YVRésumé PDF ↓

Lucknow, India

YashmitVaish

Backend engineer building LLM infrastructure and async systems that stay up when real people show up.

See the workPrefer a shell? There's a terminal versiontry it
  • 5,000+Peak concurrent users
  • 30%LLM cost cut
  • 100+Lectures generated
  • 9.28CGPA

About

Hi, I'm Yashmit.

Tap the card.

I build backend systems and LLM infrastructure: RAG pipelines, async job queues, and the deployment work that keeps them up when real people show up.

Computer Engineering at Thapar Institute of Engineering and Technology, class of 2028, with a CGPA of 9.28.

Right now I'm the sole engineer on the AI architecture at Clyra, and General Secretary of Thapar's ACM Student Chapter, where I also end up running the deployments.

On the side I'm building DigiTwin, a local-first AI assistant with long-term memory, and writing Go for a Redis-style server and a peer-to-peer compute experiment.

Toolbox

What I build with

Languages

PythonGoJavaScriptC++C

LLM engineering

RAGGraphRAGLangChainLangGraphCrewAIvLLMOllamallama.cppGGUF quantization

Backend

FastAPIDjangoGinRedisCeleryMicroservices

Data

PostgreSQLNeo4jElasticsearchMongoDBSQLiteMySQLSupabasepgvectorPinecone

Infra

AWSAzureDockerNginxCloudflareCI/CD

Tools

GitGitHubVercelLocust

Dec 2025 – presentRemote

Sole engineer on the AI architecture of an AI-powered content platform.

  • Designed and shipped production RAG pipelines on LangGraph, Elasticsearch and Neo4j, with async job processing, retries and error handling. Load-tested with Locust to sustain 30–40 content files a day.
  • Cut LLM inference cost by 30% and redundant re-processing by 70–80% with model routing and response caching: simple tasks go to cheaper models, and repeated retrieval calls are cached.
  • Raised RAG output quality by 30% (ROUGE) by redesigning chunking, tuning retrieval across Elasticsearch and Neo4j, and reworking agent orchestration. Validated on a held-out content benchmark.
LangGraphElasticsearchNeo4jLocustPython

2026Patiala

Built and deployed the Campus Permission Portal, Thapar's night permission system, in a two-person team.

  • Gathered requirements from faculty and administration, scoped multi-role approval workflows, then built and deployed the platform. Observed a peak of 5,000+ concurrent users in usage logs, with zero downtime.
  • Async FastAPI and a Redis job queue behind an Nginx reverse proxy absorb bursts in the portal's daily 4 to 5 hour active window, at roughly 3,000 page views a day.
  • Transactional email with direct status links cut follow-up questions to faculty from 50+ to 1–2 a week.
  • Unit tests cover the core approval logic. Deployed on AWS through the admission-cycle traffic peak.
FastAPIReactPostgreSQLRedisNginxAWS

Aug 2026 – presentThapar Institute

Running a 100+ member chapter, and the infrastructure underneath it.

  • Lead 100+ members across coordinated sub-teams. Planned and ran 4 chapter-wide events, including a freshman orientation, a hackathon and several technical workshops.
  • Own event operations end to end: planning, logistics, team coordination and on-ground execution.
  • Ran deployment for the chapter's recruitment test in Oct 2026 on AWS EC2, Azure VMs and S3: 1,000 students, 500 concurrent at peak, 570 ms p95 and no incidents. Deployment was my part; I did not build the application.
  • Manage the chapter's Cloudflare account for DNS and email routing, and handle its deployments on AWS EC2, Azure VMs and S3.
CloudflareAWS EC2S3Azure VMNginxDocker

Selected work

Four things worth opening

Each one has a full write-up: what it does, how it's built, and the numbers.

01 / 2026 · Live

Campus Permission Portal

Thapar's night permission system

Multi-role approval workflows for an entire campus, built by a two-person team and live in production.

FastAPIReactPostgreSQLRedis
Read the write-up

02 / 2025 · System

Pdf2Video

Academic PDFs into narrated, animated lectures

A fully offline multi-agent pipeline that has turned academic PDFs into 100+ animated lectures with Manim. No GPU rental, no cloud bill.

🏆 Became EDUSCENE: Innovation Award, Ulster University

CrewAILangChainManimPiper
Read the write-up

03 / 2026 · In progress

DigiTwin

A personal AI assistant with long-term memory

An assistant that remembers your goals, preferences and life, gives you a daily brief and keeps you on track. It runs on a local model by default and only reaches for the cloud when you let it.

FastAPIOllamaChromaDBNeo4j
Read the write-up

Also built

Contest Portal Backend

Whole backend for the FCC contest portal, solo. Go and Gin on Supabase, served 500 people.

Go · Gin · Supabase · LocustDredis

A Redis-style server written in Go. Handles a basic command set, early stage.

Go · ExperimentP2PCS

Peer-to-peer compute sharing: run your code on someone else's machine. Signalling service plus executor.

Go · In progress

The rest of it

Education, awards, what's running

Education

  • B.E. Computer EngineeringThapar Institute of Engineering & TechnologyCGPA 9.28 · 2024 – 2028

Awards

  • Innovation AwardAgentic AI Hackathon 2025, Ulster University (UK)EDUSCENE · Team .gitignore

Running right now

  • LiveCampus Permission Portal5,000+ peak concurrent, zero downtime
  • LiveClyra RAG platform30–40 content files a day
  • DeliveredACM test portal500 concurrent, 570 ms p95
  • BuildingDigiTwinLocal-first AI with long-term memory

Let's buildsomething.

Open to backend and LLM infrastructure roles. Send me the problem and I'll reply with how I'd approach it.

contact@yashmitvaish.me
GitHub LinkedIn LeetCode Résumé

Rather read this from a shell?

The same portfolio, as a terminal. Type help, poke around, run status to see what's live.

There may be a few commands that aren't in the help text.

Open the terminal

Works best with a keyboard.