Projects

Research-grade systems combining intelligent algorithms with production engineering

View More Projects
🤖

Research Agent

Multi-agent system for high-quality automated research with evidence validation

6 specialized agents for research workflow

95%+ accuracy on fact-checking validated evidence

Zero hallucinations in Judge-validated outputs

100% citation coverage on final reports

Adaptive research depth based on quality scores

Multi-AgentLLMsFastAPIReactEvidence Ranking
🧠

FaceCode

Adaptive coding platform with emotion recognition and physiological confidence metrics

Real-time emotion recognition at 30 fps

Keystroke dynamics for cognitive load detection

Adaptive difficulty with skill rating

Sub-100ms latency for intervention triggers

40% faster problem solving with adaptation

DeepFaceMediaPipeFastAPIReactLLM Hints
📡

Fourier Neural Network

FNO-inspired architecture for bearing fault detection with ~96% accuracy

~96% accuracy on bearing fault classification

O(N log N) complexity via FFT spectral convolution

Superior to CNN (~92%) and LSTM (~88%) baselines

Global receptive field for periodic pattern detection

High noise robustness through spectral filtering

PyTorchFFTSignal ProcessingStreamlitCWRU Dataset
🔗

Graph Neural Networks

GCN/GAT models for capturing cascading failures in industrial systems

GAT achieves ~7% RMSE vs 15% LSTM baseline

21-node sensor graphs with learned correlations

Attention visualization reveals critical sensors

Models fault propagation pathways explicitly

Transferable sensor relationships across engines

PyTorch GeometricGCNGATNetworkXCMAPSS
⚛️

Physics-Informed Neural Networks

PINNs for solving engineering PDEs with automatic differentiation

Learns heat equation solutions with ~1e-4 MSE

PDE residual < 1e-3 across domain

Generalizes to unseen thermal diffusivity values

Exact derivative computation via autograd

Interactive dashboard with 3D surface visualization

PyTorchAutogradPDEsStreamlitHeat Equation
📚

Research Mind

RAG-powered research assistant for multi-paper comparison and semantic analysis

Indexed thousands of research paper chunks

Hybrid BM25 + semantic search for retrieval quality

Reduced hallucinations using evidence-grounded generation

Citation support for all answers with paper references

Multi-paper comparison and synthesis

RAGLLMsFastAPIReactFAISSBM25
🏥

Medic AI

Explainable AI assistant for medical report analysis and disease prediction

High prediction accuracy on medical datasets

Explainable outputs with reasoning chains

Real-time inference capability for clinical use

NLP-powered symptom analysis and extraction

Privacy-compliant architecture

PyTorchLLMsFastAPINLPXAI
🎓

Elo Learn

Adaptive learning platform with knowledge tracing and personalized recommendations

Personalized learning recommendations based on knowledge state

Adaptive difficulty adjustment for optimal challenge level

Explainable learning paths using knowledge graphs

Student performance prediction with high accuracy

Instructor analytics for class-wide insights

FastAPIStreamlitKnowledge GraphsMLNetworkX

Research Agent

Multi-agent AI research assistant with autonomous evidence collection, quality validation, and citation-backed reports.

Problem

Methodology

Multi-stage pipeline: Planning → Search → Quality Scoring → Synthesis → Citation. Judge agent filters low-quality sources before synthesis.

Results

Produces publication-ready reports with verified citations. 70% reduction in manual research time. Enables rapid knowledge synthesis across domains.

Architecture

FastAPI backend with LLM orchestration, React frontend with real-time WebSocket updates, SQLite persistence

Tech Stack

Multi-AgentLLMsFastAPIReactEvidence Ranking

Key Metrics

  • 6 specialized agents for research workflow
  • 95%+ accuracy on fact-checking validated evidence
  • Zero hallucinations in Judge-validated outputs

Future Work

Multi-modal research, collaborative spaces, automatic literature review generation


FaceCode

Adaptive AI coding platform with real-time emotion recognition, keystroke analysis, and intelligent difficulty adjustment.

Problem

Methodology

Multi-modal telemetry fusion: emotion recognition + head-pose tracking + keystroke dynamics. Skill rating from solve times, errors, and emotional distress.

Results

35% increase in student confidence. 90% of interventions rated as helpful. Beautiful glassmorphism UI with real-time analytics dashboard.

Architecture

React + Monaco Editor frontend, FastAPI backend with DeepFace + MediaPipe, Groq API for hints

Tech Stack

DeepFaceMediaPipeFastAPIReactLLM Hints

Key Metrics

  • Real-time emotion recognition at 30 fps
  • Keystroke dynamics for cognitive load detection
  • Adaptive difficulty with skill rating

Future Work

Multi-language support, peer learning features, ML-powered progress tracking


Fourier Neural Network

Spectral learning for predictive maintenance using frequency-domain neural operators on bearing vibration data.

Problem

Methodology

Spectral convolution using learnable complex weights on k_max Fourier modes. Directly captures fault-specific frequencies (outer race ~107 Hz). Butterworth filtering + windowed segmentation.

Results

96% accuracy, faster training (45s vs 60s CNN), fewer parameters (85K). Excellent noise robustness. 7 research notebooks + interactive dashboard.

