Aravindkumar K S — AI/ML & Software Engineer, Cork

About

I'm an AI/ML engineer in Cork. Most of my work so far has been on retrieval systems and the tooling around LLMs: getting the right documents in front of the model, checking what comes out, and keeping it running.

I spent two years at LTIMindtree in Bengaluru, starting as a graduate engineer trainee. Generative AI took off just after I joined, so I learned a lot of it on the job: a production RAG assistant for insurance documents, the extraction pipeline and knowledge graph behind it, and Studio, the team's model portal, with the gateway and guardrails around it, all running on AWS.

In 2024 I moved to Ireland for an MSc in Mathematical Modelling and Machine Learning at UCC, because I wanted to understand the models better than I did. My research looked at making Kolmogorov–Arnold Networks smaller and faster to train.

Since finishing, I've built projects of my own (an agentic RAG assistant, a self-hosted vLLM inference service and an LLM evaluation harness), earned the AWS ML Engineer certification, completed LangChain and Anthropic MCP courses, and mentored first-time programmers at Django Girls Cork. I also work with coding agents such as Claude Code, including MCP servers, skills, hooks and subagents.

Experience

Independent Projects & Professional Development

Cork, Ireland

AI/ML Engineering Oct 2025 – Present

Designed and built production-grade AI systems to deepen applied LLM and MLOps expertise — an agentic RAG assistant, a self-hosted vLLM inference service, and a CI-integrated LLM evaluation harness (detailed under Projects).
Earned the AWS Certified Machine Learning Engineer – Associate certification and completed LangChain and Anthropic Model Context Protocol (MCP) courses, formalizing cloud ML and agent-tooling skills.
Mentored first-time programmers as a Django Girls volunteer and stayed engaged with the AI engineering community while pursuing full-time AI Engineer / Forward Deployed Engineer / Software Engineer roles

LangChain
vLLM
RAGAS
DeepEval
Langfuse
AWS
Model Context Protocol (MCP)

LTIMindtree (now LTM)

Bengaluru, India

Senior Consultant – Package ImplementationInsurance R&D Oct 2022 – Aug 2024

I worked on GenAI projects in insurance R&D, from first proof of concept to production.

Led the proof-of-concept and co-built a production insurance RAG assistant (LangChain, FastAPI), cutting bulk policy-document processing time ~30% including manual extraction-quality review, at a 4–5s p95 time-to-first-token.
Owned an event-driven extraction pipeline (Textract, S3, SQS with dead-letter queues, Lambda consumers, Step Functions) parsing 20+ vendor formats into a Neo4j knowledge-graph insurance ontology at 95%+ accuracy, auto-quarantining malformed inputs.
Co-built Studio, the team's central model portal (React, FastAPI) providing model access, management and governance — rate limiting, PII redaction, content moderation, domain gating and application-based access control.
Contributed to evaluation baselines using a RAGAS harness and Langfuse tracing, wiring context-relevance and faithfulness checks into CI as release gates to catch retrieval regressions.

Built ingestion, chunking and embedding pipelines feeding Chroma-based hybrid retrieval (dense vectors + BM25 via Reciprocal Rank Fusion), sustaining bulk re-index throughput through SQS-buffered Lambda workers, with versions tracked in MLflow.
Independently built a VS Code extension on Studio bringing LLM chat, code generation, review and editing, model switching and history into the IDE for the engineering team.
Contributed to the LLM inference gateway behind Studio (FastAPI edge, internal gRPC/protobuf), adding a Redis semantic cache that cut redundant LLM calls ~5% and helping enforce GDPR-compliant PII handling.
Contributed to migrating infrastructure-as-code from CloudFormation to Terraform, building reusable ECS/ECR/Lambda/S3 modules with least-privilege IAM and encryption, and cutting AWS spend ~10% through ECS right-sizing and autoscaling.
Contributed to containerized blue-green ECS releases via GitHub Actions and CodeBuild, with SonarQube and Trivy security gates enabling low-risk deployments across Dev, Test and Prod.

