ak@neural:~
AVAILABLE_FOR_HIRESTUDENT AT NEWCASTLE UNIVERSITY

AI/ML Engineer · Cloud Architect · UI/UX Designer. MSc Advanced Computer Science student at Newcastle University. B.Tech Computer Science & Engineering (AI Specialisation) from AKTU. Ex-IBM AI/ML Engineering Intern (Completed). Building clinical-grade models and agentic AI infrastructure.

95.43%Best Acc.
0.9941Macro AUC
102+HF DL
3Live Models
model_card.jsonLIVE
entityAnimesh Kumar
statusStudent @ Newcastle University
locationNewcastle upon Tyne, UK
focusSeeking: Grad AI/ML · Research roles
84K+OCT SCANS
102+HF DL
3LIVE MODELS
best_modelEfficientNetV2L + TransUNet
best_acc95.43% · AUC 0.9941
edge_latency<100ms CPU (ONNX INT8)
mcp_servers13 live — FRIDAY OS
STATUS: ACTIVE · SEEKING GRAD ROLES2025
Research Journey / Compute to Deployment

From Compute to Clinical Deployment

// Four phases — RTX 4060 to live HuggingFace production

00COMPUTE & SCALE

Training at Scale on RTX 4060

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

84,495
Training Images
5
Random Seeds
FP16
Mixed Precision
CUDAAMPPyTorchOptuna
01ARCHITECTURE

Hybrid CNN-Transformer Design

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

95.43%
Accuracy
0.9941
Macro AUC
0.9244
Macro F1
EfficientNetV2L4×MHAXGBoost
Live Demo on HuggingFace Spaces ↗
02CLINICAL SAFETY

Explainability & Safety Engineering

Grad-CAM, SHAP, UMAP, MC Dropout (20 passes), Mahalanobis OOD. Temperature-scaled calibration: ECE 0.0024. UCUS triage: Monitor / Review / Urgent.

0.0024
ECE Calibration
1.34×
Uncertainty Ratio
3
Triage Bands
Grad-CAMSHAPUMAPMC DropoutMahalanobis OOD
03●LIVE IN PRODUCTION

INT8 Quantisation & Deployment

V2L: 510MB → 132MB (3.9×). V2S: 91MB → 24MB. Live Streamlit dashboard on HuggingFace Spaces. OCT Complete Pipeline end-to-end.

132MB
V2L INT8
3.9×
Compression
<150ms
Local Inference
ONNX INT8StreamlitHuggingFace
OCT Complete Pipeline ↗
Explainable AI / Retinal Informatics

Retinex Exploded CNN Layer Architecture

Research focused on safety-critical computer vision models. By stacking multi-scale CNN activations, self-attention maps, and Monte Carlo dropout uncertainty, we open the "black box" of AI diagnostic systems.

●Scroll down to expand CNN model layers along the Z-axis
✦ Click any layer card to cycle it to the front focus
LAYER_00: INPUT_OCT_SCAN
FOVEAL_DIP
SCALE: 1:1RESOLUTION: 512x512
LAYER_01: FEATURE_ACTIVATIONS
f_0
f_1
f_2
f_3
f_4
f_5
f_6
f_7
f_8
f_9
f_10
f_11
f_12
f_13
f_14
f_15
f_16
f_17
f_18
f_19
f_20
f_21
f_22
f_23
f_24
KERNEL: 3x3FILTERS: 64
LAYER_02: GRAD-CAM_SALIENCY_MAP
FOCUS: FLUID_ACCUMULATIONGRADIENT: dY/dA
LAYER_03: MULTI-HEAD_ATTENTION
HEADS: 4QKV_DIM: 64
EXPLAINABLE AI OUTPUTCALIBRATED
DIAGNOSTIC AUC (84K IMAGES)0.9941
EXPECTED CALIBRATION ERROR0.0024

Temperature scaling prevents overconfident predictions.

Key Projects / Case Files

Research → Production

// Live deployments with verifiable metrics

Hybrid CNN-Transformer

— Retinal OCT Classification

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 + XGBoost hybrid. Mahalanobis OOD, MC Dropout, 5-seed CV, temperature calibration.

// RESULT

Live Streamlit dashboard on HuggingFace. Sub-150ms GPU inference. Grad-CAM explainability. 102+ downloads.

// Confusion Matrix · 5-seed CV mean · row = actual · col = predicted
A↓ P→
CNV
DME
Drusen
Normal
CNV
97.1
0.8
1.3
0.8
DME
0.6
96.5
1.4
1.5
Drusen
1.8
1.1
92.8
4.3
Normal
0.5
0.9
2.8
95.8

Macro AUC 0.9941 · Macro F1 0.9244 · ECE 0.0024 · Calibrated via temperature scaling

TF/KerasEfficientNetV2LXGBoostGrad-CAMSHAPUMAPOptuna

Attention-Guided TransUNet

— OCT Fluid Segmentation

0.916IRF DSC
0.856SRF DSC
1.34×Uncertainty
3.9×INT8 Compress
// PROBLEM

Pixel-level segmentation of IRF, SRF, PED retinal fluids across 4 independent OCT datasets.

// APPROACH

Dual AttentionTransUNetL (127M params) with Source-Adaptive BatchNorm, Focal Tversky loss, novel UCUS triage.

