Muhammed Rasin AI and data engineer Kozhikode, India
I love problems that don’t make sense yet. Always refactoring what I think I know.
I’m Rasin. I have spent a year and a half in industry building LLM backends and data pipelines, most recently working remotely for DearPriceHunter, a Swedish startup. I also built Misata, an open-source synthetic data tool used by more than 50 people. I want a full-time role in engineering, product or go-to-market, on a hard problem, with a team I can learn from.
Muhammed Rasin
Work
What I’ve built, and how it measured up.
Mostly solo projects, each tested, usually on a public dataset. The label under each name says what it is: a product, a research project, a prototype or a demo.
Misata (opens in a new tab)
Synthetic data · 2025
Open source and product
Generates realistic test data with several linked tables from a short description of a business. If you say revenue should grow from 10,000 to 50,000 over a year, the generated rows add up to exactly that, and every link between tables is valid.
- Key choice
- A language model may suggest how the data should look, but ordinary code checks every target you stated before a dataset is accepted.
- Result
- Meets all 136 checks in its public test suite, where a typical Faker script meets 78. Misata Studio, the hosted version, has more than 50 users and one pilot, and was selected for Claude for Startups.
- Python
- NumPy
- DuckDB
- MCP
- pytest
Ratio (opens in a new tab)
Legal search · 2026
Research project
Answers legal questions about Indian court judgments, with a citation for every claim. It searches 10,588 judgments and tells apart what the court decided from what a lawyer argued.
- Key choice
- Send questions that contain a case citation to an exact lookup instead of a semantic search, which kept confusing similar citation numbers.
- Result
- Finds the exact case asked for 89% of the time on 313 questions I labelled by hand, up from 59% with keyword search and 4% with semantic search alone.
- Python
- BM25
- Embeddings
- Reranking
Entify (opens in a new tab)
Duplicate detection · 2026
Open source
Finds duplicate customer records in messy data, such as the same person entered twice with a typo. It sets itself up from the data, explains why it matched two records, and needs no specialist tuning.
- Key choice
- Let the tool choose its own matching rules, then test it on a standard dataset it had never seen.
- Result
- On the FEBRL benchmark, every pair it flagged was a real duplicate, and it found 98.7% of them.
- Splink
- DuckDB
- TypeScript
Nimbal (opens in a new tab)
Energy forecasting · 2026
Research project
Forecasts how much power wind and solar farms will produce the next day, as a range of likely outcomes rather than a single number, and turns the forecast into bids for the UK electricity market.
- Key choice
- Bid the expected output rather than a cautious low estimate, after working out that this earns the most on average.
- Result
- Replayed against the 2024 HEFTCom forecasting competition, it would have placed 2nd of 66 teams on accuracy and 1st on trading revenue.
- Python
- LightGBM
- Quantile regression
Harara (opens in a new tab)
Heat safety · 2026
Research project
Hour-by-hour work and rest plans for outdoor workers in the Gulf, based on heat forecasts and Qatar’s heat-stress rules.
- Key choice
- Kept the raw weather forecast after testing showed that a learned correction made it worse.
- Result
- Compared on the same hours, the ECMWF forecast missed 12 to 18% of the hours when work should have stopped; NOAA GFS missed 40 to 45%.
- Python
- Weather data
- LLM evaluation
Mitate (opens in a new tab)
Recommendations · 2026
Research project
Product recommendations for small online shops with a few hundred products and a few dozen regular customers. It works from the catalogue on day one and improves as orders come in.
- Key choice
- Fixed the comparison against fifteen tuned methods before running it, so the result could not be cherry-picked.
- Result
- Beat the best single method at every shop size on the public OTTO shopping dataset.
- Python
- LambdaMART
- FastAPI
Argus (opens in a new tab)
Maintenance assistant · 2026
Prototype
An assistant for factory maintenance teams that shows how close a machine is to failing. Every number it gives is calculated by code, not written by the language model.
- Key choice
- The language model only explains numbers; it is never allowed to produce one.
- Result
- Passes all of its test cases on the UCI AI4I 2020 machine-failure dataset.
- Python
- DuckDB
- FastAPI
Footnote (opens in a new tab)
Source checking · 2026
Prototype
Checks AI search answers against the web pages they cite. Every claim must point to a source, and every source is actually read.
- Key choice
- The checks use fixed rules rather than another AI model, so the same answer always gets the same verdict.
- Result
- Runs can be replayed from saved results without new searches.
- Python
- SerpApi
- Groq
Yaqzan (opens in a new tab)
Disaster response · 2026
Demo, built in two days
A disaster-response simulation based on Kerala’s 2018 floods. An AI model sends boats and trucks where they are needed, and a checker blocks any order that names a resource that does not exist.
- Key choice
- The model may only use resources that are actually in the simulation.
- Result
- Built for the MBZUAI and IFM K2 showcase.
- TypeScript
- Next.js
- K2 Think
The Quantum Canvas (opens in a new tab)
Generative art · 2025
Experiment
Type a name and it becomes a small planetary system, drawn from the results of a five-qubit quantum circuit.
- Key choice
- The orbits come from measured quantum results, not from a random number generator.
- Result
- Runs in the browser with Qiskit and p5.js.
- Qiskit
- p5.js
About fifty more repositories on GitHub, unfinished ones included.github.com/rasinmuhammed
Experience
Where I’ve worked
A year and a half in industry: data engineering at JMAN Group in Chennai, then remote backend and LLM work for DearPriceHunter, a Swedish startup, from December 2025 to August 2026. Alongside both, freelance work and my own open-source projects.
Dec 2025 – Aug 2026 · Sweden (remote)
Software developer, contract, DearPriceHunter AB
A Swedish startup that analyses what companies spend with their suppliers. I worked remotely on the backend and its LLM features.
