AI Engineer · MLOps Engineer
Building intelligent systems that understand language, see the world, and reason like humans.
01 / About
"I love building systems end to end, from the model to the backend to the cloud it runs on, and watching all the pieces come together into something people rely on."
I'm an Founding AI & Software engineer based in Edinburgh, where I build production systems at ZOT Engineering: a full-stack platform and the AI that runs inside it. I hold a Master's in Artificial Intelligence from the University of Edinburgh, where I focused on machine learning, NLP, and large language models, and where my dissertation with Amazon explored privacy and memorisation in LLMs.
Most of my work lives at the point where a model stops being an experiment and starts carrying real load. I take ideas from research through to production, building LLM, computer-vision, and forecasting systems along with the full stack around them, from data and backend to cloud and MLOps, so they hold up once people depend on them.
What draws me is the meeting point of applied AI and thoughtful engineering: reliable, robust systems that solve real problems and make everyday work and life run a little smoother. If you're working on similar ideas, or think we might build something well together, I'd be glad to connect.
Core Specializations
By the numbers
MSc Distinction
University of Edinburgh
Production Systems
Built & shipped 0→1
BE CGPA / 10
Anna University
Education
MSc Artificial Intelligence
University of Edinburgh
Sep 2023 – Aug 2024
Distinction · 82%BE Electronics & Communication
Anna University
Aug 2019 – Jun 2023
CGPA 9.50 / 1002 / Experience
Software Application Engineer
Founding Engineer · Intern → Full-time · PCB & Precision Manufacturing · United Kingdom
"Founding engineer at a UK manufacturer: built the digital foundation 0→1, then turned it into a platform that ships AI and full-stack features across the business at speed."
Scaling the platform: breadth at velocity
Scaled the platform company-wide. Grew the foundation into a unified ERP spanning every division, deploying 15+ production systems on Microsoft Azure and giving the business a single, extensible source of truth.
Generative AI & LLMOps. Delivered an LLM-powered payroll assistant with prompt caching and multi-tool orchestration, plus a privacy-first, multi-agent recruitment pipeline with automated candidate scoring.
Applied ML & computer vision. Engineered a PCB-defect detection system (RANSAC, DETR, DINOv3, SAM2), LayoutLM-based document intelligence, and Times-FM demand forecasting.
Self-serve data & analytics. Built natural-language-to-SQL graph inference and 100+ live executive dashboards, surfacing organisation-wide insight on demand.
Paperless manufacturing initiative. Identified the opportunity and delivered end-to-end a production system that digitises and logs every manufacturing step, reducing paper usage by 80% and advancing the company’s transition to environmentally friendly, paperless operations.
Backend & cloud infrastructure. Architected the backend serverless Azure Functions (114+ endpoints), integrating three legacy on-premise ERP systems over an encrypted VPN, owning DevOps and MLOps throughout.
Cloud architecture & FinOps. Designed the complete Azure cloud architecture from the ground up, creating and managing resource groups across environments and owning FinOps cost governance to keep spend predictable as the platform scaled.
CI/CD & developer experience. Set up the company’s GitHub organisation, repositories, and GitHub Actions workflows, and standardised local quality gates, ESLint, unit tests, and Cypress E2E, behind automated build, test, and deploy pipelines.
Observability & reliability. Instrumented distributed tracing, monitoring, and alerting across production systems, running ongoing performance evaluations, bug fixes, and improvements to keep the platform fast and dependable.
Query performance optimisation. Profiled the report-generation workflow to isolate its bottlenecks and re-engineered the underlying query structures, cutting execution time from 55 seconds to 3.4 seconds, a 16× improvement.
Stack
Laying the 0→1 foundation
Established the software function 0→1. Joined as the first software engineer and set up the company’s entire technical foundation, translating operational needs into a delivery roadmap.
Discovery & requirements. Partnered with stakeholders across three business divisions to map manual workflows and convert user requirements into well-defined deliverables, tracked end-to-end in ClickUp.
Stack & architecture. Selected the full technology stack and designed the relational database schema underpinning the platform.
Foundational product. Delivered the public company website and the first ERP modules and analytics dashboards on an Angular 19 SSR frontend and Python REST API.
Reusable delivery pipeline. Standardised a secure path from database to API to UI, with MSAL/JWT authentication, RBAC, CI/CD, and automated testing at 98% coverage, enabling rapid delivery of every feature that followed.
Stack
Industry Collaborated Dissertation
MSc Research Partnership · Edinburgh, United Kingdom
"MSc thesis, in collaboration with Amazon: measuring and mitigating how code-based LLMs memorize their training data, and the privacy and copyright risks it creates."
Research question. Investigated memorization in code-based large language models, the tendency to reproduce training data verbatim, which drives privacy, PII-leakage, and copyright risk in AI code generation. Awarded a distinction (82%).
Cross-architecture study. Evaluated all three model families, encoder-only, decoder-only, and encoder-decoder (CodeBERT, CodeGPT, CodeT5), on the CodeSearchNet dataset across four programming languages (Python, Java, JavaScript, Ruby), processing 30,000 samples per language.
Novel evaluation methodology. Designed two data-extraction attacks (masked-token prediction and prefix–suffix generation) and proposed an extended CodeBLEU metric with custom privacy components that detect exact reproduction of variable names, values, string literals, and comments via AST parsing.
Key findings. Showed that identifiers and string literals are the most memorized tokens (up to 79.3% exact match) and that verbose languages such as Java memorize more than concise ones like Python; 8-bit quantization consistently reduced exact memorization while preserving code quality.
