SYSTEM ONLINE | ⬣ IBM AI/ML INTERN 2025 | MSc NEWCASTLE UNIVERSITY | OPEN TO ROLES

AI/ML
Researcher
& Engineer

298 researchers downloaded my retinal OCT classifier.
The accuracy is 95.43% — but what they're actually using is the
safety layer underneath: uncertainty flags, OOD detection,
calibration to ECE 0.0024.

0 % ACCURACY
0 K+ OCT IMAGES
0 HF DOWNLOADS
0 CERTS
EMAIL ME LINKEDIN ↗
GITHUB 🤗 HUGGINGFACE ⬢ ORCID
ANIMESH-LOGIC-V4 // NEURAL CARTOGRAPHY LIVE
EMPLOYERIBM Corp. — AI/ML Engineering Intern (2025)
DEGREEMSc Advanced Computer Science — Newcastle University, UK
ARCHEfficientNetV2L + 4×MHA + Learnable Positional Enc.
STACKPyTorch · TensorFlow · CUDA · AWS · HuggingFace
STATUSOPEN TO GRADUATE ROLES (UK / GLOBAL)
0% ACCURACY
0MACRO AUC
0MACRO F1
ECE calibration0.0024
Dataset84,495 OCT retinal images
Cross-validation5-seed reproducible
ExplainabilityGrad-CAM · SHAP · UMAP
Safety layerMC Dropout · Mahalanobis OOD
Inference (local)<150ms on RTX 4060
FIELD LOG / PROFESSIONAL EXPERIENCE

Industry & Research Roles

// Chronological record of deployments, research contributions, and technical engagements

JUN — AUG 2025 IBM Corp. Remote
AI/ML Engineering Intern ⬢ IBM
IBM · AI Research Division · Remote
  • Developed and fine-tuned Large Language Models (LLMs) and Transformer architectures for enterprise-grade NLP tasks within IBM's AI research division.
  • Engineered prompt-tuning and inference optimisation workflows, contributing to production-ready Generative AI pipelines.
  • Collaborated cross-functionally with senior researchers on model evaluation, benchmarking, and MLOps deployment — from prototype to production.
LLMsTransformersPrompt TuningGenAIMLOpsEnterprise NLP
MAY — JUN 2023 IIT Kanpur India
Cloud & AI Research Intern
IIT Kanpur (E&ICT Academy) · Kanpur, India
  • Architected a Virtual Private Cloud on AWS with Elastic Load Balancing — 99.9% uptime under high-traffic simulation.
  • Built a CNN pipeline for retinal disease detection, contributing to early-stage medical imaging research.
  • Developed a Loan Approval Prediction system (Logistic Regression + Random Forest), automating risk assessment workflows.
AWS VPCELBCNNsMedical Imaging
JUN — JUL 2023 C3iHub Remote
Cybersecurity Research Intern
C3iHub, IIT Kanpur · Government-funded (DST)
  • 8-week programme: vulnerability assessment, system hardening, and network threat intelligence analysis.
SecurityHardeningVuln Assessment
OCT 2022 MedTourEasy Remote
Data Analyst Intern
MedTourEasy · Remote
  • Built a Logistic Regression model predicting blood donation behaviour across 10,000+ patient records.
  • Designed dashboards (Matplotlib, Seaborn) to communicate donor retention insights.
PythonPandasScikit-LearnTPOT
CREDIBILITY / ACADEMIC & CERTIFICATION RECORD

Education & Certifications

// ACADEMIC RECORD

MSc Advanced Computer Science

2025–2026
Newcastle University · UK

Distinction track. Deep Learning · Distributed Algorithms · IoT · Advanced Software Engineering. Dissertation in progress.

B.Tech CSE — AI Specialisation

2021–2025
AKTU · India

First Division, CGPA 7.11/10. Neural Networks · Computer Vision · Big Data Analytics · IoT.

