Skip to content
Pratyush Vishal
Resume (PDF)

Software engineer · AI researcher

Pratyush VishalSystems that survive the crash.Models that survive the noise.

Based in
Bangalore, India
Studying
B.Tech ECS, VIT Vellore ’28
Focus
Backend, infrastructure, ML
Previously
SDE intern, Avior AI

I build backend and infrastructure software in Go, Python and TypeScript, and I’m first author of a deep-learning paper in Scientific Reports. Currently studying Electrical and Computer Science Engineering at VIT Vellore.

Explore projectsDownload resume (PDF)
01Selected work

Where the guarantees live.

Three projects, chosen for the engineering underneath them: fairness and durability in an ingestion path, service boundaries in a multi-role platform, and the plumbing of a secure REST API.

01Fair-share event ingestion gateway

FairGate

Open source · MIT · beta

A TCP ingestion gateway in Go that keeps one noisy producer from starving the rest, and makes every accepted event durable before acknowledging it.

Problem

When many producers share one ingestion path, a single high-rate producer can take over buffers and bandwidth, and anything held only in memory is lost if the process crashes.

InteractiveTake FairGate apart

● Pipeline explorer

Select a stage

internal/sched

DRR scheduler

The scheduler visits queues in rounds. Each backlogged producer earns quantum × weight credit and is served one event per unit of credit. A producer with an empty queue has its credit reset, so it can’t bank credit and burst ahead later.

Under contention, shares follow weights.

Round 000

  • A
    Share 0%
  • B
    Share 0%
  • C
    Share 0%

Press play or step to run the scheduler.

Illustrative · quantum 2 · capacity 12 · share over last 8 rounds

Stage details from the FairGate repository docs

Stack

  • Go
  • ClickHouse
  • Prometheus
  • Grafana
  • Docker Compose

Key decisions

  1. D1

    Group commit instead of one fsync per write

    Concurrent writers are batched into a shared fsync, so durability is paid for once per batch rather than once per event.

  2. D2

    Torn records are only tolerated at the tail

    Recovery replays segments in order and rejects a torn record anywhere except the final segment, the only place an interrupted write can legitimately appear.

  3. D3

    Shipping is decoupled from scheduling

    The shipper reads durable records straight from the WAL, checkpoints what it has delivered, and reclaims segments once they are fully shipped.

02Research management platform

Luno

Staging on AWS EC2

A full-stack platform for running academic research projects, with separate dashboards and permissions for admins, faculty and students.

Problem

Admins, faculty and students work on the same projects, files and references, but each role needs different access and a different view of the work.

Approach

  • Next.js and React frontend over FastAPI microservices.
  • Covers auth, project management, file uploads, references, notifications and role-based dashboards.
  • API Gateway in front; gRPC and Kafka for inter-service messaging.
  • Redis caching over PostgreSQL storage.
  • Docker Compose, Nginx with HTTPS, and JWT/RBAC, staged on AWS EC2.

● System overview

Simplified

Next.js

React · role dashboards

Nginx

HTTPS

API Gateway

FastAPI services

  • auth
  • projects
  • files
  • references
  • notifications
  • dashboards

gRPC + Kafkainter-service messaging

Redis

cache

PostgreSQL

storage

Docker ComposeJWT · RBACAWS EC2 · staging

  1. D1

    One gateway, many services

    Clients talk to a single API Gateway, while the services behind it exchange messages over gRPC and Kafka.

  2. D2

    Roles as a first-class concept

    JWT authentication with role-based access control drives separate dashboards for admin, faculty and student users.

  3. D3

    Analytics by role

    Usage tracking across the platform records access patterns per role, not just raw page views.

  • Next.js
  • React
  • FastAPI
  • gRPC
  • Kafka
  • Redis
  • PostgreSQL
  • Nginx
  • Docker Compose
  • AWS EC2

03Node.js / Express REST API

Patient Management System

A REST API for clinic operations: authentication, appointments, doctor and patient management, pharmacy and invoicing.

Clinic operations span sign-in, scheduling, patient and doctor records, pharmacy and billing, and all of it needs to sit behind one secure API.