Few-Shot & prompt tuning. Analysed how Few-Shot Learning shifts the balance between memorization and generalisation, uncovering non-monotonic, context-dependent effects where more examples do not always help.
Tech Pillars
Research Focus
LLM Memorization · Privacy · Copyright
Models & Data
CodeBERT · CodeGPT · CodeT5 · CodeSearchNet
Methods
MLM & NTP Attacks · Extended CodeBLEU · AST
Analysis
8-bit Quantization · Few-Shot Learning
03 / Projects
Six flagship builds spanning NLP, computer vision, robotics and machine learning, each documented end to end, from method to measured result. Nine more sit in the catalogue below.
P.01
My undergraduate final-year thesis: a Raspberry Pi self-driving car that cooperates with no central server. It parses a spoken destination, resolves an offline route across a dual-layer Folium + Google-Earth map, and plans motion with the Haversine formula, while a YOLO/Darknet perception stack detects 205 traffic-sign classes (trained on 97K+ images) and a custom model reads traffic lights at 83% accuracy, rerouting in real time on-device.
P.02
Built and benchmarked four German→English translation architectures on a low-resource Europarl subset, from an attentional LSTM baseline to a lexical model and a from-scratch Transformer with multi-head attention. The lexical model won decisively, converging in 53 epochs against the baseline's 99.
P.03
A low-cost ESP32-CAM rover that photographs an unknown maze, solves the shortest escape route with Dijkstra's algorithm over the pixel grid, converts it into a turn-by-turn instruction set, and drives it end to end, validated in a Pygame simulation before hardware deployment.
P.04
Implemented vanilla RNN and GRU language models from scratch in NumPy (forward pass, backprop and backpropagation-through-time) to study subject–verb agreement. A hidden-state cosine analysis showed the GRU far better preserves long-range syntactic dependencies.
P.05
Predicted whether a track would become a lasting chart hit from its audio features alone, across 650K+ rows of Spotify's Top-200. Benchmarked four classifiers plus K-means clustering. Random Forest led decisively, with energy and danceability the strongest drivers of success.
P.06
Built interpolated character-trigram language models from scratch (preprocessing, add-α smoothing and text generation) that identify a document's language purely by perplexity, then extended the study with four neural next-character models. The English model nailed the test document at a fraction of the rivals' perplexity.
All projects · The full catalogue
04 / Skills
Python · TypeScript · C++ · C# / .NET · Rust · SQL
PyTorch · TensorFlow · Hugging Face Transformers · OpenCV · ONNX Runtime
LangChain · RAG Pipelines · Vector Databases · Azure OpenAI · Prompt Engineering
MLflow · Azure ML · Docker · Kubernetes · CI/CD
Azure · AWS · PostgreSQL · MongoDB
React · Angular · Node.js · Git · Linux
05 / Education
Master of Science in Artificial Intelligence
Relevant Courses
Bachelor of Engineering in Electronics and Communication
Relevant Courses
06 / Volunteering
Electronics Engineers Association
Aug 2021 – Jul 2022 · Anna University · Chennai, India
Participants
National symposium
Coordinators
Across 14 events
Events & Workshops
Both symposiums
Electrofocus 2022
Ideated and organized a national-level Technical Symposium with 132 event coordinators across 14 events and 6 workshops, drawing 1,400+ participants.
Apocalypse 2021
Organized an intra-college Technical Symposium with 40+ coordinators and volunteers across 21 events and 3 workshops, with 400+ participants.
Finance & Audit
Managed financial records for the association and assisted in the end-of-year audit process.
MIT Robotics Association
Aug 2021 – Jul 2022 · Anna University · Chennai, India
Workshops Keynoted
ML · DL · ROS · RL
Robot Engineered
Swatchtha Bot
Competition Led
DEXBOT line-follower
Keynote Speaker
Delivered keynotes for the Machine Learning & Deep Learning workshop, tri-weekly Robotic Operating System (ROS) sessions, and a Reinforcement Learning workshop.
Swatchtha Bot
Led the technical team for the association's annual flagship project: a fully functional waste-collection robot.
DEXBOT Competition
Conducted and managed the line-follower robot competition. Active contributor across logistics and technical operations.
07 / Recommendations
Having worked with Vaikunth for over 6 months , originally as our intern, but now as a full time employee, he has quickly proved himself as a dedicated and very talented software engineer. From day one he has demonstrated his eagerness to learn, strong problem solving skills, and ability to learn and adapt to new challenges. A fantastic "can do" attitude, working quickly and efficiently to solve problems and handle task given to him, showing creativity and initiative to find solutions. He has quickly established himself as a key and valuable member of the Zot team.
Gordon Falconer
Director · Zot Engineering Ltd
I was one of Vaikunth's advisors for his MSc research project on LLM memorisation for code. From day one Vaikunth showed an incredible amount of passion and initiative, and, in Amazon terms, bias for action. He immersed himself into the relevant literature, and even in our first project meeting had done some preliminary experimentation. He continued at the same pace throughout the project, combining his hard work with innovative ideas that challenged existing approaches in the literature. I can easily recommend Vaikunth for any position that requires both deep exploration and technical execution.
Christos Christodoulopoulos
Principal Technology Adviser · ICO
09 / Contact
Whether it's a research collaboration, a full-time role, or an ambitious AI project, I'm always open to a conversation.
