HomeResearch

Research Labs

/experimental/reports

Deep-dive systems research in distributed architectures, fog computing, and AI/ML. Each project is backed by rigorous simulations, technical reports, and quantifiable metrics.

Distributed Systems

Fog Computing, Edge Networks

UAV Networks

Mobility, Caching, RL Agents

AI/NLU

NER, Intent Classification

Project Completed#01

LOKI (NLP Engine)

Local-First Voice Assistant with Hybrid NLU

0.9977

Weighted NER F1

Local

Inference

Embeddings

Intent Fast-Path

Overview

A privacy-centric voice assistant that performs 100% local inference. It processes speech and commands locally using a dual-layer intent classification system: utilizing high-speed vector embeddings for common commands and falling back to a local Quantized LLM for complex semantic understanding.

System Architecture

Interactive Architecture

Technical Decisions

Decision	Trade-off	Outcome
Hybrid Classifier Strategy	System Complexity vs. Responsiveness	Implemented a 'Fast Path' (Cosine Similarity > 0.6) for common commands, reserving the heavy LLM only for complex, novel queries.
CRF for NER	Modernity vs. Efficiency	Chose Conditional Random Fields over BERT for Named Entity Recognition to minimize CPU footprint while maintaining 99.77% F1-score on parameter extraction.
Threaded Architecture	Dev Overhead vs. UX	Decoupled audio acquisition (VAD) and inference workers from the UI thread, ensuring the application remains responsive during heavy processing.

Tech Stack

PythonFaster-WhisperSentence-TransformersOllama (Dolphin-Phi)sklearn-crfsuite

Key Metrics

Weighted NER F10.9977

InferenceLocal

Intent Fast-PathEmbeddings

Resources

Source Code

Simulation Study#02

MUCEDS

Multi-UAV Cost-Efficient Deployment Scheme

18%→78%

Cache Hit Ratio

↓53%

Avg Task Latency

~300%

Profit vs Baseline

Overview

A comprehensive optimization framework for UAV-assisted Vehicular Edge Computing. It integrates high-fidelity traffic simulation (SUMO) with a Hierarchical Reinforcement Learning (HRL) agent to dynamically position UAVs and a Spatial-Temporal LSTM to predict content demand. In the reported simulation evaluation, cache hit ratio improved from 18% to 78% and average task latency fell by 53% compared with the baseline. These are simulation results, not production measurements.

System Architecture

Interactive Architecture

Technical Decisions

Decision	Trade-off	Outcome
Hierarchical RL	Convergence Speed vs. Control	Decomposed the problem: DDQN manages the 'Strategic' fleet size, while MADDPG handles the 'Tactical' velocity vectors, stabilizing training.
Predictive Caching (LSTM)	Compute Overhead vs. Backhaul Load	Proactive content caching based on predicted vehicle trajectories improved Cache Hit Ratio from 18% (Zipf baseline) to 78%, with 53% lower average task latency.
SUMO Physics Integration	Simulation Speed vs. Accuracy	Replaced synthetic Random Waypoint models with real OpenStreetMap data (Delhi/Mumbai), to evaluate the approach with modeled urban traffic flows.

Tech Stack

PythonPyTorchSUMO (TraCI)DDQN & MADDPGSpatial-Temporal LSTM

Key Metrics

Cache Hit Ratio18%→78%

Avg Task Latency↓53%

Profit vs Baseline~300%

Resources

Source Code

Testbed Validated#03

Fog Computing Testbed

Gateway Validation Module Placement (GVMP)

95%

Network Reduction

~16ms

Edge Latency

~5600ms

Cloud Latency

Overview

A 4-tier distributed Fog Computing virtual testbed built with Docker to evaluate IoT application placement strategies. The study validates the 'GVMP' heuristic, which prioritizes horizontal resource sharing (sibling nodes) over vertical cloud offloading, using realistic network emulation.

System Architecture

Interactive Architecture

Technical Decisions

Decision	Trade-off	Outcome
Sibling-Aware Placement	Search Latency vs. Bandwidth	Validating resources on neighbor nodes via the gateway reduced core network usage by 95% compared to standard Edge-Ward placement.
Kernel-Level Emulation	Setup Complexity vs. Realism	Used Linux 'tc' and 'netem' within containers to inject real jitter/packet loss, proving the strategy holds up under degraded network conditions.
Cgroup Monitoring	Implementation Effort vs. Accuracy	Built custom monitors reading /sys/fs/cgroup to calculate normalized CPU load, ensuring accurate performance metrics across heterogeneous node types.

Tech Stack

Docker ComposeLinux Traffic Control (tc)Prometheus/GrafanaPython (Flask)Redis

Key Metrics

Network Reduction95%

Edge Latency~16ms

Cloud Latency~5600ms

Resources

Source Code