// RESULT

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

PyTorchTransUNetFocal TverskyMC DropoutONNX INT8

Plant Disease Detection

— EfficientNetV2S Pipeline

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 two-stage transfer learning, pHash dedup, MC Dropout, Grad-CAM, family-aware splits.

// RESULT

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

EfficientNetV2STFLiteGrad-CAMMC DropoutUMAP
SYSTEMS · AGENTIC AI

Snitch MCP Server

— Agentic AI Infrastructure

MCPProtocol
<50msLatency
13Servers Live
LocalExecution
// PROBLEM

LLM agents lack direct, low-latency access to local OS environments — file systems, shell, and private vaults.

// APPROACH

Custom Model Context Protocol (MCP) server bridging local OS environments with LLMs. Exposes typed tool definitions via JSON-RPC 2.0 over stdio/SSE transports, enabling AI agents to read files, run processes, and query private knowledge bases.

// RESULT

Production-deployed across FRIDAY OS — powers 13 live MCP integrations including vault search, code execution, and real-time context injection. Demonstrates advanced systems engineering and agentic AI infrastructure design.

// LIVE MCP EXCHANGE
snitch-mcp — bashONLINE
agent >
MCP ProtocolJSON-RPC 2.0TypeScriptNode.jsstdio/SSEPythonObsidian API

AI Caption Generator

— RAG Pipeline

GITHUBSOON
// PROBLEM

Automating context-aware caption generation across LinkedIn, Twitter, and Instagram at scale.

// APPROACH

RAG pipeline with LangChain + FAISS + fine-tuned Llama-2-7b. Cross-platform tone adaptation.

// RESULT

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

Llama-2-7bRAGLangChainFAISS
Methodology / Build Process

How I Build AI Systems

// From raw data to clinical-grade production

01

Data & ETL

Preprocessing, augmentation, pHash deduplication, Optuna hyperparameter search.

PandasNumPyAirflowOptuna
02

Architecture

EfficientNetV2 backbones, Multi-Head Attention layers, XGBoost ensemble heads.

PyTorchTF/Keras4×MHA
03

Optimisation

INT8 quantisation, ONNX export, temperature calibration, uncertainty estimation.

ONNXTensorRTFP16INT8
04

Deployment

Dockerised FastAPI services, Streamlit dashboards, HuggingFace Spaces CI/CD.

DockerFastAPIStreamlitHuggingFace
Field Log / Experience

Where I've Shipped

// Verifiable deployments, not just titles

Jun 2025 — Aug 2025 (Completed)Remote / UK
<100ms latencyINT8 3.9× compressVPC/ELB

AI/ML Engineering Intern

IBM
  • Deploy ONNX FP32/INT8 models on AWS with VPC/ELB architecture for sub-100ms edge CPU inference.
  • Build production MLOps pipelines: CI/CD model versioning, A/B traffic splitting, automated rollback.
  • Network threat analysis using C3iHub cybersecurity tooling — full-stack security awareness.
  • Integrate LLM inference APIs with enterprise IBM Cloud services; validate latency SLAs.
AWSONNXDockerIBM CloudPythonMLflow
Jun 2023 — Jul 2023Kanpur, India
DST Govt. funded8-Week Programme

Cybersecurity Research Intern

C3iHub, IIT Kanpur
  • Completed an 8-week DST government-funded programme: vulnerability assessment, system hardening, and network threat intelligence analysis.
  • Conducted systems hardening and network threat analysis using vulnerability assessment tooling.
CybersecurityThreat IntelligenceSystem HardeningNetwork Analysis
May 2023 — Jun 2023Kanpur, India
99.9% UptimeRetinal CNN

Cloud & AI Research Intern

IIT Kanpur (E&ICT Academy)
  • Architected an AWS Virtual Private Cloud (VPC) with Elastic Load Balancing (ELB), achieving 99.9% uptime under high-traffic load testing.
  • Designed and trained a CNN pipeline for retinal disease detection, contributing to early-stage medical imaging research.
  • Built a Loan Approval Prediction system (Logistic Regression, Random Forest) to automate risk assessment workflows.
AWSVPCELBCNNPythonscikit-learn
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
  • Uncertainty Quantification & Model Calibration

// Certification Record

Featured — Oracle · AWS · Azure · Google

Skills / Capability Matrix

// Research → Systems → Production

Deep Learning

Core
PyTorchTensorFlow/KerasTransformersCNNsTransUNetEfficientNetV2

ML Engineering

Core
ONNX INT8/FP32MLflowOptunaGrad-CAMSHAPMC DropoutUMAP

GenAI / LLMs

Proficient
LangChainRAGFAISSLlama-2Prompt EngineeringHuggingFace

Cloud Infrastructure

Proficient
AWS VPCELBEC2S3IBM CloudVercelDocker

Agentic AI / MCP

Core
MCP ProtocolJSON-RPC 2.0stdio/SSETool DefinitionsAgent Orchestration

Cybersecurity

Proficient
C3iHub ToolingNetwork Threat AnalysisAnomaly DetectionSecure Software Dev

Languages

Core
PythonTypeScriptJavaScriptJavaSQLBash

Research Methods

Core
Medical ImagingStatistical TestingError Correction CodesCalibrationOOD Detection
Contact / Open to Opportunities

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

// Direct Contact

Email
LocationNewcastle upon Tyne, UK

// External Profiles

⌘KOpen command palette for quick navigation or try "run snitch" for a surprise.

// Send a Message