// Research Focus Areas

  • Deep Learning & Representation Learning
  • Computer Vision & Model Explainability (Grad-CAM, SHAP, UMAP)
  • Generative AI, LLMs & Agentic Systems
  • GPU-Accelerated Training & Inference (CUDA, TensorRT)
  • Clinical AI Safety & Out-of-Distribution Detection
  • Hyperparameter Optimisation (Optuna), Ensemble Methods
TECHNICAL SKILLS / EVIDENCE-LINKED PROFICIENCIES

Skills & Expertise

// Every claim linked to a deployed system or measurable outcome

Computer Vision & Deep LearningADVANCED
PyTorch · TensorFlow/Keras · OpenCV · EfficientNetV2 · CNNs
  • 95.43% accuracy, 0.9941 AUC on 84K OCT images — view project
  • 99.57% test accuracy on PlantVillage (EfficientNetV2S, 38 disease classes)
  • Full explainability: Grad-CAM, SHAP, UMAP, attention maps, ablation studies
  • MC Dropout + Mahalanobis OOD for clinical-grade safety layers
  • 5-seed cross-validation with temperature-scaled calibration (ECE: 0.0024)
GenAI, LLMs & RAGADVANCED
Llama-2 · Transformers · LangChain · FAISS · Gemini API
  • Enterprise LLM prompt-tuning at IBM AI research division (2025)
  • Built end-to-end RAG pipeline with LangChain + FAISS vector search
  • Fine-tuned Llama-2-7b on custom domain data — 70% content-time reduction
  • Inference optimisation workflows for production GenAI solutions
Medical Image SegmentationRESEARCH
PyTorch · TransUNet · Focal Tversky Loss · ITK · ONNX
  • AttentionTransUNetL ensemble: V2L val DSC 0.784 ± 0.006 (IRF 0.916, SRF 0.856, PED 0.581)
  • UCUS clinical triage: Monitor / Review / Urgent — HF Spaces
  • Uncertainty 1.34× higher at error pixels (p=3.77×10⁻⁵)
  • INT8 quantisation: V2L 510MB → 132MB (3.9×)
Cloud Architecture & MLOpsPROFICIENT
AWS · Docker · Azure · Oracle Cloud · FastAPI · HuggingFace Spaces
  • VPC + ELB achieving 99.9% uptime under high-traffic load testing
  • Live Streamlit dashboards on HuggingFace Spaces (<150ms local inference)
  • Containerised ML services with Docker; CI/CD via GitHub Actions
  • AWS WebScaler — zero-downtime ONNX model serving
METHODOLOGY / SYSTEM ARCHITECTURE

How I Build AI Systems

// From raw data to clinical-grade production deployments

01

Data & ETL

Preprocessing, augmentation, Optuna-driven optimisation, pHash deduplication.

PandasNumPyAirflowOptuna
02

Architecture

EfficientNetV2 backbones, Multi-Head Attention, ensemble design, XGBoost hybrids.

PyTorchTF/Keras4×MHA
03

Optimisation

Mixed precision, INT8 quantisation, ONNX export, temperature calibration.

ONNXTensorRTFP16INT8
04

Deployment

Dockerised FastAPI services, Streamlit dashboards, HuggingFace Spaces production.

DockerFastAPIStreamlitHuggingFace
RESEARCH JOURNEY / COMPUTE TO DEPLOYMENT

From Compute to Clinical Deployment

// Four phases — RTX 4060 to live HuggingFace production

00
COMPUTE & SCALE

Training at Scale on RTX 4060

Hybrid EfficientNetV2L + 4×MHA trained with Automatic Mixed Precision on NVIDIA RTX 4060. Full 5-seed cross-validation across 84,495 OCT retinal images.

84,495TRAINING IMAGES
5RANDOM SEEDS
FP16MIXED PRECISION
CUDAAMPPyTorchOptuna
01
ARCHITECTURE

Hybrid CNN-Transformer Design

EfficientNetV2L backbone + Learnable Positional Encoding + 4× Multi-Head Attention + XGBoost hybrid head. 200+ Optuna hyperparameter trials. McNemar's statistical validation.

95.43%ACCURACY
0.9941MACRO AUC
0.9244MACRO F1
EfficientNetV2L4×MHAXGBoost
Live Demo on HuggingFace ↗
02
CLINICAL SAFETY

Explainability & Safety Engineering

Grad-CAM heatmaps, SHAP attribution, UMAP embeddings, MC Dropout uncertainty (20 passes), Mahalanobis OOD detection. Temperature-scaled calibration. UCUS clinical triage (Monitor / Review / Urgent).

