Projects in repo are closer to production than they appear.
Software engineer & AI developer — everything I build is meant to ship
Scroll down to explore
I build software that takes the routine out of people's work. I think AI should
handle the boring parts of a job — so I build the pipelines, APIs and interfaces
that let it. Full-stack by trade, systems-minded by habit, AI-curious by default.
I'm Yash Tembhurnikar.
Lines of code that made it to prod:
Too many to count
Purdue · Yansons · production systems · AI projects
Distance traveled to ship better software:
0+ miles
Pune → Singapore → West Lafayette
Probability I’ll keep debugging until it works:
99.9%
the other 0.1% just needs one more commit
Capabilities
Full-stack engineering, end to end.
Six areas of depth, applied across production systems.
The languages I use across production full-stack systems, backend services and data-heavy applications.
TypeScript
JavaScript
SQL
Python
Java
Primary
TypeScript · JavaScript · SQL
Also use
Python · Java
Applied in
Purdue · Yansons · Diagrams-AI
Production interfaces built around component architecture, state, accessibility and responsive behavior.
React
Next.js
Tailwind CSS
Bootstrap
Angular
React Flow
Primary
React · Next.js
Focus
Component architecture · accessibility
Applied in
Purdue · Diagrams-AI · Yansons
APIs, business workflows, authorization and data layers designed to behave predictably in production.
Node.js
Express.js
FastAPI
Prisma
Pydantic
REST APIs
Primary
Node.js · Express.js
Also use
FastAPI · Prisma
Applied in
Purdue · Yansons · Diagrams-AI
Relational and application storage spanning production databases, ORM-backed systems and client/cloud persistence.
PostgreSQL
MySQL
Firebase / Firestore
IndexedDB
Primary
PostgreSQL · MySQL
Cloud storage
Firestore
Client persistence
IndexedDB
Containerized applications, deployment pipelines and production delivery workflows.
Docker
Kubernetes
AWS
GitHub Actions
Rancher
Harbor
Primary
Docker · Kubernetes · GitHub Actions
Cloud
AWS
Production tooling
Rancher · Harbor
The engineering principles underneath the stack: correctness, maintainability, access control and system behavior.
Natural language in. Editable software diagrams out.
A full-stack AI platform that turns natural-language requirements into editable software and system-design diagrams. It supports 14 diagram types across UML, architecture, scalable systems, DevOps and product workflows. Gemini and GPT models generate structured output that is normalized into schema-validated JSON before specialized graph, sequence and architecture renderers display it.
A solo full-stack AI platform that turns natural-language requirements into editable software and system-design diagrams — 14 UI-supported diagram types spanning System Architecture, Flowchart, Sequence, ER, Class, Component, Deployment, Data Flow, Use Case, User Flow, User Journey, Cloud Architecture, Network and CI/CD Pipeline diagrams.
Pipeline: a natural-language prompt goes through Gemini/GPT orchestration, comes back as a validated JSON intermediate representation, is rendered by a diagram-specific renderer onto an editable canvas, and persists locally or to the cloud before export. The backend is Python 3.11, FastAPI and Pydantic v2; the frontend is Next.js 14, TypeScript and React, with React Flow for graph-based diagrams. Three separate rendering approaches handle graph, sequence and architecture structures, since each has different layout rules.
Pydantic field and model validators catch malformed LLM output before it reaches the renderer — invalid edge references are dropped rather than failing the whole generation, and enum normalization absorbs schema drift between model versions. Gemini 2.0 Flash is the primary model with GPT as fallback, with exponential backoff on rate limits. Work is autosaved to IndexedDB and persisted to Firestore, and diagrams export as Mermaid (.mmd), draw.io XML or JSON IR. The app ships as a 3-service Docker Compose stack with service health checks.
14 UI-supported diagram types, schema-validated before rendering
Pydantic v2 validators — invalid edges dropped, not failed outright
In production at Yansons Engineering · Sole developer
React
Node.js
Express.js
MySQL
AWS
From spreadsheets to an operational control system.
