01 / 09Industrial ML99.63%Within ±0.1 tolerance
MSc team project, 2026
JUMO Laser Trimming
How do you correct a sensor before you know how far it will drift?
Rohit Suryaa Saravanan. I turn raw data into forecasts, AI agents, APIs, pipelines and decisions people trust.
Software developer & data scientist · Fulda, Germany · Open to opportunities
(01)Who am I?Engineer by training. Researcher by habit.
Data science meets software engineering. I build the model and the system it lives in.
I like the moment a hard problem starts to make sense.
Grew up in Chennai. Built enterprise software for a US cybersecurity startup. Now an MSc in Data Science in Fulda.
Half backend and .NET, half forecasting, ML and AI agents. My favourite problems need both.
- Based
- Fulda, Germany
- From
- Chennai, India
- Languages
- English · Deutsch
- Looking for
- Software, backend, data and ML roles
- 96%Prediction accuracyWindmill pitch model, IEEE 2023
Pitch error fell from 8.6% to 2.3%. The turbine made 21% more power.
- 110,123Orders analysedSupply chain pipeline
Every order of a real online store: 92.1% on time, 881 sellers scored for risk.
- 63KFlights modelledAviation delay model
O’Hare averaged 21.1 minutes of delay. Atlanta, 13.6.
- 02IEEE papersMachine learning and blockchain finance
Windmill pitch prediction (2023) and blockchain finance for farmers (2024).
(02)What I built
Nine projects. Real questions.
One from a real factory line. Eight with all the code on GitHub.
(03)Next stopBoarding soon
Will I make my connection? Anschluss.
- WhyOnly 60.1% of Deutsche Bahn long-distance trains were on time in 2025. ↗
- The dataDB’s open timetable and delay data for every station.
- The modelThe chance of making a transfer, with honest uncertainty.
- The twistAn agent that picks the safest connection, not the fastest.
Starting from Fulda, an ICE hub
- 01Collect
Monthly DB delay files for every station, from 2024 on.
- 02Learn
Arrival delay as a range of outcomes, not one guess.
- 03Connect
P(make it) = chance the delay fits inside the transfer time.
- 04Advise
An agent compares routes and picks the one most likely to work.
(04)How I thinkThree rules
Measure the outcome.
A model only matters if it improves a decision.
Build for production.
Clean APIs and readable logs. The boring parts earn trust.
Explain it simply.
If the team cannot follow it, it is not finished.
(05)Where I have been
Built across the stack.
- 2020 to 2024
BTech, Big Data Computer Science
SRM University · Chennai, India
Distributed data and algorithms. Both IEEE papers came from here.
- 2024 to 2025
Associate Software Developer
EigenSecure · USA, remote
Backend APIs, access control and compliance workflows for a cybersecurity startup.
- 2025 to now
MSc Data Science
Hochschule Fulda · Germany
Statistics, ML and data engineering. Most projects here grew from it.
IEEE publications
Certifications
- ✓IBM Data Science ProfessionalIBM, Coursera
(06)What I use01 / 06
From raw data to results.
Layer 01 · Raw data
It always starts messy.
Turbine logs, 110,123 orders, 63,000 flights, race telemetry. Always messy.
Layer 02 · Clean and shape
Most of the work happens here.
Join, fix, build features. Wrong data in, wrong answers out.
Layer 03 · Model
Then the patterns.
Forecasts and classifiers, always checked against a simple baseline.
Layer 04 · Build
A model nobody can call is just a notebook.
Services, APIs and agents, so people can actually use the model.
Layer 05 · Ship
Deployed, versioned, boring.
Containers, pipelines and databases that keep running.
Layer 06 · Results
The part people see.
+21% turbine power, 18.5% forecast error, an agent that never invents a price.