Intern at Bosch Global Software Technologies in Coimbatore, until March 2027
I build LLM systems, vision models and accessible AI, from medical diagnostics to safety guardrails for production chatbots.
I'm an applied AI engineer and CS student at CUSAT, interning at Bosch Global Software Technologies in Coimbatore. I've worked on LLM safety for a medical chatbot, a vision system that grades handwritten papers, and low-cost non-invasive diagnostics. Outside the lab I host AI Wednesday and mentor hackathons across Kerala.
Right now
Intern at Bosch Global Software Technologies in Coimbatore, until March 2027
Based in
Kerala, India
Internships and research, newest first. The dark panels are concept illustrations.
Jul 2026 – Mar 2027
Intern, Coimbatore
Aug 2025 – Mar 2026
AI Engineer (AI Intern)
May – Jul 2025
Research Intern
Nov 2024 – Feb 2025
Software Engineering Intern
May – Dec 2024
Software Engineer & Mentor
May – Jun 2024
Summer Research Intern
May – Jun 2023
Software Engineering Intern
Research, production and side projects. The dark panels are concept illustrations unless the caption says otherwise.
LLM safety, in production
Multilingual medical chatbot platform on Gemini / Vertex AI. My part: guardrails, prompt-injection detection and RAG + web-search prompting to keep answers grounded.
Employer product (Laennec AI), so no public repo.
Computer vision, top 5 at NCVPRIPG'24
Evaluates handwritten True/False answer sheets: Table Transformer finds the table, a TinyVGG classifier reads each cell, and a Flask endpoint returns the marks.
Signal processing, Semester 8 project
Classifies five urban sound classes from UrbanSound8K (gun shot, jackhammer, engine idling, drilling, siren) using FFT, band-energy and MFCC features. Feature importance picks out the frequency bands that drive each prediction, then bandpass filtering and an inverse FFT map them back onto the waveform. XGBoost scored best at 83.3%.
Code isn't public yet.
Bars start at 70% to show the differences.
Hardware and ML, IIT Kharagpur
A portable NIR-spectroscopy prototype that estimates hemoglobin, glucose and bilirubin, using XGBoost and Random Forest on a custom spectral dataset.
Research work; not published as a repo.
GraphRAG prototype
Generative vision, semester project
A PyTorch conditional GAN (encoder-decoder generator with skip connections) that shifts image lighting by a continuous condition from −1 to +1. A second repo uses its output to augment ResNet50, EfficientNet and ViT classifiers on the Intel Image dataset.
Generative AI
Text-to-image app with Stable Diffusion, prompt enhancement through GPT-4 or Gemini, and live progress over WebSocket. FastAPI backend, Next.js frontend.
AI agent tooling, early stage
Security skills for AI coding agents: rule packs, a stdlib-only Python scanner and Claude Code hooks that scan what the agent writes. Defense in depth, not a guarantee.
def connect(user): key = "sk-live-xxxxxxxx" q = f"SELECT * WHERE id={user}" return run(q)Accessibility, hackathon 2nd prize
Team-built accessibility platform for specially-abled users: voice navigation, sign-language recognition, a RAG-based YouTube summarizer, personalized MCQs, PPT-to-video and YouTube Braille support.
More repositories
How the Semester 8 project turns a classifier's decision into something you can see. The signal drawn here is synthetic and frequencies are not to scale; the real figures are in the project report.
Step 1 of 5
A short urban-sound clip: engine idling, drilling, jackhammer, siren or gun shot from UrbanSound8K. The waveform shows loudness over time, and little else.
Step 2 of 5
A Fourier transform turns the clip into frequency content, alongside band energies and MFCC statistics. These are the features the classifiers learn from.
Step 3 of 5
XGBoost's feature importance ranks the frequency bands behind its predictions. In the report's chart, the two lowest bands carry almost all of the weight.
Step 4 of 5
A bandpass filter keeps only those bands, and an inverse Fourier transform turns them back into a waveform you can look at.
Step 5 of 5
Laid over the original clip, the stretches where that filtered signal is strongest show which moments the model leaned on.
What I reach for most often.
Hackathons, weekly AI sessions and student workshops around Kerala.
Present
Host, Tinkerspace
Weekly sessions on the latest in AI, 20–30 people each, documented as session journals.
2026
Host, mentor and judge
Part of the organizing team for Kerala's largest women-only hackathon, with 6,000 participants.
2025 – 26
Tech Head
Led the technical team for space-education initiatives.
Feb 2025
Host, mentor and judge
Women-only hackathon with 4,000 participants at SNGCE Ernakulam and Adoor Engineering College. Also taught an HTML/CSS/JS class beforehand.
Jan 2025
Mentor
Mentored teams at a women-only hackathon.
2023 – 25
Community Manager & CUSAT Lead
Ran community engagement, events and tech initiatives for the CUSAT chapter.
Sessions I've taught
Cochin University of Science and Technology. CGPA 9.31.
GATE 2026 (Data Science & AI): Qualified.