<ARJITH/>

Arjith A R

Software EngineerMachine Learning EngineerBackend Systems

> Building neural networks from scratch.

[cf]nvimzsh
~/codeforces/1000/1342A-RoadToZero🢖
🢖python3 submit.py
Compiling... Running Test Case 1... ... Accepted Codeforces #1342A

"I don't want to just use machine learning.I want to understand every equation well enough to build it from scratch."

Software Engineering

  • Backend Architecture
  • APIs
  • Databases
  • Docker
  • CI/CD

Machine Learning

  • PyTorch
  • Neural Networks
  • Diffusion Models
  • Transformers
  • Computer Vision

Systems

  • Linux
  • Networking
  • Observability
  • Cloud
  • Containers

Experience

My professional journey.

Full Stack Developer Intern

Dom Ventas • Bangalore
Oct 2025 — May 2026
▹Built and deployed a greenfield AI agent for automating the listing process using PydanticAI and Gemini APIs.
The Challenge
Manual generation of property listing text was highly tedious, time-consuming, and expensive.
The Solution
Streamlined the text generation pipeline with production-grade LLM orchestration and observability via Logfire.
Quantified Result
Reduced listing creation time from 1 hour to 15 minutes, cutting costs from approx ₹150 to ₹26 per listing.
▹Built REST APIs and microservices for a property analysis platform serving international clients using Python and FastAPI.
The Challenge
Needed scalable asynchronous backend services to handle complex property data processing.
The Solution
Provided a robust architectural foundation for the platform's core logic and client-facing features.
Quantified Result
Successfully served global client traffic with non-blocking async execution.
▹Designed an ECR-based CI/CD pipeline with GitHub Actions and automated ECS deployments.
The Challenge
Manual deployments were error-prone and caused service interruptions.
The Solution
Enabled seamless container image builds and deployments with automated health checks.
Quantified Result
Achieved zero-downtime releases and accelerated delivery cycles.
▹Implemented Git commit-triggered code quality checks using Ruff and basedpyright.
The Challenge
Codebase lacked consistent formatting and strict type safety across the team.
The Solution
Enforced linting, formatting, and type checking on every commit automatically.
Quantified Result
Eliminated runtime type errors and dramatically improved code review efficiency.
▹Patched a data duplication bug in PostgreSQL by fixing an improper SQLAlchemy join.
The Challenge
Users were experiencing inaccurate duplicated data records on the platform.
The Solution
Restored data integrity and accuracy for the end users.
Quantified Result
Resolved 100% of data duplication reports related to the core query.
▹Diagnosed and resolved a multi-layered frontend caching bug using Trie-based caching and TanStack Query.
The Challenge
Stale data was persisting unexpectedly on the client side, causing a poor user experience.
The Solution
Implemented a robust caching invalidation strategy that properly synchronized state.
Quantified Result
Resolved a highly complex 2+ week blocking issue within 3 days.
▹Built server-rendered pages using React/Next.js SSR alongside the FastAPI backend.
The Challenge
Global customers on low-speed networks experienced severe layout shifts and slow initial loads.
The Solution
Delivered fully rendered HTML directly to the client, greatly improving perceived performance.
Quantified Result
Significantly improved First Contentful Paint (FCP) and overall load times on slow networks.

