AI/ML Developer · Building Intelligent Systems · Python & Full-Stack Engineering
I'm a fourth-year B.Tech student in Data Science & AI at SRMU, Lucknow, building AI-powered software end to end: from model training and evaluation to FastAPI backends and React interfaces. My strongest work so far pairs deep learning with explainability and monitoring, because a model is only useful if you can understand how it behaves.
Plant disease screening with MobileNetV2, Grad-CAM explainability and monitoring. Team project.
Python · TensorFlow/Keras · MobileNetV2 · FastAPI · Pydantic v2 · React 19 · TypeScript · Vite · Tailwind CSS · SQLite
Repository · No hosted demo (runs locally)
Architecture & details
Classifies a leaf image into one of 38 PlantVillage classes across 14 crop species. A scan returns ranked scores, a Grad-CAM overlay, latency and telemetry, with a clear distinction between a live backend response and a demo preview. ```mermaid flowchart LR UI["React + TypeScript dashboard"] --> API["FastAPI REST API"] API --> V["Validate + preprocess to 224x224"] V --> M["MobileNetV2 classifier, 38 classes"] M --> R["Top-5 scores"] M --> G["Grad-CAM overlay"] API --> T["Telemetry service"] T --> DB[("SQLite")] DB --> API ``` Features - Single-image upload with type and size validation, ranked top-5 predictions and a server-generated Grad-CAM overlay - Batch analysis of up to 24 images with per-file status, accepted/rejected counts and filter-aware CSV export - Monitoring dashboard: request volume, class distribution, mean score, p50/p95 latency and recent events - Model evaluation report read from a stored artifact, plus a searchable library of the 38 classes - Typed API boundary: Pydantic v2 contracts with explicit error codes Evaluation (PlantVillage held-out split, controlled benchmark) - 98.04% test accuracy on 8,146 test images; macro F1 97.47% - Weakest class in the stored report: corn gray leaf spot (F1 0.8514) - These figures are not field-condition accuracy Engineering considerations documented in the repo - A Grad-CAM failure never blocks a prediction - If the model can't load, the API returns model_unavailable instead of fabricated scores - Drift status is reported as unavailable until a verified baseline distribution exists - Class order in classes.json must match the trained model, or labels are silently wrong Limitations: closed-set 38-class model, scores are not calibrated probabilities, model artifacts are not committed, no authentication or production deployment configuration. Educational decision-support software, not a diagnostic tool.Screenshots: Overview · Scanner · Grad-CAM · Batch · Monitoring · Model report
Repository · No hosted demo
Architecture & details
Combines deterministic HR workflows with LLM-powered assistance in one dashboard. Capabilities below are as described in the repository README. ```mermaid flowchart TD UI["Streamlit dashboard"] --> API["FastAPI layer"] API --> AI["AI orchestration: LangChain, Claude, Ollama"] AI --> KB["Policy knowledge base: Pinecone RAG"] AI --> RE["Resume engine: ATS-style scoring"] API --> D["Employee data, analytics, SQLite"] ``` Features - AI HR assistant with multi-route query handling, designed around Claude with a local fallback - Resume intelligence: PDF parsing, ATS-style scoring, skill extraction, candidate ranking and downloadable reports - HR policy retrieval and semantic search using RAG over a Pinecone index - Workforce analytics: headcount, department insights, leave utilization and AI-generated summaries - Employee operations: leave balances, payroll previews and HR request workflows Roadmap listed in the repo: PostgreSQL migration, more advanced RAG pipelines, candidate-job matching, multi-tenant architecture.Status: documented architecture and roadmap; no hosted demo.
React · Vite · Tailwind CSS · Node.js · Express · MongoDB · Groq API · JWT · Nodemailer · Docker
Repository · No hosted demo
Architecture & details
```mermaid flowchart LR C["React + Vite client"] --> S["Express API"] S --> DB[("MongoDB")] S --> G["Groq API"] S --> E["Nodemailer"] ``` In the codebase - Generation of outreach messages through the Groq API - Signup, login and OTP verification flows with JWT and bcrypt - MongoDB models for users and email history - Dockerfiles and a Docker Compose configuration Roadmap (from the repo): verify MongoDB connectivity, test auth and OTP end to end, test Groq generation and email delivery, add automated tests and a CI workflow, then verify deployment.Status: under active development; some features and deployment settings still need end-to-end testing.
Browser-based ML workflow: upload a CSV, train a model, review metrics, predict.
Python · FastAPI · scikit-learn · Pandas · NumPy · React · Vite · Axios
Repository · Live demo (may take a moment to wake up)
Architecture & details
```mermaid flowchart LR FE["React + Vite"] -->|"HTTP / JSON"| BE["FastAPI"] BE --> ML["Pandas, NumPy, scikit-learn"] ``` Features - CSV upload with dataset inspection (rows, columns, feature and target columns) - Training routes for KNN, logistic regression, decision tree, random forest and linear regression - Accuracy and training metrics shown in the dashboard, with loading and status feedback - Prediction interface (a KNN prediction route is documented) Implementation notes - One route module and one model module per algorithm, plus a shared dataset service, keeps each workflow modular - A preprocessing route and module are included in the backend Roadmap listed in the repo: model persistence, more evaluation metrics, confusion matrices, hyperparameter tuning, prediction for all trained model types.Repository · Live demo
Architecture & details
```mermaid flowchart LR U["React + Vite client"] -->|"video or playlist URL"| S["Express API"] S --> Y["YouTube Data API v3"] U --> L[("Browser localStorage")] ``` Features - Converts playlists into multi-module courses and splits long videos into sections - A lesson can only be marked complete after 90% watch progress - Course progress as a completion percentage, with a dashboard of bootcamp cards - Interactive course player with sidebar navigation Implementation notes - The backend generates the bootcamp structure; the frontend saves the user, bootcamps and completed lessons in localStorage, so progress resumes in the same browser - No authentication or database yet (both listed on the roadmap, along with AI summaries and chapter titles)Repository
Architecture & details
Features - Lift logging: exercise, weight, sets, reps and notes - Progress graphs by week, month or all time - Strength rank: percentile comparison against general lifting standards for your bodyweight - Calculators: one-rep max estimator, BMI and body fat percentage (Navy method) Implementation notes - Single index.html, no framework and no build step; Chart.js is loaded from a CDN - All data stays in the browser's localStorage; nothing is sent anywhere - Planned: CSV export, bodyweight tracking over time, custom exercise names, installable PWA
| Languages | |
| AI & ML | |
| Backend | |
| Frontend | |
| Databases | |
| Tools | |
| Learning |
Where I've used each tool
- TensorFlow / Keras, MobileNetV2, Grad-CAM: Vistara AI - scikit-learn, Pandas, NumPy: ML Studio, Titanic classification (Matplotlib and Jupyter / Colab as well) - FastAPI, Pydantic: Vistara AI, Orchestrix, ML Studio - LangChain, Pinecone (RAG), Streamlit, Plotly: Orchestrix - Groq API (Llama 3): Wbot, PitchPilot - React, Vite: Vistara AI, ML Studio, LearnFlow, PitchPilot (TypeScript and Tailwind CSS in Vistara AI) - Node.js, Express: LearnFlow, PitchPilot - SQLite: Vistara AI, Orchestrix - MongoDB, Docker (still learning): PitchPilot - Vanilla JavaScript, Chart.js: Strenzz
Live data from public GitHub activity. Language percentages reflect code volume, not proficiency.
Browse my solutions → Currently Exploring
Early in the journey, building in public.