Software Developer · AI Engineer

Kshitij Parashar

What does it take to move AI from proof of concept to production?

Scroll to read

Abstract

What interests me right now is the stretch between AI research and deployment, where a model that works as a proof of concept has to become a system people can rely on. I have seen it from both sides: research at Panasonic, and a healthcare platform built from the ground up for Vyom Health. This page follows the path: the data, the model, the system.

1/Introduction

Hello, I’m Kshitij.

I train AI models and build the software around them. After a year of research at Panasonic, I am now the founding engineer at Vyom Health, building its healthcare platform alongside a Master’s at TU Darmstadt. How AI gets from research to production is what I am exploring now.

Experience

2.5 Years

Completed

19+ Projects

Worked in

6 Companies

2/Background

Doha

High school in Doha, Qatar.

2/Background

Singapore

2/Background

Darmstadt

Now a Master’s in Computer Science at TU Darmstadt. Alongside it I am the founding engineer at Vyom Health, a small team making preventive healthcare more accessible in India, building its platform from the ground up.

1. 2026–

   Apr 2026 – Present: M.Sc Computer Science, TU DarmstadtTechnische Universität Darmstadt

2. 2026–

   Jan 2026 – Present: Founding Engineer, Vyom Health

3/Method

Data

The work starts with the data. At Panasonic I designed how it was collected and studied how it varied. It taught me to ask what the data is good for before asking which model to train: where it can create real value, and not only whether a model can be made to work on it.

3/Method

Model

Then a model that fits the problem. At Panasonic I fine-tuned a U-Net for odour detection and brought the training data it needed down by half. That taught me to judge a model by what it costs to train and to run, and not by its accuracy alone.

3/Method

System

Then the software around it. At Vyom Health the architecture was mine to choose, and it is mine to live with. That taught me to weigh a decision by what it will cost to change later. It also taught me to make software easy for the people who use it: healthcare workers at Vyom who are not technical, and at Panasonic the factory workers who monitored a production line with the tool I built.

4/Experiments

Speech in noise

My bachelor thesis: LSTM models that find speech in noisy audio. I spent months on it, reading what others had tried and then training and tuning my own. It taught me what research is like from day to day: mostly patience, and many runs that change one thing at a time.

Sound as a landscape. Time runs left to right and each ridge is energy at one pitch: speech has peaks, noise is flat. The amber strip is where the detector hears speech, the pale one where speech is. Add noise and they drift apart. Schematic: the audio is synthesised and a simple rule stands in for the trained LSTM.

4/Experiments

Colour from grey

A small project for my course on deep learning and neural networks, on a topic we chose ourselves: networks that turn a grayscale photograph into a colour one. We tried many architectures and loss functions. Comparing them showed me why each design choice is there, how much the loss function decides, and that a picture can score well and still look wrong.

Left of the seam is the picture the network was given, right of it what it gave back. Both pictures are from the project.

4/Experiments

The quickest route

A robot car that plans the quickest route to face each obstacle on an open floor, photographs the picture on it, and reads from the symbol which way to go next. I led the team of eight that built it. It taught me to split the work so that eight people can build at once, and to bring the parts together early and often.

The car can photograph only from its front. Each symbol says how it goes on: left, right or turn around. Schematic: the layout and the planner are stand-ins.

4/Experiments

AI and data

• Explanation Attacks on Face VerificationPyTorch, ElasticFace, Explainable AI, Adversarial Attacks
• Horse Health Predictionscikit-learn, Kaggle
• Robust Voice Activity DetectionPyTorch, LSTM, ONNX
• Research AutomationPython, Matplotlib, Seaborn, Genomic Analysis
• Deep Image ColourizationPyTorch, CNN, Encoder-Decoder
• IntellicheckOpenCV, NLTK, SpaCy
• Olist E-Commerce AnalysisJupyter, NumPy, pandas, seaborn

4/Experiments

Software

• This PortfolioNext.js, React Three Fiber, Tailwind, AI SDK
• Do.MoveReact Native, Expo, Supabase, PostGIS
• ResumeLens AIReact, Puter.js, Zustand, Tailwind
• PreSchool GoWhereReact, Node, Express, Firebase
• Self-Navigating RobotAndroid Studio, Java
• myBikeSGReact, Node, Express, Firebase
• Restaurant Reservation SystemJava, OOP, Eclipse
• Railway Reservation SystemC++, OOP, File Handling

4/Experiments

Security

• Capture The Flag ChallengeKali Linux, CyberChef, GDB, Ghidra
• Pegasus & Heartbleed Case StudySecurity Research
• FEAL Cipher AnalysisDifferential Cryptanalysis, Boomerang Attack
• Cyber Supply Chain SecuritySecurity Research, Supply Chain Risk, NIST

5/Results

Growth

• ResearcherFrom a bachelor thesis to a year of research and development at Panasonic
• EngineerFrom internships inside other people’s applications to founding engineer of Vyom Health’s platform
• LeadFrom leading course project teams to leading the engineering at Vyom Health
• MentorFrom a Java workshop for 30+ colleagues at STMicroelectronics to mentoring three engineering interns

5/Results

Community

• 300+ attendees at the first National AI Student Developer Conference, as program director
• 100+ attendees at Deep Race Dash with AI Singapore and AWS, as program director
• 100+ participants in Codewars, a coding event run with IEEE-NTU
• Teaching children reading, maths and tennis with Youth Corps Singapore, and students Android development
• Organising interview preparation, past-paper solutions and peer tutoring at the NTU computer science club
• Building websites for the Institution of Engineers’ NTU chapter and the Developer Student Club Darmstadt

6/Peer review

Saravanan Sampath

6/Peer review

Siddharth Dangi

6/Peer review

Chandra Gunawan

7/Conclusion

That was the long answer.

The short answer is that I am still learning. This page is what I have learned so far about taking AI from proof of concept to production, in research at Panasonic and in building Vyom Health’s platform.

Further work

I want to keep learning and growing, with people who know more than I do. I’m open to graduate roles and collaborations, and the fastest way to reach me is email.