LangChain
FastAPI
RAG
Hybrid retrieval (BM25 + RRF)
Chroma
Neo4j
Knowledge graphs
Amazon Textract
AWS Lambda
S3
SQS
Step Functions
ECS
ECR
MLflow
RAGAS
Langfuse
React
gRPC/protobuf
Redis
PII redaction
Terraform
CloudFormation
GitHub Actions
CodeBuild
SonarQube
Trivy

LTIMindtree (now LTM)

Bengaluru, India

Graduate Engineer Trainee Jul 2022 – Oct 2022

Built Angular tooling for configuring data-pipeline scan jobs and reviewing classification results and privacy insights, giving analysts a self-serve interface to manage scans without engineering support.
Designed PostgreSQL data models with Spring Boot and Hibernate ORM for data sources, job-run history and classified elements, supporting reliable tracking of scan lineage across runs.
Wrote SQL queries and aggregations powering data-discovery and coverage-monitoring dashboards, giving stakeholders visibility into classification coverage across data sources.
Delivered maintainable features in sprint increments through Agile cycles and peer reviews, improving code quality and reducing review turnaround.

Java
Spring Boot
Hibernate
Angular
PostgreSQL
SQL
Agile

Projects & Research

Agentic RAG & Conversational Knowledge Assistant

An assistant that decides for itself whether to search documents, query a database or look something up on the web.

Python
LangChain
Tool calling
OpenAI
Pinecone
Hybrid search
Redis
Pydantic
FastAPI
Gradio
RAGAS

Problem: Most chatbots answer every question the same way, so they guess when the answer lives in a database or online.
Built: A LangChain tool-calling agent that routes each query across vector search, SQL lookup and web search, returning Pydantic-validated structured outputs over multi-turn, Redis-backed conversation state.
Hard part: Retrieval quality. I tuned chunking (semantic vs fixed-size, overlap) and hybrid search using OpenAI embeddings and Pinecone, and diagnosed retrieval failures like missed context through RAGAS context-recall analysis.
Result: An LLM-based response classification and filtering guardrail layer catches ungrounded or irrelevant responses before they're returned, improving answer grounding and quality.

LLM Evaluation Harness

A reusable test suite that scores models and prompts, and blocks a merge when the scores drop.

Python
pytest
RAGAS
DeepEval
Langfuse
OpenAI/Anthropic APIs
GitHub Actions

Problem: A prompt or model change can make answers worse without anyone noticing until users do.
Built: A reusable harness that scores models and prompts on curated datasets across accuracy, faithfulness, relevance and latency, with pluggable metrics and provider-agnostic adapters.
Hard part: Trusting the judge. I implemented LLM-as-judge and reference-based scoring with RAGAS and DeepEval, and calibrated the judge prompts against human labels to improve metric–human agreement.
Result: The harness runs in CI as a regression gate via GitHub Actions, blocking merges on score drops and logging per-run traces to Langfuse for diagnosis.

Optimized LLM Inference Serving

A self-hosted LLM service tuned to serve more requests, faster, on a single GPU.

Python
vLLM
FastAPI
PyTorch
Hugging Face Transformers
Docker
Prometheus

Problem: A naive Hugging Face pipeline makes poor use of the GPU: throughput is low and larger models don't fit.
Built: A self-hosted inference service with vLLM behind a FastAPI gateway, using continuous batching and PagedAttention.
Hard part: GPU memory and latency. I applied INT8/AWQ quantization and KV-cache tuning, then benchmarked throughput, latency percentiles and tokens/sec under concurrent load.
Result: ~3x throughput over a naive Hugging Face pipeline, ~50% less GPU memory and ~40% lower p95 latency, enough to fit larger models on a single GPU. Prometheus metrics drive autoscaling and catch performance regressions.

MSc Research: Wendland's CSRBF in Kolmogorov–Arnold Networks (KAN)

A smaller, faster version of a newer neural network architecture. Research project at UCC, under Dr. Andreas Amann.

PyTorch
Kolmogorov–Arnold Networks
Radial basis functions
Regularization

Repo

Problem: Kolmogorov–Arnold Networks learn their own activation functions, but they're expensive to train.
Built: KAN alternatives in PyTorch that replace B-spline and Gaussian edge activations with compactly supported Wendland radial basis functions (a lighter building block that only switches on locally), with trainable centers and support radii.
Hard part: Keeping trainable radii stable, which I did with inverse-L2 regularization, and quantifying the sparsity–accuracy trade-off.
Result: Compared across six baselines.

