Projects

Featured Projects

Scientific Literature RAG System

AI-powered introduction generator for scientific papers

Live demonstration of the Scientific Literature RAG system generating introductions with semantic search and citation extraction.

A production-ready RAG system that automatically generates well-structured, literature-informed introductions for scientific papers. The system indexes research papers and leverages LLMs to synthesize relevant literature into comprehensive introductions with properly formatted citations.

Key Features:

• Multi-Provider LLM Support: Claude (Anthropic), OpenAI GPT-4/GPT-4o, Google Gemini
• SPECTER2 Embeddings: Domain-optimized scientific embeddings for superior semantic understanding
• GPU Acceleration: Metal GPU support for Apple Silicon, CUDA for NVIDIA, automatic CPU fallback
• Smart PDF Processing: Semantic chunking (500-word chunks with 50-word overlap)
• ChromaDB Vector Search: Persistent local storage with cosine similarity search
• Automatic BibTeX Citations: Extracts and formats references with sophisticated metadata extraction
• Interactive Dash UI: User-friendly web interface for literature exploration
• ROUGE Evaluation: Built-in evaluation metrics for assessing generation quality (F1 scores >0.5)

Performance:

• Embedding: 50-100 chunks/second on Apple Silicon
• Generation: 10-30 seconds per introduction
• Indexing: ~1 minute per 100 PDFs

Technologies: Python 3.11+, GPU, LLM, ChromaDB, SPECTER2, RAG, Dash
Status: Production Ready | Apache 2.0 License
→ View on GitHub

---

Speech-to-Text with Sentiment Analysis and Translation

Real-time multilingual processing pipeline

Detailed architecture of the Transformer model showing encoder-decoder structure with multi-head attention mechanisms.

A comprehensive end-to-end system integrating speech recognition, sentiment analysis, and neural machine translation. Built with from-scratch Transformer implementation demonstrating deep understanding of attention mechanisms and encoder-decoder architectures.

Key Features:

• Real-time Speech-to-Text: Audio capture and transcription using Vosk library (English)
• From-Scratch Transformer: Complete encoder-decoder architecture without pre-trained models
• Custom Multi-Head Attention: Manually implemented attention mechanisms with configurable heads
• Positional Encoding: Hand-crafted sinusoidal position embeddings
• Sentiment Classification: Bidirectional LSTM with 95% test accuracy
• Interactive Dash Web App: Real-time processing with visual feedback and export functionality
• Hyperparameter Optimization: Automated tuning with Optuna for both models
• BLEU Score Evaluation: Translation quality metrics and model assessment

Performance Metrics:

• Sentiment Analysis (BiLSTM): 95.00% test accuracy
• Translation (Transformer): 67.26% test accuracy, 0.52 BLEU score
• Real-time processing with instant speech recognition and translation

Architecture Highlights:

• Educational implementation demonstrating core concepts
• Full control over attention mechanisms and positional encodings
• Research-grade code suitable for experimentation
• Clean, well-documented implementation

Technologies: Python 3.11+, Transformer, Bidirection LSTM, TensorFlow, Keras, Vosk, Dash, Optuna
Status: Research/Educational | Apache 2.0 License
→ View on GitHub
→ View Transformer Model on HuggingFace
→ Try Interactive Demo on HuggingFace Spaces
→ View on Kaggle

---

Active Particles in 3D Confinement

GPU-accelerated molecular dynamics simulation

Performance comparison across CPU and GPU implementations showing up to 27× speedup on Apple Silicon.

Simulation of 1000 active Brownian particles under 3D cylindrical confinement, demonstrating collective motion and particle-particle interactions with Lennard-Jones repulsion.

A high-performance C++ simulation framework for modeling active Brownian particles (ABPs) under cylindrical confinement in 3D space. Implements Euler-Maruyama algorithm for Langevin dynamics with dual CPU/GPU implementations.

Key Features:

• Dual Implementations: CPU-only (OpenMP) and GPU hybrid (Metal) versions
• GPU Acceleration: Up to 27× speedup on Apple Silicon for large systems (N ≥ 1000)
• Langevin Dynamics: Proper stochastic noise generation with Euler-Maruyama integration
• 3D Cylindrical Confinement: Reflective boundary conditions for realistic physical constraints
• Lennard-Jones Interactions: Repulsive particle-particle forces with configurable strength
• OpenMP Parallelization: Multi-threaded CPU optimization for all particle counts
• Automatic CPU/GPU Selection: Smart threshold-based switching for optimal performance
• Numerical Stability: Robust safeguards prevent NaN overflow at high particle densities

Performance Benchmarks (Apple M2, 1000 timesteps):

Particles	CPU (1 thread)	CPU (OpenMP 6)	GPU (Metal)	GPU vs 1 CPU	GPU vs OpenMP
100	0.026s	0.103s	0.096s	0.3×	0.9×
200	0.075s	0.125s	0.115s	0.7×	1.1×
500	0.401s	0.269s	0.225s	1.8×	1.2×
1,000	1.514s	0.747s	0.445s	3.4×	1.7×
2,000	5.638s	2.389s	0.601s	9.4×	4.0×
5,000	34.393s	13.903s	1.262s	27.3×	11.0×

Physical Model:

• Active Brownian particles with self-propulsion
• Translational and rotational diffusion
• Repulsive Lennard-Jones potential
• Applications: bacterial suspensions, cell motility, colloidal systems

Technologies: C++17, OpenMP, Metal GPU Status: Research-Grade | Apache 2.0 License
→ View on GitHub

---

Personality Classification with Ensemble Learning

Production ML pipeline achieving Kaggle top 5%

Interactive dashboard for real-time personality classification with model explainability and performance metrics.

