Projects
Selected systems and applied ML work
A mix of application engineering, model development, data analysis, MLOps, LLMOps, and computer vision projects.
QC-Leap
AI-Powered Manufacturing Quality Control System
Next.js, React, TypeScript, FastAPI, PostgreSQL, Redis, Celery, PyTorch, OpenCV, Docker
- Built a full-stack quality-control dashboard for defect detection model management, training, inference, and reporting.
- Implemented ML workflows with segmentation masks, Copy-Paste augmentation, class balancing, and MobileNetV3-Small classification.
- Developed FastAPI services with JWT auth, PostgreSQL models, Redis/Celery jobs, and real-time training progress updates.
SiteSage AI
Construction procurement intelligence copilot
Python, Streamlit, LangChain, Ollama, FAISS, SQLite, Scikit-learn, Plotly, Docker, Kubernetes
- Built a multi-page procurement dashboard for document upload, cited AI question answering, delay prediction, vendor risk, and alerts.
- Implemented local RAG with document chunking, HashingVectorizer embeddings, FAISS retrieval, SQLite persistence, and grounded-answer fallbacks.
- Developed Random Forest delay-risk scoring, vendor reliability ranking, schedule-impact alerts, demo datasets, tests, Docker, and Kubernetes manifests.
Satellite Image Cloud Removal
Generative AI reconstruction for remote sensing imagery
Python, PyTorch, ResUNet, GAN, OpenCV, NumPy, Pandas, Matplotlib, Kaggle GPU
- Built a cloud-removal pipeline mapping cloudy optical imagery and masks to reconstructed cloud-free satellite outputs.
- Prepared RICE1/RICE2 manifests, trained a mask-guided ResUNet with cloud-aware L1 and SSIM loss, and exported visual predictions.
- Reached about 22.74 dB PSNR and 0.796 SSIM on RICE2 testing, then explored GAN refinement and Bhoonidhi LISS-IV data automation.
Multilingual Speech Recognition
Routed ASR pipeline under compute constraints
Python, PyTorch, Hugging Face, Whisper, Faster-Whisper, Indic Conformer, Librosa, Torchaudio, Kaggle
- Built an English, Hindi, and Tamil transcription pipeline for a Kaggle speech recognition challenge.
- Added language identification and dynamic model routing across Whisper large-v3-turbo and Indic Conformer models.
- Implemented transcript normalization, WER evaluation, beam-search inference, FP16 GPU execution, caching, and Kaggle submission generation.
Multilingual Sentiment Analysis
LLM fine-tuning for Indian-language sentiment classification
Python, Hugging Face Transformers, Gemma, LoRA, PEFT, bitsandbytes
- Fine-tuned google/gemma-3-1b-it with LoRA/PEFT for binary sentiment classification across 13 Indian languages.
- Built preprocessing, label encoding, tokenization, training, validation, diagnostics, and Kaggle submission workflows.
- Used memory-efficient training with bitsandbytes and Hugging Face Trainer, reaching 0.9778 local validation Macro F1.
Low-Light Image Restoration
Denoising, illumination correction, and 4x super-resolution
Python, PyTorch, NumPy, Pandas, PIL, Matplotlib, Seaborn, U-Net, PixelShuffle, CUDA
- Built a PyTorch restoration pipeline for the RELLISUR/Kaggle dataset with denoising, illumination correction, and 4x super-resolution.
- Designed a custom SR U-Net with encoder-decoder blocks, skip connections, PixelShuffle upscaling, and refinement layers.
- Reduced private RMSE from a 41.8 baseline to about 18.7 using EDA-driven modeling, loss tuning, mixed precision, TTA, and calibration.
End-to-End MLOps Pipeline
Cloud-native ML deployment and observability
Python, Scikit-learn, MLflow, DVC, GCS, Feast, BigQuery, FastAPI, Docker, GKE, OpenTelemetry
- Built an Iris classification pipeline with cloud artifact storage, data versioning, feature store setup, and experiment tracking.
- Automated CI/CD with GitHub Actions, pytest, CML reports, Dockerized FastAPI serving, Kubernetes deployment, HPA, and load testing.
- Added model explainability, fairness checks, drift monitoring, guardrails, and LLMOps/MLSecOps experiments.
RAG Document QnA
Local retrieval-augmented document assistant
Python, FAISS, Streamlit, Gemini, NLP chunking, semantic embeddings
- Built a local RAG system that synthesizes grounded answers from proprietary documents with refusal behavior for unsupported queries.
- Optimized CPU-bound vector similarity search with FAISS flat L2 indexing and overlapping text chunking.
- Created a Streamlit dashboard for PDF ingestion, session-state memory, constrained Gemini prompting, and citation-aware responses.
System Threat Forecaster
System vulnerability classification
Python, Scikit-learn, Pandas, Matplotlib
- Built a classification model to predict system vulnerabilities.
- Implemented preprocessing, evaluation, and model performance analysis.
Household Services Platform
Role-based services marketplace
Vue.js, JWT, Celery, Redis, Mailhog
- Built a full-stack application with admin, customer, and professional roles.
- Developed APIs, authentication, async task handling, and workflow automation.
Influencer Marketing App
Brand and influencer collaboration backend
Flask, SQLite, Python
- Designed database schema and REST APIs for campaign and user data handling.
- Built backend flows to connect brands with influencers.
Business Data Analysis
Local dairy shop sales analysis
Python, Pandas, Excel, Matplotlib, Seaborn, SQLite, Flask
- Analyzed sales and customer data to identify trends and improve inventory decisions.
- Created visualizations to communicate insights clearly.
