Ashmit JaiSarita Gupta

• Led simultaneous migrations of LLM infrastructure (Azure OpenAI → Vertex AI Gemini 2.5 Flash) and file storage (Azure Blob → GCS) across 3 repos - centralised all AI calls into a single llmService.js with per-feature token budgets, native JSON schema enforcement, and SignOz observability tracking prompt tokens, output tokens, and latency per feature. • Shipped 5 complete products end-to-end: Resume Builder, Cover Letter Generator, AI Mock Interview, GoodSpace Accelerate, and a Manage Subscription & Credits system with atomic credit deduction and a master products API - all Razorpay-integrated. • Built Lumi, a streaming AI chatbot powered by Gemini Live API - implemented LLM tool calling for real-time job search, word-queue streaming animation, rolling context compression reducing per-session token cost ~85%, isolated WebSocket management to prevent collision with the interview engine, and a persistent sidebar that survives page navigation via sessionStorage. • Overhauled the payment system from one-time Razorpay orders to recurring subscriptions with webhook-primary activation - idempotency to handle Razorpay retries, PRO → GOLD upgrade via Razorpay subscription update API, atomic credit deduction in DB transactions; built Refer & Earn with Shlink URL shortening, cookie + localStorage dual attribution, and server-side commission calculation. • Drove multiple performance and reliability sprints: eliminated N+1 queries reducing a critical admin API from 10s → 800ms; replaced a TOCTOU-vulnerable Redis lock pattern with atomic SET NX; eliminated 3× cascading duplicate API fetches per page with React Query; and resolved production errors from SignOz traces.

• Implemented Google OAuth 2.0 (Authorisation Code flow) for hirers' end-to-end - JWKS token verification, business email domain validation, a google_oauth_pending holding table for multi-step onboarding, and profile picture upload to GCS; later extended the same architecture to job seekers with a unified callback URL driven by OAuth state params. • Designed and shipped the Organisation/Company Page - a LinkedIn-style company profile with three-tier RBAC enforced at the API middleware layer, one-time invitation tokens with TTL, bottom-up join request flows, diff-based updates to reduce write load, and team management across multiple hirers under one org. • Built a multi-format resume extraction pipeline (PDF via pdf-parse/unpdf, DOCX via mammoth, DOC via antiword, image-based resumes via Google Vision OCR) with optimistic DB locking to prevent concurrent processing, a retry cron for failed extractions, and source tracking foundational infrastructure powering all AI resume features on the platform. • Upgraded the React app from v18 → v19 and optimised the build pipeline, reducing the JS bundle from 55MB → 21MB via route-based code splitting and multi-stage Docker builds; refactored the auth system to use HTTP-only cookie JWTs (XSS-resistant), access + refresh token rotation, and account switching between hirer and job seeker modes without session loss. • Delivered platform-wide SEO infrastructure: dynamic XML sitemaps (jobs, skills, closed jobs), JSON-LD structured data, React Helmet on all pages, SVG → WebP migration for Core Web Vitals (LCP), and skill-specific landing pages for long-tail organic traffic; also built admin revenue analytics dashboards with time-series SQL aggregations, repeat-customer stats, and the AI fit score pipeline for candidate ranking.

The Technical Steering Committee (TSC) is responsible for the oversight of the AsyncAPI Initiative. Maintainers (aka committers) make decisions at the given repository/project level. The TSC helps to make decisions on a higher level, or when maintainers cannot find a consensus.

I maintain the Modelina Website and the AsyncAPI Website handling the issues, PRs and other things related to the Modelina Website.

Developing a UI Kit for the AsyncAPI website using Storybook in Next.js and Typescript.

I am an active contributor to the AsyncAPI Initiative. Some of my major contributions are: • Migrated the AsyncAPI Website from JavaScript to TypeScript and Next.js, improving code maintainability, type safety, and performance. • Redesigned and built the Modelina Playground, an interactive tool enabling visitors to experiment with Modelina directly on the website without local configuration, significantly improving user engagement and ease of use. • Migrated the Modelina website codebase from class-based React components to functional components, enhancing code readability, performance, and alignment with modern React practices.

I handle Industry and Institute Interaction and Database.

Leading the volunteers of Physteo to organise the National Physics Conference 2024 at the National Institute Of Technology Hamirpur and work on a yearly real world project.

Physteo is the Physics Society of the National Institute Of Technology Hamirpur under the Department of Physics and Photonics Science.

Beta Microsoft Learn Student Ambassador Certificate

• Organized and conducted the Microsoft Student Ambassadors Azure AI Challenge in my department, engaging 30 attendees in hands-on learning and practical application of Azure AI technologies.

• Completed the Technical Onboarding Process, gaining proficiency in Git, GitHub, and Static Web Apps, essential tools for modern software development.

Designed a hybrid quantum neural network to classify the MNIST dataset, integrating quantum computing techniques with traditional machine learning methods, achieving 50% test accuracy. The goal of this project was to understand how we can utilize the quantum capabilities to create quantum layers.

Researched and developed an application for Optimising QUBO Relaxation Parameter Via Learning Solver Surrogate. - Created a Pipeline to read and analyze data from the TSPLIB dataset. - Implemented QUBO formulation of TSP using PyQubo. - Used D-wave neal to solve instances from TSPLIB and Study the relation between relaxation parameters, the problem instance, feasibility, and the objective energy. - Explored different ways to select the range of parameters for each instance based on the feasibility of the solutions returned by the solver. - Generating a training dataset for a surrogate model consisting of 300 instances and 30,000 parameter values, taking an approximate compute time of 180 hours to generate. - Currently exploring feature extraction methods for TSP and VRP instances, such as using Graph Convolutional Neural Networks or Reinforcement Learning.