Tab. 1 — benchmarks
Benchmark	Result
MNIST vs FC-KAN	97.6% vs 97.7% accuracy, 16% fewer parameters, 21–23% faster training
Fashion-MNIST	86.7% → 88.7% accuracy
Activation density vs Fast-KAN (fixed radii)	~25% vs 88%

More projects

Serverless Observability on AWS

A small serverless service where I break things on purpose to check the monitoring catches it.

Terraform
AWS Lambda
API Gateway
SQS
DynamoDB
CloudWatch
SNS
KMS
IAM
Python

Problem: Serverless apps are easy to deploy and hard to debug when they fail at 3 a.m.
Built: A modular Terraform stack for a serverless order service with a sync path (API Gateway → Lambda → DynamoDB) and an async path (SQS → worker Lambda + DLQ), on an S3-locked remote backend.
Hard part: Making failures visible, through structured JSON logging, metric filters and EMF custom metrics, with saved Logs Insights queries for latency percentiles, cold starts and request tracing.
Result: Fault injection (error rate, latency, poison messages) reproduces failures and validates metrics and alerting end to end. Least-privilege IAM, PII masking via CloudWatch Logs data protection, and KMS-encrypted SNS alerting with scoped key and topic policies keep it secure.

Predictive Modeling for Customer Churn

A churn model that also explains why customers leave.

Python
pandas
scikit-learn
XGBoost
Optuna
SHAP
MLflow

Problem: Knowing who might churn isn't enough; a retention team also needs to know why.
Built: An end-to-end churn-prediction pipeline on 50k+ customer records, with behavioral and temporal features. I trained and compared logistic regression, random forest and XGBoost, tuned hyperparameters with Optuna and tracked experiments in MLflow.
Hard part: Class imbalance, handled with SMOTE and stratified cross-validation.
Result: 0.89 ROC-AUC. SHAP surfaced the top churn drivers, which I turned into retention recommendations for non-technical stakeholders.

Education & credentials

Education
Degree	Institution	Dates	Highlight
MSc Mathematical Modelling & Machine Learning	University College Cork, Cork, Ireland	Sep 2024 – Sep 2025	First Class Honours (1.1) Coursework: Deep Learning (ANN, RNN, VAE, Transformers, RL), Statistical Methods (Data Analysis, EDA, Regression, Classification, Unsupervised Models)
BE Mechanical Engineering	Kumaraguru College of Technology, Coimbatore, India	Jul 2018 – Jun 2022	First Class with Distinction (8.8/10.0 CGPA) Publication: “Design Optimization of Grids in Grid Fins for Spacecrafts”, International Journal of Vehicle Structures and Systems (IJVSS), 2022

Certifications
Certification	Issuer	Date
AWS Certified Machine Learning Engineer – Associate	Amazon Web Services	Sep 2026
Foundation: Introduction to LangChain – Python	LangChain	May 2026
Model Context Protocol: Advanced Topics	Anthropic	Apr 2026
Introduction to Model Context Protocol	Anthropic	Mar 2026

Awards

Shooting Star Award, LTIMindtree, Bengaluru (Apr – Jun 2024): recognised as a top team contributor for consistently delivering high-impact features on the Studio platform during Q1.
Rising Star Award, LTIMindtree, Bengaluru (Sep 2023): Young Achiever award for exemplary contribution in my first year at the company.

CommunityVolunteer Mentor, Django Girls Cork (Sep 2026). Coached a group of 3 first-time programmers through building and deploying a full Django blog application to PythonAnywhere in a single-day workshop.

Contact

I'm based in Cork and open to AI Engineer, Forward Deployed Engineer and Software Engineer roles in Ireland. LinkedIn is the best place to reach me, and my code is on GitHub.

Cork, Ireland · Stamp 1G, eligible to work in Ireland · Hybrid, onsite or remote within Ireland · English, Tamil

LinkedIn GitHub CV (PDF)