Shivam Singh

Product / Full-Stack Software Engineer building Calwyn, an AI job-hunt workspace I developed and operate end to end. My work covers product decisions, AI pipelines, job orchestration, multi-tenant Postgres, billing and metering, testing, and production operations. Its predecessor, HireRank, reached 100+ daily active users and paying subscribers across eleven solo production releases. My path into software was less conventional. Before writing production code, I spent two years running retail operations, managing a 10-person team and handling daily excise compliance with state regulators. I later worked at Tata Technologies on Python-based ML for manufacturing safety data, then joined Veritas, an AI fellowship led by Harvard graduates, where I trained neural networks on a meteorology dataset.

Experience

  • Built and shipped HireRank through eleven production releases, reaching 100+ daily active users and paying subscribers.
  • Rebuilt and relaunched the product as Calwyn after a naming conflict, preserving the core idea while expanding its scope.
  • Built the application end to end, including a multi-provider AI pipeline with automatic fallbacks, durable background jobs with real-time updates, secure multi-tenant isolation, usage metering, billing, and GDPR tooling.
  • Expanded Calwyn beyond résumé scoring into application tailoring and Fit scoring, job tracking, interview preparation, outreach drafting, salary negotiation, public profiles, and an OAuth-protected MCP server that lets AI agents handle the job hunt user's behalf.
  • As the sole developer, I also take care of the less visible parts of running Calwyn: CI, database migrations, webhook reliability, automated testing, monitoring, billing operations, GDPR requests, and customer support.
  • Worked through manufacturing safety data: incident reports, PPE compliance records, near-miss logs. Found where and when violations clustered.
  • Built Python tooling to clean, analyze, and visualize the data using NumPy, TensorFlow, and Matplotlib.
  • Presented the findings to operations leadership, with several recommendations incorporated into the following safety protocol update.
  • Managed day-to-day operations for a 10-person retail team, including scheduling, training, and performance reviews.
  • Handled daily excise compliance, licensing documentation, and reporting for a heavily regulated spirits retail business processing thousands of customer transactions each month.
  • Worked with the Department of Excise and Prohibition to improve licensing and reporting workflows, reducing paperwork and speeding up end-of-day close.

Projects

Visit
  • Calwyn is built around the full job hunt, not just résumé tailoring. It keeps applications, contacts, outreach, interviews, offers, reminders, and career information together in one workspace.
  • For résumé tailoring, I built the analysis as a background workflow that scores fit against the role, checks that suggested changes are backed by the candidate’s actual experience, and can switch AI providers if one goes down.
  • Added an OAuth-protected MCP server so Claude, ChatGPT, and other agents can work with the same career data, update applications and contacts, prepare materials, and pick up where a previous session left off.
Next.jsTypeScriptPostgresPrismaInngestAI
CodeVisit
  • Turns recorded lectures into searchable notes students can come back to later. Audio is transcribed with AssemblyAI, then Gemini generates summaries, titles, follow-up questions, and embeddings for search.
  • Built semantic search over that history with Gemini embeddings and MongoDB Atlas Vector Search, so a student can look for an idea without remembering the exact words used in the lecture.
  • Supports signed-in, guest, and shared use. Signed-in users keep their lecture history, guests can use the app without an account, and any analysis can be shared through a public link. Guest data is cleared after 24 hours.
Next.jsTypeScriptMongoDBAI
Code
  • Used 145K+ Australian weather observations to predict whether it would rain the following day. Compared logistic regression with a TensorFlow neural network, with the simpler logistic model reaching 85% testing accuracy.
  • Cleaned and prepared the data in Python, including missing-value handling, categorical encoding, feature scaling, and experiments with under- and oversampling to deal with the imbalance between rainy and non-rainy days.
  • Built the project during the Veritas AI fellowship under the supervision of three technical mentors, where we used the model comparisons to understand the tradeoff between overall accuracy and performance on the minority rain class.
PythonTensorFlowPandasScikit-learn

Skills

Core

TypeScriptJavaScriptReactNext.jsNode.jsPostgreSQLPrismaPythonSQL

Backend & Systems

MongoDBInngestRedisQStashOAuthAuth.jsTypstwebhooksMCPCI/CD

AI & Tools

Vercel AI SDKAssemblyAIElevenLabsGenerative AIClaude CodeCodexBrowser UseGit

Testing & Infrastructure

VitestPlaywrightVercelCloudflare R2NeonPostHogBetterStackDocker

Frontend & UI

Tailwind CSSshadcn/uiTiptapFramer MotionGSAPLucideThree.js

Education

SRM Institute of Science and Technology, Kattankulathur, TN · Bachelor of Computer Applications (BCA), Data Science

May 2025

St. Xavier’s School, Hazaribagh, JH · Higher Secondary (CBSE), Mathematics with Computer Science

May 2021