- Built a document ingestion pipeline as FastAPI services that turns uploaded CSV, Excel, PDF and JSON files into structured spending data.
- Built a detection engine that flags unusual spending, with LLM-written reports in English and Swedish. Prompt caching cut inference cost by about 90%, and Langfuse tracing showed what the models were doing.
- Ran models on AWS Bedrock in an EU region so client data stayed within data-residency requirements.
Mar 2025 – Present · Remote
Backend and AI developer, Independent
Freelance work for Onepane AI, and my own open-source projects, mainly Misata.
- Onepane AI: built a report generator with LangGraph, FastAPI and the OpenAI API that summarises cloud billing data.
- Peer reviewer for Springer Nature’s Network Modeling Analysis in Health Informatics and Bioinformatics.
Jul 2024 – Feb 2025 · Chennai, India
Software engineer, data engineering, JMAN Group
My first job after graduating, in the data engineering team of a data and analytics consultancy. The synthetic data work there is where Misata started.
- Built a probabilistic record-linkage model with Splink and PySpark on Databricks for an entity-resolution product, tuning blocking rules and match thresholds for precision and recall.
- Generated synthetic datasets so consultants could test ideas without touching production data.
- Built an LLM lead classifier and Power BI dashboards; trained in Snowflake, dbt and Azure Data Factory.
About
How I work
Before I build something, I decide how I will test it. Then I publish the results, including when a simpler approach turns out to be just as good.
I use AI coding tools a lot, and I am open about it. On these projects I chose the problem, how to test it, what to compare it against, and what to ship or cut. Each repository records those choices, and I am happy to walk through any of them.
Beyond the code, I care about why someone would want a product, how they would find it, and whether it actually helps them. Running Misata Studio, I do some of that work already, so I am as open to product and go-to-market roles as to engineering ones, especially where the two meet. I am early in my career and still learning, and I am looking for a team where I can keep doing that for years.
- Name
- Muhammed Rasin O M
- Based in
- Kozhikode, Kerala, India (IST)
- Working since
- July 2024
- Roles I fit
- AI Engineer, Machine Learning Engineer, Data Engineer
- Education
- B.Tech, Computer Science and Engineering (data science), Mar Athanasius College of Engineering, 2024
- Also open to
- Forward Deployed Engineer, Solutions Engineer, GTM Engineer, Technical Product Manager. Remote, on site or relocating
- Contact
- rasinbinabdulla@gmail.com
Now · November 2026, Kozhikode
- Running
- Misata Studio, the hosted version of Misata: more than 50 users and one pilot
- Maintaining
- The open-source Misata library, about 2,000 downloads a month
- Reading
- How data products find their first hundred users
- Looking for
- A full-time role in AI and data, in engineering, product or go-to-market, on a problem I can work on for years
Tools I have shipped with
- AI and LLMs
- Large language models (LLMs), RAG, LangGraph, LangChain, OpenAI, Claude and AWS Bedrock APIs, Ollama, Langfuse, Prompt caching, MCP
- Data
- Python, SQL, pandas, NumPy, PySpark, ETL pipelines, Databricks, Snowflake, dbt, Azure Data Factory, DuckDB, PostgreSQL, Splink, Power BI
- Machine learning
- scikit-learn, LightGBM, PyTorch, NLP, Model evaluation, Statistics and A/B testing
- Engineering
- FastAPI, REST APIs, SQLAlchemy, pytest, Git, Docker, GitHub Actions (CI/CD), TypeScript
From the founder I worked with
“Muhammed is exceptional. He consistently delivers far beyond expectations, not just in speed, but in quality of thinking, ownership, and execution.
He doesn’t just do what you ask; he actively improves the solution, flags risks early, and proposes better approaches when needed. He has an amazing way to sort out the most messy and fluffy visions into real life details and functions. If I could give 7 stars, I would.”
Josephine GustingerFounder & CEO, DearPriceHunter AB
Open source
Open-source packages
Python packages, most of them on PyPI, each written because I needed it first. 116 GitHub stars across them, most of them on Misata.
- misata75
The Python library behind Misata. Describe the totals a dataset must hit and get rows that add up to them exactly, with a report you can rerun to check.
- indicfuzznew
Matches Indian names that sound alike but are written differently, across scripts, word order, initials and titles, for KYC checks and record matching. Scores 0.987 ROC AUC on its test set, against 0.699 for the popular rapidfuzz library.
- rag-tui21
A terminal tool for checking how documents are split up for retrieval (RAG), with live previews, batch tests and export to LangChain or LlamaIndex.
- opsdeck20
An admin panel for FastAPI apps that use async SQLAlchemy, with permissions, per-row access rules and audit logs.
Also: pytest-misata (Misata test data in pytest), mangograd (an autograd engine in NumPy), node-canon (entity resolution for knowledge graphs).
Writing
Essays and notes
Essays on synthetic data, LLM systems and building products, written as I learn. New ones arrive by email.
Contact
Get in touch
If you are hiring, I would like to hear from you. If you can introduce me to a team or a person I should meet, I would be grateful for that too. I read every message, and I will tell you honestly if I am not the right fit.
I’d especially like to hear from
- Teams hiring an AI, machine learning or data engineer for a full-time role.
- Teams hiring for product or go-to-market roles that need someone technical: forward deployed, solutions or GTM engineering, or technical product management.
- Early-stage founders who want an engineer who also cares about how the product reaches its users.
- Researchers working on synthetic data or evaluation who would like to discuss Misata or work on something together.
- Fintech and compliance teams working on Indian name matching for KYC and AML.
Write to me directly
Open to full-time roles, remote or on site. It is --:-- in Kozhikode right now.