Approach

  • Twilio-based OTP verification and Nodemailer for email.
  • PostgreSQL and MongoDB side by side through Sequelize and Mongoose.
  • Redis plus local in-memory caching.
  • JWT, bcrypt, Helmet, CORS, rate limiting and compression.
  • Docker Compose, with Kubernetes and Azure deployment manifests.
~/patient-managementoverview
$ describe service
runtime   node · expressdomains   [auth, appointments, doctors, patients, pharmacy, invoicing]stores    [postgresql (sequelize), mongodb (mongoose)]cache     [redis, in-memory]security  [jwt, bcrypt, helmet, cors, rate-limit]messaging [twilio otp, nodemailer]deploy    [docker compose, kubernetes manifests, azure manifests]
Project summary · not a repository file
02Research & publications

Reading machine health from vibration.

In plain terms

Rotating machinery usually fails at its bearings, and the vibration signal carries the warning. The paper compares deep-learning approaches for telling bearing faults apart when the machine’s load and speed keep changing, then asks how the best models behave under heavy noise and what they would cost to deploy.

  • Published during the second year of undergraduate study.
  • Cited twice within six months of publication.

Contributions

  1. 01Built and evaluated a deep-learning pipeline for a 5-class, 45-condition bearing-fault test rig, processing more than 18.9 million raw vibration data points.
  2. 02Designed a hybrid transfer-learning architecture: frozen EfficientNetB0, MobileNetV2 and InceptionV3 backbones on CQ-NSGT spectrograms, feeding custom trainable residual blocks.
  3. 03Benchmarked CNN and Vision Transformer approaches on training time, GFLOPS, latency and memory footprint for deployment trade-offs.

● Method

Hybrid transfer learning

  1. 01Raw vibrationTest rig · variable load and speed
  2. 02CQ-NSGTTime–frequency spectrograms
  3. 03Frozen backbone
    • EfficientNetB0
    • MobileNetV2
    • InceptionV3
  4. 04Residual blocksCustom, trainabletrainable
  5. 05Fault class5 classes
Why frozen? Partial fine-tuning of the backbones raised training cost by 30–100% with no accuracy gain.
Five-fold CV accuracy
99.83%
Best of the five architectures evaluated
Accuracy at 1 dB SNR
98.8%
Under noise stress testing
Friedman test
p = 0.0005
χ²(4) = 20.00; gains statistically significant
03Experience

On a product team.

  1. to

    6 months

    Avior AI

    Software Development Intern

    • PostgreSQL
    • API design
    • React
    • Refactored a proof-of-concept microservice’s PostgreSQL schema and API design, cutting request latency by 2–2.5×.
    • Modularised the React frontend and removed unnecessary re-renders to improve application performance.
    • Built secure file storage and authentication for a production service, and redesigned the platform’s landing page.
04Toolkit & education

What I work with.

Languages
Python, JavaScript, TypeScript, Go, Javabasic proficiency, Kotlinbasic proficiency
AI / ML
PyTorch, TensorFlow, Scikit-Learn, NumPy, Pandas
Backend
FastAPI, Express.js, Flask, gRPC, Kafka, Redis
Databases
PostgreSQL, MongoDB, MySQL, ClickHouse
Frontend
React, Next.js, React Native, Expo
Cloud / DevOps
Docker, AWS

basic = basic proficiency

Also used in project work: Prometheus, Grafana, Nginx, Kubernetes manifests, Azure manifests, Twilio

Education

Jul 2024 — Jul 2028

Vellore Institute of Technology, Vellore

B.Tech in Electrical and Computer Science Engineering

CGPA8.49

05About

What connects the work.

Most of what I build sits underneath an interface: ingestion paths, storage formats, service boundaries. I’m drawn to the questions a system has to answer under stress: what happens when a producer floods the queue, when a disk write is cut off halfway, or when one service slows down for everyone else.

My research comes from the same instinct, applied to models. In the bearing-fault work, the interesting questions weren’t only which architecture scored highest, but whether that score held up under noise, whether fine-tuning was worth what it cost, and what each model would demand once deployed.

Off the clock
CricketBadmintonPhotography
Based in
Bangalore, Karnataka, India

06Contact

Have a hard problem in systems or ML? Let’s talk.

pratyush.vishal15@gmail.com
Download resume (PDF)