B.Tech Computer Science and Engineering
PES University, Bengaluru
Dec 2020 to Aug 2024
Thesis: Generating Images from Text Descriptions using Diffusion Models and CLIP.
Episode 0: the pilot
Shrinidhi KJ
AI and ML engineer in Glasgow, UK
I build LLM applications, retrieval systems and computer vision models, and I test that they work before I trust them.
Prologue
Models that are right for the right reason
I'm an AI and machine learning engineer based in Glasgow. I have just finished an MSc in Artificial Intelligence at the University of Stirling, after a B.Tech in Computer Science at PES University, Bengaluru.
I work on LLM applications, retrieval-augmented generation, agents and computer vision. What I care about most is evaluation: a model that looks right is not always right for the right reason, so I build the checks alongside the system. I use Claude Code every day, and I review and test what it writes before I rely on it.
- Document chunks indexed in my MSc retrieval study
- 34,502
- MRI scans in the BRISC 2025 dataset I trained on
- 4,793
- Endpoints in the Stripe spec twin-api was run against
- 414
- Answers labelled and checked by two human annotators
- 120
Episode 1
Origin
Bengaluru, Dec 2020 to Aug 2024
B.Tech thesis, 2024
Text-to-image diffusion model
Generates images from text descriptions using CLIP, a VAE, a UNet and a diffusion process, trained on Flickr30k.
- Built
- Built and trained the system, with training run on AWS EC2.
- How I checked it
- Evaluated image quality with FID and Inception Score.
- PyTorch
- CLIP
- Diffusion
- AWS EC2
Episode 2
First missions
Where I've worked: internships in Bengaluru, 2023 to 2024
-
Oct 2023 to Dec 2023, Bengaluru
Data Analyst Intern
Nestlé
- Collected, cleaned and explored data in Python, which helped raise the defect resolution rate from 30% to 40%.
- Automated reporting workflows and dashboards, making reporting about 20% faster.
-
Feb 2024 to Mar 2024, Bengaluru
Data Science and Analytics Intern
Zidio Development
- Analysed large datasets from several sources to find trends and give practical recommendations.
- Built Python data pipelines with Pandas, NumPy, SQL and R that cut computation time by about 30%.
- Turned findings into recommendations with people from other teams.
Episode 3
Journey to Scotland
Education: University of Stirling, Jan 2025 to Sept 2026
MSc Artificial Intelligence
University of Stirling, Scotland
Jan 2025 to Sept 2026
- Distinction80%Machine Learning
- Distinction76%Stochastic Processes and Optimisation
Also: Deep Learning for Vision and NLP, Mathematical and Statistical Foundations. Dissertation: Do RAG Systems Retrieve or Memorise? Graduation ceremony November 2026.
Episode 4
The twist
MSc dissertation, 2026: do RAG systems retrieve or memorise?
My first question set leaked its answers.
So I rebuilt it by hand.
And the conclusion reversed.
MSc dissertation, 2026
Do RAG systems retrieve or memorise?
When a retrieval system answers correctly, is it reading the documents or reciting what the model memorised? A five-condition experiment that changes only the documents the model sees.
- Built
- 501 open-access papers parsed with GROBID into 34,502 chunks in ChromaDB with BGE embeddings. Five document conditions, two Llama models used through the Groq API, 120 labelled answers.
- How I checked it
- My first question set leaked its answers, so I rebuilt it by hand and the conclusion reversed. The LLM judge agreed with two blind human annotators on 94% and 95% of answers, with an NLI model from a different family cross-checking grounding.
- Result
- Correctness rose from 33% with no documents to 92% with retrieval (McNemar p = 0.0001). Giving the model the exact source paper showed no detectable further gain (p = 0.5). Told to use the documents, both models followed a fluent false one in 24 of 24 tests. Limits: 12 questions, one topic, one model family.
- Python
- LangChain
- ChromaDB
- Llama 3.1 / 3.3
- DeBERTa NLI
- GROBID
Giving the model the exact source paper showed no detectable further gain over standard retrieval (p = 0.5).
Episode 5
Building now
Selected work, and how I checked it. Every project lists what I built, how I tested it, and what the numbers actually show.
Open source, 2026
twin-api
Turns any OpenAPI 3.x spec into a running mock REST API, so integrations and AI agents can be tested without touching a live service.
- Built
- Dynamic route registration, recursive $ref resolution, schema-valid synthetic responses, authentication from securitySchemes, and route ordering so parameterised paths don't hide specific ones.
- How I checked it
- Ran it against large real specs: Stripe (414 endpoints, 7.6 MB) and the GitHub REST API.
- Python 3.11
- FastAPI
- Pydantic
- Faker
- Prance
- Docker
Computer vision, 2026
Brain tumour segmentation
Fine-tuned YOLOv8n-seg and YOLO26n-seg to outline brain tumours on MRI scans from BRISC 2025 (4,793 scans, 3 tumour classes).
- Built
- A training pipeline with a reproducible mask-to-polygon conversion, and an interactive demo on Hugging Face Spaces.
- How I checked it
- Compared both models on mask accuracy, parameter count and post-processing time.
- Result
- Mask mAP@50 of 0.905 and mAP@50-95 of 0.670. YOLO26n-seg used 18% fewer parameters and was 46% faster at post-processing.
- PyTorch
- YOLOv8
- Gradio
- Hugging Face Spaces
Applied research, 2026
Evidence report drafting pipeline
An LLM pipeline that drafts sections of an expert evidence report from scientific papers, with a person reviewing the output.
- Built
- An LLM client with response caching, retryable versus fatal errors, and per-call token, latency and truncation logging, covered by a 50-check self-test. Built with Claude Code, reading and testing its output.
- How I checked it
- Froze a 38-finding gold set, with a recorded hash and a re-runnable regression check, before any generation ran.
- Result
- All 42 citations traced back to sources the model was given, but only 4 of 19 reachable findings were covered (11 counting partial matches). Well cited, but incomplete.
- Python
- LLM APIs
- Retrieval
- Claude Code
Private repository
Interlude
Tools I work with
LLMs and RAG
- Claude Code
- Prompt and context engineering
- Structured outputs
- LangChain
- ChromaDB
- BGE embeddings
- Hugging Face Transformers
- Llama 3.1 / 3.3
- Groq
- GROBID
- LLM-as-judge evaluation
- Evaluation set design
ML and deep learning
- PyTorch
- TensorFlow
- scikit-learn
- YOLOv8
- Computer vision
- NLP
- McNemar test
- Cohen's kappa
Programming and data
- Python
- SQL
- R
- JavaScript / TypeScript
- Pandas
- NumPy
Engineering
- FastAPI
- Pydantic
- REST APIs
- OpenAPI
- Docker
- Git
- CI
- pytest
- Linux
- AWS EC2
- GCP
- Hugging Face Spaces
- Gradio
How I work with AI coding tools
Plan
Write down what "done" looks like and how I'll test it before any code.
Build
Work with Claude Code in small, reviewable steps.
Read
Review what it writes until I can explain every part of it.
Verify
Tests, frozen gold sets and regression checks, not trust.
Next episode
Your team
Let's build something that works.
I'm looking for graduate and junior AI / ML engineering roles. The quickest way to reach me is email.
Based in Glasgow and open to relocating within the UK. Eligible to work full-time in the UK, with no sponsorship needed.