A production-ready machine learning pipeline for personality classification using ensemble learning, achieving top 5% (200/4,329) in the Kaggle Personality Classification Competition. Features modular architecture, automated hyperparameter optimization, and interactive visualization.

Key Features:

• Six Ensemble Stacks: Complementary ML algorithms (Random Forest, XGBoost, LightGBM, CatBoost, Neural Networks)
• Advanced Data Augmentation: SDV Copula-based synthetic data generation with quality control
• Automated Hyperparameter Optimization: Optuna-powered tuning for each stack (200+ trials for blending)
• Interactive Dash Dashboard: Real-time predictions, model explainability, and performance visualization
• Meta-Learning Approach: Logistic Regression ensemble combiner with optimized blending weights
• Out-of-Fold Predictions: Unbiased ensemble training with stratified cross-validation
• Production-Ready: Full test coverage, CI/CD pipeline, comprehensive logging

Ensemble Architecture:

Stack	Focus	Algorithms	Special Features
A	Traditional ML (Narrow)	RF, LR, XGBoost, LightGBM, CatBoost	Stable baseline
B	Traditional ML (Wide)	Same as A	Extended search space
C	Gradient Boosting	XGBoost, CatBoost	Tree specialists
D	Sklearn Ensemble	Extra Trees, Hist GB, SVM, Gaussian NB	Diverse mix
E	Neural Networks	MLPClassifier, Deep architectures	Non-linear patterns
F	Noise-Robust	Same as A + label noise	Improved generalization

Competition Results:

• Rank: 200/4,329 (Top 5%)
• Dataset: ~18,000+ training samples (with augmentation), 8 personality dimensions
• Cross-Validation: Stratified 5-fold CV for robust evaluation

Technologies: Python 3.11+, scikit-learn, XGBoost, LightGBM, CatBoost, Optuna, Dash, Plotly Status: Production Ready | Apache 2.0 License
→ View on GitHub

---

Industrial Audio Anomaly Detection

Unsupervised anomaly detection for industrial machine audio

Upload a .wav recording to the Gradio app and get a verdict (normal/anomalous), the pooled Mahalanobis anomaly score, and its percentile against the calibration distribution of normal clips.

Per-machine-ID mean AUC-ROC for all detector families across fan, pump, and valve, averaged over 3 seeds with ±1 std error bars. The embedding model dominates on valve and fan; the Dense AE leads on pump.

ROC curves on valve comparing oracle per-ID scoring against ArcFace pooled Mahalanobis scoring. The curves overlap: the deployment constraint of knowing the machine ID is removed without meaningful loss in detection performance.

Unsupervised anomaly detection on industrial machine audio, benchmarked on the MIMII pump dataset and DCASE 2020 Task 2 (fan, pump, valve at 0 dB SNR). Three detector families are compared in increasing order of sophistication: a Dense Autoencoder, a Transformer-VAE scored by ELBO and KL divergence, and an ArcFace-trained embedding model scored by Mahalanobis distance to the stationary distribution of normal embeddings. The deployed valve model requires no known machine ID at inference and reaches 0.947 AUC.

Key Features:

• Three Detector Families: Dense Autoencoder, Transformer-VAE (ELBO and KL scoring), and ArcFace embedding + Mahalanobis distance
• From-Scratch DSP: STFT, triangular mel filterbank, and patch extraction implemented in NumPy/SciPy, no librosa
• ArcFace Angular Margin Loss: Tightens per-machine-ID clusters on the hypersphere for reliable scoring without a known machine ID
• Pooled Mahalanobis Deployment Scoring: Single LedoitWolf distribution fitted across all machine IDs, no oracle ID required at inference
• Gradio Web App: Upload a .wav recording and get a verdict, anomaly score, and percentile against the calibration distribution
• HuggingFace Hub Integration: Models downloaded and cached automatically on first run
• CLI + Python API: iaad score, iaad ensemble, iaad evaluate commands, plus AnomalyDetector / EnsembleDetector classes
• Full Test Coverage + CI: pytest suite, Ruff linting, pre-commit hooks, GitHub Actions CI

Results (per-machine-ID mean AUC-ROC, chance = 0.500):

Detector	Fan	Pump	Valve
Dense AE	0.501	0.671	0.632
Transformer-VAE (recon)	0.501	0.667	0.591
Transformer-VAE (KL)	0.491	0.631	0.439
EMB per-ID (oracle)	0.714	0.645	0.966
EMB pooled Mahalanobis (deployed)	0.730	0.738	0.947