An internal project and employee-performance platform I designed and built from scratch as the sole developer. The system centralized project planning, task assignment, dependency tracking, submission and review workflows, performance scoring, reporting and management analytics across 8 departments — 40+ REST API routes and 8+ automated operational workflows, deployed on AWS EC2 and Amplify.
or swipe ↔
Overview
Performance Monitoring System
Management needed one system to replace fragmented spreadsheet-based tracking of employees, departments, projects, tasks, dependencies, deadlines, submissions, approvals, performance and reporting. I architected and built the entire platform from scratch as sole developer — React frontend, Node.js/Express backend, MySQL — translating an operational problem into a data model and workflow set with no prewritten technical spec.
Workflows cover employee/task assignment, dependency-based scheduling, submission and review, approval, performance/credit scoring, email notifications, OTP/password reset, reporting and stale-task cleanup — 40+ REST API routes and 8+ automated operational workflows in total. An analytics dashboard gives managers project-level task data, planned-vs-actual Gantt-style timelines, completion metrics, overdue-task and delay-reason analysis, employee contribution and total credit/performance views.
Deployment: the React frontend runs on AWS Amplify, the Node/Express API on AWS EC2, with a MySQL database and a custom domain. Deployed internally across 8 departments, with 80–100 total production tasks tracked in the system over roughly 18 months of development and maintenance.
Solo project · research prototype, not a diagnostic tool
TensorFlow
FastAPI
Angular
scikit-learn
librosa
Three inference pipelines. One multimodal interface.
A multimodal machine-learning prototype that analyzes patterns associated with Parkinson's disease from drawing and speech inputs. Two image-based CNN pipelines handle spiral and wave drawings, while an audio-classification pipeline extracts 22 speech features and aggregates predictions from a 9-model ensemble. FastAPI exposes inference services to Angular interfaces for drawing and audio input.
or swipe ↔
Overview
Parkinson's Classification Prototype
A solo multimodal ML prototype combining two image-based CNN pipelines for spiral and wave drawings with an audio-classification pipeline, built with TensorFlow/Keras, scikit-learn and XGBoost.
The audio path extracts 22 speech features with librosa and aggregates predictions from a 9-model ensemble. FastAPI exposes inference services to Angular interfaces for drawing and audio input, so a user can submit either input type and get a prediction back through the same interface.
This is a research prototype exploring pattern classification from drawing and speech data — it is not a clinical diagnostic system and was not validated for medical use.
Web Developer (Graduate Intern)Purdue University · Center for Career Opportunities
Aug 2026 — May 2027 · West Lafayette, Indiana Current
Developing and maintaining production web systems for Purdue's Center for Career Opportunities as part of a six-developer team, working across frontend, backend, databases, deployment and production debugging.
Web Developer (Graduate Intern)
Purdue University · Center for Career Opportunities
Aug 2026 — May 2027 · West Lafayette, Indiana Current
Situation
The CCO maintains production web applications supporting career-service workflows, event administration and public-facing content. The codebase requires ongoing feature development, debugging, modernization and deployment.
Task
Build and maintain full-stack features, investigate production issues independently and improve reliability across CCO Calendar, Outreach, administrative tools and migrated web content.
Action
Timezone/calendar architecture: reengineered date/time handling across CCO Calendar and Outreach after diagnosing a hard-coded −4 hour UTC offset that only worked during EDT, plus browser-local Date behavior that shifted displayed times by viewer timezone. Built ~12 centralized DST-aware helper functions across ~6–8 files, replacing scattered timezone hacks with consistent handling of PostgreSQL timestamp(6) fields for event creation, editing, querying and calendar display.
Outreach email re-delivery: built 2 REST API endpoints and 2 admin UI resend controls covering 4 email flows/templates, with status-aware routing across 3 event states, multiple facilitator recipients, ICS calendar attachments, validation and confirmation controls to prevent accidental sends.
Authorization/notification bug: diagnosed a production defect where View-only Outreach users could receive Admin-only notifications, traced to a permission lookup using startsWith('/admin/outreach') that matched both the admin and view-only routes. Fixed by adding optional tool-ID support and exact tool-ID matching so Admin notifications resolve specifically to Admin permission.