0.0024ECE CALIBRATION
1.34×UNCERTAINTY RATIO
3TRIAGE BANDS
Grad-CAMSHAPUMAPMC DropoutMahalanobis OOD
03
LIVE IN PRODUCTION

INT8 Quantisation & Deployment

V2L model: 510MB → 132MB (3.9× compression). V2S: 91MB → 24MB. Two-tier Streamlit dashboard running locally at <150ms and publicly on HuggingFace Spaces.

132MBV2L INT8 SIZE
3.9×COMPRESSION
<150msLOCAL INFERENCE
ONNX INT8StreamlitHuggingFace Spaces
OCT Complete Pipeline ↗
KEY PROJECTS / CASE FILES

Research → Production

// Selected work with verifiable metrics and live deployments

Hybrid AI — Retinal OCT Disease Classification

GITHUB LIVE ↗
95.43%ACCURACY
0.9941MACRO AUC
0.9244MACRO F1
0.0024ECE
// PROBLEM

Diagnosing retinal diseases from 84,495 OCT scans at clinical-grade accuracy with full interpretability.

// APPROACH

EfficientNetV2L + 4×MHA + Learnable Positional Encoding + XGBoost hybrid. Mahalanobis OOD, MC Dropout, 5-seed CV, temperature calibration.

// RESULT

Live Streamlit dashboard on HuggingFace Spaces. Sub-150ms GPU inference. Full Grad-CAM explainability. 294+ downloads.

TensorFlow/KerasEfficientNetV2LXGBoostGrad-CAMSHAPUMAPOptuna

Attention-Guided TransUNet — OCT Fluid Segmentation

GITHUB LIVE ↗
0.916IRF DSC
0.856SRF DSC
1.34×UNCERTAINTY
3.9×INT8 COMPRESS
// PROBLEM

Pixel-level segmentation of IRF, SRF, and PED retinal fluids from OCT volumes across 4 independent datasets (DUKE, AROI, UMN AMD, UMN DME).

// APPROACH

Dual AttentionTransUNetL ensemble (V2L, 127M params) with Source-Adaptive BatchNorm, Focal Tversky loss, 2.5D multi-slice input. Novel UCUS clinical triage score.

// RESULT

V2L val Dice 0.784±0.006. SRF volume r=0.778, PED r=0.841. INT8: 510MB→132MB. Zenodo preprint archived.

PyTorchTransUNetEfficientNetV2LFocal TverskyMC DropoutONNX INT8

Plant Disease Detection — EfficientNetV2S

GITHUB LIVE ↗
99.57%TEST ACC
99.48%MACRO F1
38CLASSES
~45MBTFLITE
// PROBLEM

Early detection of crop diseases across 38 categories to prevent agricultural yield loss.

// APPROACH

EfficientNetV2S with two-stage transfer learning, pHash deduplication, family-aware splits, MC Dropout, Grad-CAM explainability.

// RESULT

McNemar p=3.27×10⁻¹⁸². TFLite float16 ~45MB. Interactive UMAP 3D on GitHub Pages. Live HuggingFace Spaces.

EfficientNetV2STFLiteGrad-CAMMC DropoutUMAPHuggingFace

AI Caption Generator — RAG Pipeline

GITHUB DEPLOYING SOON
// PROBLEM

Automating context-aware, platform-specific caption generation at scale.

// APPROACH

RAG pipeline with LangChain + FAISS + fine-tuned Llama-2-7b. Cross-platform tone adaptation (LinkedIn, Twitter, Instagram).

// RESULT

70% reduction in manual content-writing time. Continuous human-feedback evaluation loop.

Llama-2-7bRAGLangChainFAISSTransformers
CONTACT / OPEN TO OPPORTUNITIES

Let's Build Something

// Graduate AI/ML Engineering · Data Science · Research roles — UK & Global

// Direct Contact

EMAIL
kranimesh2004@gmail.com
LOCATION
Newcastle upon Tyne, UK
ORCID
0009-0003-0608-7004

// External Profiles

LinkedIn Profile GitHub Code 🤗 HuggingFace Models 🏅 Credly Badges 🏆 HackerRank Algo 🏆 CodeChef CP ☁️ Google Cloud Skills ⬢ ORCID Research
© ANIMESH KUMAR // NEWCASTLE UPON TYNE, UK akumar-tech.me