Skip to content
View akash335's full-sized avatar
👋
Welcome To My World
👋
Welcome To My World

Block or report akash335

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
akash335/README.md

Porumamilla Akash

Software Engineer · Full-Stack & GenAI · Final-year CS undergrad at IIT Bhilai (Class of 2027)

I build full-stack software and GenAI applications end to end: REST backends in FastAPI and Express, React and Next.js frontends, PostgreSQL schemas designed around real access patterns, and retrieval and agent systems with LangGraph and RAG, deployed with CI. Currently researching backdoor attacks on neural code models for my final-year thesis.

📍 Vijayawada, Andhra Pradesh, India  •  📧 porumamillaakash@gmail.com  •  🌐 Portfolio


Projects

AI Knowledge Assistant — Production RAG System

End-to-end RAG application for multi-document PDF question answering with grounded responses and source/page attribution. A LangGraph workflow rewrites conversational follow-ups, retrieves chunks via MMR-based FAISS search, and generates answers with Groq GPT-OSS 120B. FastAPI backend, Streamlit frontend, deployed separately.

Python LangGraph FAISS FastEmbed Groq FastAPI Streamlit

Support Triage Agent — Retrieval-Grounded LLM Agent

Classifies support messages into 10 intents, retrieves historical resolutions, drafts grounded replies, and routes auto-handle vs. escalation. An escalate-only guardrail means a prompt injection in a customer message cannot force auto-handling. Evaluated on 200 stratified examples against two baselines with bootstrap confidence intervals and LLM-as-judge: auto-handles 74.5% of traffic at 0.91 precision.

Python GPT-OSS 120B scikit-learn TF-IDF retrieval LLM-as-judge

CKD Guardian — Clinical Risk Prediction Platform

60-endpoint FastAPI backend (~6,100 lines of Python) over a 17-table PostgreSQL schema with 20 foreign keys, 6 uniqueness constraints and 14 composite indexes tuned to real access patterns. Gradient Boosting model stages kidney disease risk into four KDIGO-aligned levels. Includes WebSocket doctor–patient chat, push notifications, and a PDF/OCR pipeline turning unstructured lab reports into structured records. Shipped to Android and iOS via Capacitor.

FastAPI PostgreSQL SQLAlchemy Alembic scikit-learn Next.js Capacitor

Backdoor Attacks on Neural Code Models — B.Tech Thesis (in progress)

Replication of AFRAIDOOR (IEEE TSE 2024) at IIT Bhilai. Reproduced fixed, grammar-sampled and adaptive gradient-based trigger attacks on an LSTM sequence-to-sequence model trained on CodeSearchNet-Python; fixed and grammar triggers reached 98.1% and 96.5% attack success at 10% poisoning. Fixed 9 blocking defects in the upstream research code and cut trigger generation from 140 to 5.4 seconds per batch (26× faster).

Python PyTorch seq2seq tree-sitter adversarial ML

Shipment Analytics Pipeline — Batch Data Pipeline

Moves 500,000 synthetic shipment records from PostgreSQL into partitioned Parquet with PySpark, using broadcast joins and window functions. Warehouse layer modelled in dbt (6 models, 11 schema tests), orchestrated as a 3-task Airflow DAG, validated by GitHub Actions CI on every push.

PySpark Airflow dbt Parquet PostgreSQL GitHub Actions

Team Task Manager — REST Backend

Multi-user task platform in Express and PostgreSQL. 8 endpoints covering registration, login and full task CRUD with status transitions. JWT auth with bcrypt hashing and route-level middleware enforcing per-user ownership. Deployed on Render against managed PostgreSQL. (frontend)

Express.js PostgreSQL JWT bcrypt React

AKIR Restaurant — Full-Stack Web Application

Nine-page restaurant site in React and TypeScript with shadcn/ui and a Three.js 3D hero. Express API handles reservations and contact submissions, delivering transactional email through the Brevo API with retry on transient failures, behind a per-IP sliding-window rate limiter. (backend)

React TypeScript Vite Tailwind CSS Three.js Express.js


Tech

Languages   Python · Java · C++ · JavaScript · TypeScript · SQL · C

Backend   FastAPI · Express.js · Node.js · REST APIs · WebSockets · JWT · SQLAlchemy · Alembic

Frontend   React · Next.js · Vite · Tailwind CSS · shadcn/ui · Streamlit

GenAI & Agents   LangGraph · RAG · FAISS · FastEmbed · prompt engineering · structured outputs · guardrails · LLM evaluation · OpenAI and Groq APIs

ML   PyTorch · scikit-learn · pandas · NumPy

Data   PostgreSQL · schema design · indexing · migrations · PySpark · Airflow · dbt

Cloud & DevOps   Docker · GitHub Actions CI · Git · Linux · Postman · Vercel · Render


Elsewhere

Portfolio  •  LinkedIn  •  LeetCode

Open to full-time Software Engineering and AI Engineering roles, and six-month internships from January 2027.

Pinned Loading

  1. AI-Knowledge-Assistant AI-Knowledge-Assistant Public

    AI-powered RAG assistant for querying PDF documents using LangGraph, FAISS, FastEmbed, and Groq.

    Python 1

  2. hiver-support-agent hiver-support-agent Public

    Python 1

  3. CKD_app CKD_app Public

    CKD detection and monitoring platform with ML predictions and with AI Guidance, alerts, telemedicine, Android, and iOS apps.

    TypeScript

  4. Team-task-manager-Frontend Team-task-manager-Frontend Public

    JavaScript 1

  5. Web_Restaurant_frontend Web_Restaurant_frontend Public

    Full-stack restaurant website (React + Vite + TypeScript, Tailwind/shadcn) with reservations & contact forms, Brevo API, deployed on Vercel/Render.

    TypeScript

  6. Humanness-android-intern Humanness-android-intern Public

    Humanness Sample Task

    Kotlin