CMS/accessibility migration: audited and modernized ~20–25 production pages — alt text, heading hierarchy, removing problematic accordions, font/body consistency, broken links, accessible link colors, table structure, keyboard accessibility, missing images and Bootstrap/layout standardization — including replacing legacy "MyCCO" terminology with "Career Gateway".
Work across Next.js, React, TypeScript, Node.js, PostgreSQL and Prisma, deployed via Docker, Kubernetes, GitHub Actions, Rancher and Harbor, with Git feature branches and PR review.
Result
Improved consistency and reliability across calendar, notification and administrative workflows.
Separated Admin-only Outreach notifications from View-only access, closing an authorization/notification-routing defect.
Modernized 20–25 production pages for accessibility and migrated content.
Sole developer of an internal project-performance platform built from a high-level business requirement to replace fragmented spreadsheet-based tracking with centralized workflows, automation and management analytics.
Result
Deployed internally across 8 departments, with 80–100 total production tasks tracked in the system, replacing fragmented spreadsheet workflows with one operational platform.
Software Developer
Yansons Engineering Pvt. Ltd.
Jan 2024 — Jul 2025 · Pune, Maharashtra, India
Situation
Management needed one system to track employees, departments, projects, tasks, dependencies, deadlines, submissions, approvals, performance and reporting — previously spread across spreadsheets.
Task
Translate the operational problem into the architecture, database model, frontend, backend workflows and deployment without a prewritten technical specification.
Action
Architected and built the entire system from scratch as sole developer — React frontend, Node.js/Express backend, MySQL database.
Built 40+ REST API routes and 8+ automated operational workflows covering dependency scheduling, project/employee data synchronization, review/approval flows, performance credit scoring, notifications and reporting.
Designed an analytics dashboard showing planned-vs-actual Gantt-style timelines, completion/delay analysis and employee contribution metrics.
Deployed the React frontend on AWS Amplify and the Node/Express API on AWS EC2, with MySQL and a custom domain.
Result
Deployed the application internally across 8 departments, with 80–100 total production tasks tracked in the system, replacing fragmented spreadsheet workflows with one operational platform.
Worked on transactional-data preprocessing and recommendation-system experimentation using Python and Pandas.
Result
Delivered an evaluated recommendation-system prototype and improved the reliability of the preprocessing/modeling workflow.
Data Scientist Intern
Werardt Systems Pvt. Ltd.
Jul 2023 — Sep 2023 · Pune, Maharashtra, India
Situation
The team needed cleaner transactional data and a working recommendation-system approach to evaluate against the existing workflow.
Task
Preprocess transactional datasets and develop/evaluate recommendation-system experiments under a senior data scientist.
Action
Cleaned and preprocessed datasets using Python and Pandas.
Developed and evaluated recommendation-system experiments, working with both collaborative and content-based approaches.
Implemented validation and evaluation checks for model outputs.
Collaborated with a senior data scientist on debugging and iteration.
Result
Delivered an evaluated recommendation-system prototype and improved the reliability of the preprocessing/modeling workflow.
Python · Pandas · Recommendation Systems · Data Preprocessing · Data Analysis
Academic Intern — Data Analytics using Deep LearningNational University of Singapore (NUS)
Jun 2023 · Singapore
Global Academic Internship Programme in Data Analytics using Deep Learning — 70 hours of coursework and projects across data analytics, machine learning and deep learning.
Result
Completed the program with Grade A and 42/50 on the final project assessment.
Academic Intern — Data Analytics using Deep Learning
National University of Singapore (NUS)
Jun 2023 · Singapore
Situation
NUS's Global Academic Internship Programme offers an international, project-based introduction to data analytics and deep learning.
Task
Complete 70 hours of coursework and project work applying data analytics, machine learning and deep-learning techniques.
Action
Completed coursework and projects spanning data analytics, machine learning and deep learning.
Worked in an international, project-based academic cohort over 70 total hours.
Result
Completed the program with Grade A and 42/50 on the final project assessment.
Python · Deep Learning · Machine Learning · Data Analytics