Datasets:

• MIMII pump (Purohit et al., 2019): single machine type, ~380 normal and 136 anomalous clips at 16 kHz
• DCASE 2020 Task 2 (Koizumi et al., 2020): fan, pump, and valve at 0 dB SNR, 4 physical machine units per type

Technologies: Python 3.11+, TensorFlow, Keras, Transformer, ArcFace, Gradio, HuggingFace Hub
Status: Production Ready | Apache 2.0 License
→ View on GitHub
→ Read the Blog Post
→ View Model on HuggingFace
→ View on Kaggle: MIMII Pump Baseline
→ View on Kaggle: DCASE 2020 Benchmark

---

Meeting Notes from Raw Transcripts

Two from-scratch transformers that turn a raw meeting transcript into notes

The Dash app: Model 2 restores punctuation and casing in a voice note, then Model 1 turns a raw speech-to-text transcript into decisions, action items (with owners) and key points, with the importance score of every sentence.

From a raw transcript to meeting notes: punctuation restoration (Model 2), sentence splitting, sentence scoring (Model 1) and fixed selection rules.

Two small transformer models built entirely from scratch with Keras 3 on the JAX backend: own WordPiece tokenizer, own architectures, no pretrained weights. Model 2 (7.3M parameters) restores punctuation and casing in lowercase speech-to-text output. Model 1 (20.9M parameters) reads a whole meeting and picks the sentences that matter, sorting them into decisions, action items with owner and deadline, and key points. Model 1 is trained on a gold-labelled meeting corpus I built for this project, with real AMI meetings as the headline test set. Both models are published on Hugging Face.

Model 1 is hierarchical: a sentence encoder (pretrained by masked-language modelling) reads each sentence on its own, then a meeting encoder lets the sentences attend to each other before the per-sentence heads.

Model 2 is a token-classification transformer: a 4-block pre-LayerNorm encoder reads a window of lowercase words and predicts one casing label and one punctuation label per word, without ever adding, removing or correcting words.

Key Features:

• From Scratch: WordPiece tokenizer, pre-LayerNorm transformers and custom Keras layers, no pretrained weights or tokenizers
• Hierarchical Meeting Model: Sentence encoder + meeting encoder over up to 256 sentences per window, with speaker and speaker-change embeddings
• Gold-Labelled Corpus: meeting-notes-extractive-v1, about 1,900 meetings and 260,000 labelled sentences (hand-written synthetic meetings with hard negatives, plus real AMI meetings)
• Teacher Distillation Experiment: Qwen3-0.6B (reasoning labels) and a fine-tuned RoBERTa-Large as teachers, compared with gold-only training
• Owners, Deadlines and Replies: Span head for owner / deadline, and rules that attach short replies (“OK, I will email him tonight.”) to the request they answer
• Interactive Dash App: Punctuation restoration and meeting notes side by side, with an importance chart, served by gunicorn in Docker
• Self-Contained Releases: Each Hugging Face repo ships its own modeling.py, so the models run with only keras, jax and tokenizers
• Production Practices: Every hyperparameter in versioned configs, 3 seeds per run, model selection on validation only, full test suite, pre-commit and GitHub Actions CI

Results (Model 1, test set, importance F1, mean ± std over 3 seeds):

System	Params	Pretrained	AMI (real meetings)	Synthetic meetings
Lead-k (first sentences)	–	–	0.319	0.337
Qwen3-0.6B (prompted)	600M	yes	0.318	0.403
DistilBERT (fine-tuned)	66M	yes	0.580	0.870
RoBERTa-Large (fine-tuned)	355M	yes	0.634	0.903
Ours + distillation (E4)	20.9M	no	0.567 ± 0.006	0.884 ± 0.006
Ours, gold labels only (released)	20.9M	no	0.578 ± 0.006	0.878 ± 0.013

Key Findings:

• From scratch and with a third of the parameters, Model 1 matches a fine-tuned DistilBERT on real meetings, 5.6 points below RoBERTa-Large (17× larger)
• Distillation did not help: training with the Qwen and RoBERTa teachers was slightly worse than gold labels alone, so the released model uses no teacher
• Model 2 reaches 0.789 punctuation macro-F1 and 0.965 casing accuracy on DialogSum (rule baseline: 0.093), and stays at 0.662 with 20% of the words corrupted
• Limits, stated plainly: the notes are extractive (a decision spread over several turns is not written out), and decision / action-item F1 on real AMI meetings stays below 0.1 for every system tested

Technologies: Python 3.11+, Keras 3, JAX, Transformer (from scratch), WordPiece tokenizer, Dash, Docker, Hugging Face Hub
Status: Production Ready | Code: MIT License · Models: CC BY-NC-SA 4.0
→ View on GitHub
→ View Meeting Notes Model on HuggingFace
→ View Punctuation & Casing Model on HuggingFace
→ View Dataset on HuggingFace
→ View Dataset on Kaggle

---

Open Source Contributions

All public projects available on GitHub