Software engineer with a strong foundation in Android development (Java & Kotlin), now building reactive, multi-tenant backend systems in Kotlin with Spring Boot and WebFlux, alongside TypeScript + Next.js frontends. I also build and evaluate machine learning models in Python, with a focus on applied, production-minded ML rather than notebooks that never leave the notebook.

Also shipped production services in Node.js (NestJS, Express) and MongoDB before moving to the Kotlin/Spring stack.

The Kotlin + Spring Boot service behind FarmerXential, built on WebFlux and designed for tenant isolation from the request boundary upward.

• Non-blocking request pipeline with WebFlux and Kotlin coroutines
• Tenant and request context propagated through the Reactor Context so identity survives every async hop
• WebFilter-driven context seeding, with Spring Security handling authentication at the edge
• Owns auth, tenancy and rate limiting for the ML service, wrapping every inference call in Resilience4j circuit breakers, retries and bulkheads — degrading to a static fallback when the model is unreachable
• Distributed tracing and correlation via OpenTelemetry

Kotlin · Spring Boot · WebFlux · Project Reactor · Resilience4j · R2DBC · PostgreSQL

An early-warning platform for agricultural extension officers, pairing a typed web application with a dedicated ML inference service.

Web platform

• Dashboard, alerts, risk map, reports and settings, built on a typed data layer
• Full farmer management module — searchable table, detail panel, add-farmer flow and full-screen profile view
• Regional focus with location-aware risk modelling

Next.js 14 · TypeScript · React · Tailwind CSS

ML service — AgroTech-ml-service

FastAPI inference service that turns a farmer record — or a bare pair of GPS coordinates — into a decomposed agronomic risk assessment.

• Risk composed as Hazard × Exposure × Vulnerability, so each term is separately computable and checkable: drought/disease/heat hazards from FAO-56 water balance and BLITECAST epidemiology, exposure priced in USD, vulnerability from weighted coping capacity
• Calibrated HistGradientBoostingClassifier contributes a vegetation hazard term — trained on 7,527 real samples across 185 sites and 6 Sub-Saharan clusters, evaluated with leave-one-cluster-out and forward-chaining CV against three baselines
• Sentinel-2 optical and Sentinel-1 radar remote sensing plus ERA5 reanalysis, with a nightly precompute job so a live prediction is a database read and a model call
• Prediction ↔ outcome audit trail in Postgres that closes the loop back into training labels
• Protected attributes are rejected by the vulnerability model, not merely unused — a system allocating extension visits and credit must not route them by gender or education

Python · FastAPI · scikit-learn · pandas · PostgreSQL · Sentinel-2/1 · Docker · GCP

• Restaurant Order Classifier — predicts delivery success from order and logistics features
• Exploratory Analysis Notebooks — feature engineering, model comparison and evaluation with pandas, scikit-learn and seaborn

Granular access control service — role hierarchies, menu-level permissions and DI-driven service composition.

More on my GitHub repos →

• 📧 Email: iamshankarbhandari@gmail.com
• 💼 LinkedIn: linkedin.com/in/shankarbhandari2057
• 🖥️ Stack Overflow: stackoverflow.com/users/20598818/shankar-bhandari

Thanks for stopping by — always open to learning, collaborating, and building something meaningful 🚀