ANUJ
DATA×AI×ENGINEERING
INITIALIZING SYSTEM
ANUJ MUNDU
RETURN TO MAIN PORTFOLIO
LAB STATION // LIVE INFERENCE SANDBOX
// EXPERIMENTAL DIGITAL LABORATORY · CLUSTER 04

APPLIED AI LAB

Where theoretical computer science and deep learning mathematics meet production systems engineering. An interactive cybernetic testbed for live tensor operations, dynamic loss convergence simulations, 3D lattice topological projections, and computational kernels.

FRAMEWORK:PyTorch 2.5 + CUDA 12.6
ACCELERATION:TensorRT 10.4 FP16
ACTIVE EXPERIMENTS:6 Interactive Modules
01 // 3D TOPOLOGICAL NEURAL LATTICE PROJECTION
SYSTEMS IN MOTION // 3D NEURAL LATTICE SIMULATION

Click and drag or tilt mouse to rotate the multi-layer spatial tensor topology in 3D coordinate space.

DRAG TO ROTATE 3D PERSPECTIVE
5 LAYERS · 27 SYNAPTIC NODES
01 // INTERACTIVE CONVOLUTION & KERNEL ENGINE

Real-time 2D spatial cross-correlation testbed. Select an input signal tensor and convolution kernel to inspect feature map activations and receptive field projections.

OUTPUT RESOLUTION: 5×5 TENSOR

ReLU: f(x) = max(0, x)

Input Tensor I (7×7)Values: 0.0 - 1.0
0.1
0.1
0.1
0.9
0.9
0.9
0.9
0.1
0.1
0.1
0.9
0.9
0.9
0.9
0.1
0.1
0.1
0.9
0.9
0.9
0.9
0.1
0.1
0.1
0.9
0.9
0.9
0.9
0.1
0.1
0.1
0.9
0.9
0.9
0.9
0.1
0.1
0.1
0.9
0.9
0.9
0.9
0.1
0.1
0.1
0.9
0.9
0.9
0.9

Hover any output cell on the right to project its 3×3 receptive field.

Kernel K (3×3)W(m,n)
-1
0
1
-2
0
2
-1
0
1

Calculates horizontal spatial gradient; detects vertical structural edges.

Activated Feature Map (5×5)σ(I * K)
0
3.2
3.2
0
0
0
3.2
3.2
0
0
0
3.2
3.2
0
0
0
3.2
3.2
0
0
0
3.2
3.2
0
0
Hover cell to view receptive field60 FPS Real-time Math
02 // LIVE LOSS CONVERGENCE & OPTIMIZER TESTBED

Simulate PyTorch gradient descent dynamics under varied optimizer kernels and objective penalty criteria. Step through epochs to observe asymptotic loss descent and accuracy saturation.

STEP 3 / 10 ACTIVE

Decoupled Weight Decay Adam

Down-weights easy negatives (γ=2.0)

Training Loss: 0.64Val Accuracy: 83.1%
Max Epoch: 10
Epoch 1 (Cold Start)Epoch 5 (Inflection)Epoch 10 (Convergence)
Training Loss
0.64
Asymptotic L2 bound
Val Accuracy
83.1%
Hold-out test set
Gradient ||g||
0.94
Clipping threshold: 1.0
LR Schedule
Cosine
Warmup: 300 steps

Distributed Edge Inference &
Tensor Acceleration

SYSTEM TELEMETRY CONSOLE v4.2

Enterprise Neural Vision Pipeline

Real-time multi-agent inference tracking with 99.4% F1-score and distributed WebGPU tensor streaming.

14.2 ms

Inference Latency

99.4%

Precision Rate

1.2M+

Params Active

04 // FORWARD PROPAGATION TENSOR ACTIVATION
INTERACTIVE DEEP NEURAL NETWORK VISUALIZER

Hover input nodes or click simulate to observe forward feed-forward tensor activation propagation.

[INPUT TENSORS: 4-D]
x₀: Bounding IoU0.88
x₁: Edge Contrast0.74
x₂: Motion Velocity0.42
x₃: Feature Norm0.95
[HIDDEN DENSE (GELU): 6 NODES]
h_0
b: 0.42
h_1
b: 0.45
h_2
b: 0.07
h_3
b: -0.38
h_4
b: -0.48
h_5
b: -0.14
[SOFTMAX PROBABILITY]
ŷ₁: Target Detected (True)92.4%
Class: High Confidence DetectionIoU Overlap > 0.75
Framework: PyTorch Custom nn.Sequential with Cosine Learning Rate ScheduleLoss: Complete IoU (CIoU) + Binary Cross Entropy
05 // CONTEXTUAL SYSTEM ASSISTANT & KNOWLEDGE RETRIEVAL
ANUJ'S AI LAB // CONTEXTUAL SYSTEM QUERY

Query specific architectural trade-offs, data pipelines, or modeling decisions directly against project telemetry.

