourself. It reads your documents, executes Python in a network-disabled Docker container, searches a knowledge base that never leaves the machine, and writes the .docx, .xlsx and .pptx files you actually have to hand over — then reopens each one and reports its real contents back from disk.

Every role is a chain, not a choice. When a free tier rate-limits, the run doesn’t stop — it walks down to the next link, and the last link is a model on your own hardware with no quota at all.

Groq→Gemini→OpenRouter→NVIDIA→Ollama · local
22 tools · runs offline · no paid API
Python LangGraph FastAPI Docker sandbox PP-OCRv4 RAG Ollama
Native running a query: the plan, the tool call with its arguments and duration, and the answer

works with the wifi off

Every run shows its working — what it planned, which tool it called with which arguments, how long it took, and what came back.
↗
MCP server · Open source · MIT

ExcelMCP

A live Excel intelligence layer for AI agents. Point it at a OneDrive folder and an agent can ask questions about those spreadsheets in plain English — against the numbers that are in them right now, not a copy taken last Tuesday.

Most spreadsheet integrations embed your cells into a vector store and answer from that snapshot. This one splits the problem: structure is cached, data never is. Every response carries a fetched_at stamp and is_cached: false, so the model can see in-band that the number is fresh.

Question→Embed→Lexical rerank→Live Graph fetch→pandas

Answers come back with provenance — file, sheet, cell. When two sheets disagree it returns conflict rather than picking one, and a query matching nothing returns diagnostics instead of a confident empty set. The setup wizard detects and configures 12 agent hosts, backing up whatever was there first.

11 tools · 12 agent hosts · nothing served from disk
Python FastMCP MS Graph API FastEmbed NumPy cosine Pandas MSAL
→ lookup("raw material stock, plant 2")
{
  "value": 18450,
  "unit": "kg",
  "file": "Stock_Register.xlsx",
  "sheet": "Plant 2",
  "cell": "D27",
  "is_cached": false,
  "fetched_at": "2026-09-10T09:02:41Z"
}
sample response · numbers made up

never answers from a stale copy

↗
Python package · PyPI · Aug 2026

engraph

Long-term memory for AI coding assistants. It gives Claude Code a memory that lasts between sessions, so the next session doesn’t start from zero.

# one line to install
pip install engraph
Python PyPI Claude Code Agent memory
↗
ML pipeline · 2025

House Price Prediction

Benchmarked eleven models on the same split, then tuned the best one until it priced most houses within a tenth of their sale price.

79% of houses within 10%
scikit-learn XGBoost LightGBM Pandas

a weekend favour

A tall stack of messy printouts turning into a neat bar chart and spreadsheet
Data tool

CBSE Gazette Analyzer

Turns the board’s messy text gazette files into clean Excel and PDF analytics for teachers. 600 students, 2 schools.

Python Pandas
↗

argues with itself

Three little blob characters around a table: one researching, one writing, one critiquing
Multi-agent

Researcher · Writer · Critic

A three-agent pipeline that researches, drafts, and critiques its own output until it’s good enough to ship.

Python LLMs
↗

smells like coffee

Roastly's homepage: a latte-art cup under a big serif Roastly wordmark
Website

Roastly

A website for a coffee shop: a latte-art hero, a full menu you can order ahead from, and reviews from the neighbourhood.

Next.js Frontend
↗

Fixes in other people’s code.

Read a big codebase until the bug makes sense, then send the smallest change that fixes it. two merged inside two weeks ↓

ipython/ipython

IPython

A fix to how images are displayed, in the interactive shell underneath Jupyter.

see the merged PR ↗ merged
git-cola/git-cola

git-cola

Better error reporting in the Git GUI, so a failed command says what went wrong.

see the merged PR ↗ merged
collective/icalendar

icalendar

A fix in review at the library Python projects use to read and write calendar files.

see the open PR ↗ in review
scikit-learn/scikit-learn

scikit-learn

A fix in review at Python’s best-known machine-learning library.

see the open PR ↗ in review

What I write.

~ five languages, no padding ~
Python
Python
the first love
TypeScript
TypeScript
for the web bits
SQL
SQL
37 tables deep
C++
C++
DSA territory
Java
Java
OOP, done properly
toolkit.py 
toolkit = {
    "ai":    ["LLM agents", "RAG", "embeddings", "vector search",
              "OCR", "prompt engineering", "LLM evaluation",
              "MCP", "LangGraph"],
    "ml":    ["scikit-learn", "XGBoost", "LightGBM", "PyTorch",
              "Pandas", "NumPy"],
    "tools": ["FastAPI", "Node.js", "PostgreSQL", "Supabase",
              "Docker", "Git", "GitHub Actions", "pytest",
              "Google Cloud"],
}

A few things I’ve written.

Tooling · Mar 2026

“A weekend tool, 600 students.”

The CBSE gazette analyzer started as a one-weekend favour. Then teachers at two schools used it on the year’s board results.

Read on LinkedIn
MCP · Jun 2026

“Structure cached, data live.”

The one architectural rule behind ExcelMCP — and why letting agents read stale spreadsheet data is worse than no access at all.

Read on LinkedIn

Let’s build something.

Open to internships, collaborations, and interesting problems. Quickest replies over email.

linkedin.com/in/karunya-muddana github.com/Karunya-Muddana