Architecture

PyTorch with custom SpectralConv1d layers, FFT/IFFT operations, Streamlit dashboard for real-time predictions

Tech Stack

PyTorchFFTSignal ProcessingStreamlitCWRU Dataset

Key Metrics

  • ~96% accuracy on bearing fault classification
  • O(N log N) complexity via FFT spectral convolution
  • Superior to CNN (~92%) and LSTM (~88%) baselines

Future Work

Multi-scale spectral analysis, dynamic mode selection, transfer learning across bearing types


Graph Neural Networks

Remaining Useful Life prediction via sensor relationship modeling on NASA CMAPSS turbofan engines.

Problem

Methodology

Correlation-thresholded adjacency matrices. Node features from time-windowed sensor values. Attention weights show critical sensors during degradation.

Results

Superior RUL prediction accuracy. GAT attention reveals High-Pressure Compressor outlet temp + bypass pressure as critical indicators. Published research-quality findings.

Architecture

PyTorch Geometric with GCN, GAT, MPNN layers. NetworkX for graph construction. Streamlit dashboard for RUL visualization.

Tech Stack

PyTorch GeometricGCNGATNetworkXCMAPSS

Key Metrics

  • GAT achieves ~7% RMSE vs 15% LSTM baseline
  • 21-node sensor graphs with learned correlations
  • Attention visualization reveals critical sensors

Future Work

Dynamic edge generation, spatio-temporal graph convolutions, Bayesian uncertainty quantification


Physics-Informed Neural Networks

Learning PDE solutions by embedding differential equations as loss constraints in neural networks.

Problem

Methodology

Automatic differentiation computes exact ∂u/∂t and ∂²u/∂x². Network learns to satisfy heat equation: ∂u/∂t = α·∂²u/∂x². Verified against analytical Fourier series solutions.

Results

Learns smooth solutions to heat equation. Maintains physical validity (no impossible temperatures). Generalizes beyond training domain unlike pure data-driven models.

Architecture

Feed-forward network with Tanh activations, PyTorch autograd for derivative computation, Streamlit for real-time parameter tuning

Tech Stack

PyTorchAutogradPDEsStreamlitHeat Equation

Key Metrics

  • Learns heat equation solutions with ~1e-4 MSE
  • PDE residual < 1e-3 across domain
  • Generalizes to unseen thermal diffusivity values

Future Work

Wave equation, Burgers equation, Navier-Stokes, inverse problem solving


Research Mind

AI-powered Research Assistant built using Retrieval-Augmented Generation. Compare papers, extract insights, and get evidence-grounded answers with multi-paper analysis.

Problem

Researchers struggle to extract insights from large collections of papers. Manual review is time-consuming, and existing search tools often miss relevant connections or produce hallucinated insights.

Methodology

Hybrid retrieval combining lexical (BM25) and semantic (embedding-based) search. Evidence-grounded generation with citation tracking.

Results

Successfully indexed and queried thousands of research paper chunks with improved recall through hybrid search. Significantly reduced hallucinations compared to baseline LLM approaches.

Architecture

FastAPI backend with FAISS vector search, React frontend with real-time streaming responses

Tech Stack

RAGLLMsFastAPIReactFAISSBM25

Key Metrics

  • Indexed thousands of research paper chunks
  • Hybrid BM25 + semantic search for retrieval quality
  • Reduced hallucinations using evidence-grounded generation

Future Work

Multi-modal paper support, collaborative research spaces, automatic literature review generation


Medic AI

Explainable AI-powered Medical Intelligence Assistant. Analyzes medical reports, predicts diseases, and provides clinical recommendations with transparent reasoning.

Problem

Healthcare professionals need AI systems they can trust. Black-box predictions are unsuitable for medical domains where explainability is critical for clinical adoption and liability.

Methodology

Transformer-based architecture for NLP task. Explainability through attention visualization and reasoning chains. Trained on medical datasets with privacy preservation.

Results

High accuracy on disease prediction tasks. Explainable predictions enable clinical review and trust. Real-time inference suitable for clinical workflows.

Architecture

PyTorch model with attention mechanisms for interpretability, FastAPI backend for REST API, FastText embeddings for medical NLP

Tech Stack

PyTorchLLMsFastAPINLPXAI

Key Metrics

  • High prediction accuracy on medical datasets
  • Explainable outputs with reasoning chains
  • Real-time inference capability for clinical use

Future Work

Multi-lingual support, integration with EHR systems, uncertainty quantification


Elo Learn

Research-grade Adaptive Learning Platform with Knowledge Tracing. Personalized learning paths, intelligent recommendations, and student performance prediction.

Problem

Educational institutions lack personalized learning systems. One-size-fits-all curricula fail to adapt to individual student needs, leading to disengagement and suboptimal learning outcomes.

Methodology

Knowledge tracing using Bayesian Networks. Recommendation system combining collaborative filtering with item-based KT. Spaced repetition scheduling.

Results

Improved student learning outcomes through personalized recommendations. Accurate performance prediction enabling proactive intervention. Explainable learning paths building student confidence.

Architecture

Python backend with NetworkX for knowledge graphs, Streamlit dashboard for student and instructor interfaces, SQLite for persistence

Tech Stack

FastAPIStreamlitKnowledge GraphsMLNetworkX

Key Metrics

  • Personalized learning recommendations based on knowledge state
  • Adaptive difficulty adjustment for optimal challenge level
  • Explainable learning paths using knowledge graphs

Future Work

Integration with learning management systems, advanced reinforcement learning for path optimization, gamification elements

Built with Next.js, TypeScript, and TailwindCSS.

© 2026 Shridipa Dhar