SUGGESTED ARCHITECTURAL INQUIRIES:
RESPONSE FOR: "Why did you select YOLOv5s over heavier modern detectors like YOLOv8x?"COMPUTER VISION

Production hardware constraints dictated the decision. Edge CPU nodes have a strict 30ms latency budget. While YOLOv8x provided a marginal +2.1% mAP gain, it increased tensor parameter size fourfold and consumed 3.8x more CPU cycles. YOLOv5s quantized to INT8 with ONNX Runtime executes deterministically in 24.8ms P95 latency with 89.4% mAP.

RELEVANT TECH:YOLOv5sONNX RuntimeINT8 QuantizationCPU Optimization
03 // REPOSITORIES & EXPERIMENTAL MICRO-PROJECTS
3 Interactive Production Kernels
EXP 01NumPy · State-Space

Vectorized Kalman Filter Tracker

Linear discrete-time stochastic estimation to interpolate occluded object centroids through 200ms sensor blackout drops.

Camera tracking active (100% visibility)
State: [x=120.4, y=45.2, vx=14.2, vy=-2.1]
P = diag([0.05, 0.05, 0.2, 0.2])
σ² = 0.041
EXP 02Cython · TreeSHAP

TreeSHAP Acceleration Kernel

Exact Shapley additive feature attribution optimized via recursive subtree pruning over 50,000 tabular instances.

Naive KernelSHAP:124.0 ms
Optimized TreeSHAP:8.4 ms
Click benchmark to profile speedup
14.7× GAIN
EXP 03ONNX · TensorRT

INT8 Quantization Benchmark

Post-training asymmetric calibration mapping FP32 tensors to 8-bit integers with minimal accuracy degradation.

Precision Format:FP32 Single-Precision
VRAM Footprint:168.4 MB
Inference Latency:28.4 ms
Accuracy Delta: Baseline 100%
-75% RAM
04 // HIGH-ASSURANCE SYSTEMS & ARCHITECTURAL BENCHMARKS
Rigorous Test Suites & Clean Architecture
419 TESTS (100% PASS)SOURCE REPO

Anuj AI Lab — Local Agentic Platform

Production-grade local AI engineering platform for agentic RAG, semantic retrieval, conversation memory, and autonomous tool-calling workflows.

FastAPIReact 19TypeScript 5OllamaChromaDB
128 TESTS (100% PASS)SOURCE REPO

RF Signal Classification & AMC Suite

Clean Architecture deep learning framework for Automatic Modulation Classification (AMC) across noisy wireless channels (Rayleigh fading & AWGN).

PyTorchSciPy SignalI/Q SpectrogramsRay / Optuna
HEXAGONAL ARCHITECTURESOURCE REPO

FuelEU Maritime Decarbonization Platform

Full-stack maritime compliance ledger calculating route GHG intensity, Compliance Balance (CB), and Articles 20/21 banking and pooling.

TypeScriptNode.jsPostgreSQLPrisma ORMReact
RABBITMQ + POSTGRESQLSOURCE REPO

Distributed Hyperparameter Tuning Framework

Decoupled scheduler-worker architecture distributing intensive hyperparameter trials across worker nodes with durable queues and fault tolerance.

RabbitMQPostgreSQLDockerPyTorchScikit-Learn

EXP 04 // GENERATIVE RF SPECTROGRAM & DSP WATERFALL

TECH-ART CANVAS

Live mathematical Fourier transformation synthesizing RF IQ constellations, AWGN channel noise, and dynamic spectrogram waterfall heatmaps.

QPSK•SNR: 14 dB•FC: 960 MHz
SPECTROGRAM: FFT BINS 96 · TIME STEPS 70
CHANNEL SNR (AWGN):14 dB
-6 dB (High Noise)+26 dB (Clean IQ)
CARRIER FREQUENCY:960 MHz
200 MHz1800 MHz
05 // FUTURE HORIZONS & EMERGING ARCHITECTURES
Autonomous Agents · Edge AI · Regulatory Tech
ACTIVE R&DSOURCE

Autonomous Agentic RAG & Dynamic Tool-Routing

Developing local autonomous agents capable of self-healing tool execution, iterative syntax error correction, and multi-hop semantic graph traversal via ChromaDB and local Ollama kernels.

KEY VALIDATION METRICS:
• < 42ms Gateway Routing• 419 Automated Tests• Zero Cloud Data Leakage
BENCHMARKEDSOURCE

Sub-Millisecond Edge Quantization & TensorRT

Converting complex PyTorch convolutional and transformer vision backbones into 8-bit integer (INT8) tensor representations via ONNX Runtime and TensorRT, cutting inference memory by 75%.

KEY VALIDATION METRICS:
• 3.01x Inference Speedup• -75% VRAM Footprint• < 0.4% Accuracy Delta
SPEC COMPLIANTSOURCE

FuelEU Maritime 2025–2050 Decarbonization Engine

Building domain-driven clean hexagonal architectures calculating GHG emission intensity limits, dynamic compliance balances, banking/borrowing surpluses, and pooled fleet penalty optimizations.

KEY VALIDATION METRICS:
• Hexagonal Architecture• Strict EU 2023/1805 Rules• Deterministic Financial Models
RETURN TO MAIN PORTFOLIO
ENGINEERED WITH RIGOR · ANUJ MUNDU // NIT BHOPAL