Backend Developer

Paradox • IIT Madras
Aug 2024 — Sep 2025
▹Designed REST API endpoints and optimized ORM queries using Python Django.
The Challenge
High-traffic routes were suffering from database bottleneck issues.
The Solution
Greatly reduced database load and improved overall endpoint response times.
Quantified Result
Optimized queries on critical paths ensuring stable high-traffic performance.
▹Fixed a severe N+1 query issue in the Django ORM involving missing FK prefetches.
The Challenge
The system was making excessive sequential database calls for related objects.
The Solution
Streamlined database access by fetching related objects efficiently in a single query.
Quantified Result
Reduced database requests from 150 down to just 2 per request.
▹Acted as first responder to diagnose and mitigate a production DDoS attack.
The Challenge
The platform was overwhelmed by malicious traffic disrupting legitimate users.
The Solution
Analyzed Nginx logs and journalctl to identify the attack vector and secure the infrastructure.
Quantified Result
Restored service availability quickly by applying UFW rules and Cloudflare reverse proxy.
▹Automated EC2 provisioning with Docker and auto-scaling cluster configurations using Bash.
The Challenge
Manual server provisioning was slow and inconsistent during traffic spikes.
The Solution
Created reproducible infrastructure deployments with all dependencies pre-configured.
Quantified Result
Enabled seamless auto-scaling to handle variable load efficiently.
▹Containerized local development environments using Docker.
The Challenge
New developers faced setup issues which slowed down velocity.
The Solution
Standardized the local stack, ensuring environment parity with production.
Quantified Result
Reduced developer onboarding time from several days to just a few hours.
▹Introduced comprehensive unit and integration tests across the Django codebase.
The Challenge
The project lacked automated testing, making refactoring highly risky.
The Solution
Established a culture of testing and safeguarded critical business logic against regressions.
Quantified Result
Drove test coverage from 0% to 33%.

Projects

theML-AIClan

Building neural networks, autograd engines, and optimizers entirely from scratch in PyTorch.

PythonPyTorchNumPyMathematics

Implemented

  • Linear Regression
  • Logistic Regression
  • Neural Network
  • Momentum Optimizer
  • Manual Gradient Descent
  • Manual Autograd
  • Training Visualization

Roadmap

  • Transformers
  • Diffusion
  • GAN
  • CUDA
  • Triton

Current Focus

What I'm learning, building, and debugging.

Manual Autograd Engine

Building a custom automatic differentiation engine and computational graph from scratch.

Key Learnings
  • ▹Computational graphs and topological sorting.
  • ▹Chain rule application across dynamic graphs.

Linear Regression

First ever model I built from scratch using PyTorch.

Key Learnings
  • ▹How gradients and learning rates mathematically converge.
  • ▹Fundamental tensor manipulation and PyTorch operations.

Logistic Regression

Learned the critical importance of loss functions. Compared MSE vs BCE for classification tasks.

Key Learnings
  • ▹Difference in gradient shapes between MSE and BCE.
  • ▹How log likelihoods define classification bounds.
Bug Faced
Loss and gradients exploding to NaN.
Root Cause
The BCE log term `log(y - y^)` approached 0, resulting in `log(0) = -inf`.
The Fix
Clamped the predictions using a small epsilon `1e-7` to prevent perfect 0s and 1s inside the log.

Custom Optimizers

Implementing classic gradient descent optimization algorithms completely from scratch.

Key Learnings
  • ▹How momentum helps escape local minima.
  • ▹Adaptive learning rates via AdaGrad and RMSProp.
Sub-components Breakdown
Gradient Descent
Momentum
AdaGrad
AdaDelta
RMSProp
Adam
Nesterov
* Math intuition missing. Parked for now.

XOR Net

Taught me the fundamentals of building a basic multi-layer Neural Network.

Key Learnings
  • ▹Forward and backward passes through multiple layers.
  • ▹Activation function (ReLU, Sigmoid) derivatives.
Bug Faced
Model failed to classify properly (Underfitting).
Root Cause
Used too few hidden neurons, mathematically limiting the network from learning the non-linear XOR boundary.
The Fix
Increased hidden layer capacity and debugged intermediate layer outputs.

Tech Stack

Languages

PythonTypeScriptSQL

Frontend

Next.jsReactTailwind CSS

Backend

FastAPIFlaskDjangoExpress

Infrastructure

DockerGitHub ActionsAzureGCPAWSPostgreSQL

Observability

AlloyLogFireOpenTelemetry

Machine Learning

PyTorchNumPyPandasTransformers

GitHub Activity

Repositories
41
Stars Earned
1
Followers
3

Recent Projects

Codeforces Profile

Competitive programming stats.

Current Rating
890
Max Rating
998
Current Rank
newbie
Max Rank
newbie

Let's build something interesting.

Currently open to new opportunities, collaborations, and discussions about machine learning, systems engineering, or just cool tech.