I build backend systems that hold up in production, land fixes in the open-source frameworks other engineers run on, and compete at CodeChef 5★.
Scalable systems, engineered from first principles.
From BIT Mesra to UKG, Yuvraj bridges Electronics and Software Engineering with the instinct of a systems thinker. The foundation is precise: CGPA 7.59, backed by an ECE degree, hardened through algorithms, backend architecture, and product-grade execution.
Systems stack, exposed.
Languages, backend frameworks, cloud and DevOps tooling, AI engineering and competitive programming. Hover any node to see where it was used.
C++
Competitive ProgrammingMain contest language: CodeChef max rating 2034 (5★) and LeetCode max rating 2042 (Knight).
Enterprise systems with measurable release signals.
UKG
Automated API failure detection and Apigee delivery on GCP
- Built a Python monitoring pipeline over Nginx logs, Splunk exports and Grafana metrics that flags sustained 5xx failures, with a separate threshold for each API scaled to its call volume from Apigee Analytics.
- Routed flagged failures to Claude Code through a Jira MCP integration that files tickets and skips duplicates: 62 defects filed in 4 months, about 65–70% confirmed as real by senior-engineer review.
- Maintained the Jenkins CI/CD pipelines that deploy Apigee API proxies to dev, staging and production on GCP, debugging failed builds and supporting zero-downtime revision releases through the Apigee Management API.
ECE SIPAR
Medical image analysis: drug efficacy and dental X-ray segmentation
- Measured wound healing for a RIMS Ranchi drug-efficacy study on diabetic rats (experimental drug vs. standard drug vs. natural healing): annotated wound boundaries as polygons, computed their areas with the Shoelace formula, and plotted percent healed over 13 days. By Day 13 the experimental drug group was 98.7% healed, against 83.9% for the standard drug and 56.2% for natural healing.
- Trained a custom U-Net to segment third molars (wisdom teeth) on panoramic dental X-rays, using 800–900 images annotated by my professor, and reached about 72% IoU after a warm-up phase on circular objects from a COCO-derived dataset.
Upstream fixes, merged and shipping.
9 pull requests merged into 4 widely used projects. Each one started from a real bug or gap: reproduce it, fix it, cover it with tests, and get it through maintainer review.

NestJS
nestjs/nestProgressive Node.js framework for building server-side applications.

Puppeteer
puppeteer/puppeteerJavaScript API for controlling Chrome and Firefox.

Stirling-PDF
Stirling-Tools/Stirling-PDFSelf-hosted toolkit for editing, converting and securing PDFs.

Uptime Kuma
louislam/uptime-kumaSelf-hosted uptime monitoring with alerts to 90+ notification services.
fix(smsmanager)#8174Fixed SMSManager alerts ignoring the configured gateway type.
Merged Nov 16, 2026+29−1fix(signl4)#8173Fixed SIGNL4 alerts using a nonexistent monitor ID; they now send the real monitor ID as the alert's external ID.
Merged Nov 16, 2026+43−2test(ping)#8148Fixed the IPv6 bracket ping test so it passes on machines that have IPv6.
Merged Nov 12, 2026+4−2
Product systems, built to be inspected.
AI Mock Interview
A GenAI interview simulator: Gemini writes role-specific questions, you answer out loud, and every answer comes back with a rating and written feedback.
Practising interviews alone gives no feedback. The app fixes that, but its first version called the database and Gemini straight from the browser, which exposed both secrets in client JavaScript, and its interview flow had bugs that cut answers short and saved broken questions.
- Server actions with ownership checks
- Gemini JSON mode + response schema
- Auto-restarting speech capture
- CI: lint, test, build on every PR
- Automated GitHub Releases
- v1.0.0 release
Architecture
- Next.js 15 App Router with Clerk authentication; all data access runs in server actions that check the signed-in user owns the interview.
- PostgreSQL through Drizzle ORM; deleting an interview removes its answers in the same transaction.
- Gemini in JSON mode with a response schema for both question generation and answer ratings (whole numbers from 1 to 5); questions are validated before they are saved.
- Browser speech recognition restarts automatically when the browser pauses it, so long answers are kept in full; re-recording replaces the previous answer.
Engineering Decisions
- Moved secrets server-side (DATABASE_URL, GEMINI_API_KEY) and rotated the keys the old client bundle had exposed.
- Asked Gemini for schema-checked JSON instead of parsing free text, so a malformed response can no longer create an unusable interview.
- Made the question count configurable (INTERVIEW_QUESTION_COUNT, 1–10) after finding it came from an undefined variable.
- Added unit tests, ESLint, a CI workflow, Dependabot and tag-driven GitHub Releases, then shipped the work as v1.0.0 with a changelog.