Skip to content

Building agentic AI systems in production

Arun Karthik BT — AI Engineer | LLM Agents, RAG & Voice AI

// ai engineer @ 10xScale.ai

I build production LLM systems — multi-agent backends, RAG, and real-time voice agents — and own the path from prompt and model choice to the FastAPI, PostgreSQL and task-queue infrastructure that serves them.

View work ↓Resume / LinkedIn ↗Get in touch
LANGGRAPH, MCP, RAG, GEMINI, FASTAPI, LIVEKIT, POSTGRESQL, QDRANT
Arun Karthik BT — AI Engineer portrait

[01] — About

Agents that work outside the demo.

Agents that ship. Retrieval that's right. Inference that doesn't burn the budget.

I'm an AI Software Engineer at 10xScale.ai in Hyderabad, building agentic backends with LangGraph, AgentFlow and MCP/FastMCP — and the extraction, retrieval and voice pipelines around them.

My bias is toward what works in production: reliable, measurable systems over impressive demos. I reach for deterministic logic wherever accuracy matters, use models where they genuinely add value, and keep latency and cost in check as things scale.

I studied Electrical and Electronics Engineering at BITS Pilani, and co-authored a peer-reviewed paper on phosphorene electrodes for supercapacitors in ACS Applied Energy Materials.

  • LangGraph · MCP · RAG
  • 10xScale.ai — since Jul 2026
  • Hyderabad / Remote

[02] — Skills

The stack I reach for.

01

GenAI & Agentic AI

Agents, retrieval and evaluation that hold up in production.

  • LangGraph
  • LangChain
  • CrewAI
  • AgentFlow
  • MCP / FastMCP
  • LlamaIndex
  • Multi-Agent Orchestration
  • RAG & Hybrid RAG
  • Prompt Engineering
  • Structured Output
  • Tool Calling
  • Cost Optimisation

02

Models & Voice

Picking the right model and keeping latency low.

  • Gemini 2.5 / Vertex AI
  • Groq (Llama 3)
  • OpenAI-compatible APIs
  • Google Cloud STT & TTS
  • LiveKit
  • VAD

03

Retrieval & Data

Turning messy documents into queryable knowledge.

  • Qdrant
  • pgvector
  • ChromaDB
  • Embeddings
  • Semantic Search
  • Knowledge Graphs
  • PyMuPDF
  • spaCy NER
  • Entity Resolution

04

Backend & Infrastructure

The services and queues that serve the models.

  • FastAPI
  • PostgreSQL
  • Tortoise ORM
  • Redis
  • TaskIQ / Celery
  • Apache Airflow
  • BigQuery
  • DBT
  • WebRTC / WebSockets

05

Languages

Python first, with TypeScript for the edges.

  • Python
  • SQL
  • TypeScript
  • JavaScript
  • C
  • Java (familiar)

06

Cloud & Dev Tools

Shipping and testing it all.

  • GCP
  • AWS
  • Docker
  • Git
  • GitHub Actions
  • Pytest
  • Pandas
  • NumPy
  • Streamlit

[03] — Experience

Where I've been.

  1. Jul 2026 — Present

    AI Software Engineer

    10xScale.ai · Hyderabad

    Nirnaya, a legal investigation platform: FastAPI and TaskIQ/Redis pipelines that parse 1,000+ court and registry PDFs into PostgreSQL, deterministic parsers in place of per-document LLM calls, identity resolution with hard vetoes before fuzzy scoring, and a Chrome extension for data collection. Also an AI tutoring platform: low-latency LiveKit voice, adaptive multi-agent routing on AgentFlow, and MCP tools via FastMCP.

  2. Jul 2025 — Dec 2025

    AI Intern

    10xScale.ai · Hyderabad

    Built a dual-mode medical chatbot with multimodal report analysis (Gemini 2.5 Flash, AgentFlow), a database agent with MCP tools on LangGraph, and a pgvector RAG chatbot over 200+ PDFs.

  3. Apr 2024 — Jun 2024

    AI Intern

    FourKites · Chennai

    Built a multi-agent CrewAI workflow converting XML, CSV and Excel inputs into structured JSON, and used Virtuoso for AI-powered application testing.

  4. Education

    B.E. (Hons.) Electrical and Electronics Engineering

    BITS Pilani, Hyderabad Campus · Hyderabad

    Class XII: Sri Vignesh Vidyalaya (84.6%). Class X: Alpha Wisdom Vidyashram (91.8%).

[04] — Selected work

Built to ship, built to scale.

FastAPITaskIQGeminiPostgreSQLWXT

Nirnaya — Legal Intelligence

Problem
Court and registry records are scattered across portals as unstructured PDFs and HTML.
Role
Built the extraction pipelines, identity resolution, report generators and the Chrome collection extension.
Outcome
1,000+ documents from eCourts, Supreme Court, NCLT, DRT, CERSAI and MCA normalised into PostgreSQL, with deterministic parsers cutting LLM cost and latency.
LiveKitVertex AIAgentFlowFastMCPQdrant

AI Tutoring Platform

Problem
Students needed real-time voice tutoring that adapts to what they are doing.
Role
Built the LiveKit voice pipeline, AgentFlow routing backend and FastMCP tool layer.
Outcome
Low-latency WebRTC audio with VAD-based cost control and agents that switch between video Q&A, live coding and study planning.
LangGraphMCPSQL

Database AI Agent

Problem
Non-technical users can't query diverse database schemas.
Role
Designed the multi-node LangGraph agent and MCP tool layer.
Outcome
Natural-language querying across schemas with a query-understanding preprocessing layer.
pgvectorGemini 2.5 FlashAgentFlow

RAG Chatbot & Medical Assistant

Problem
Answers needed to be grounded in 200+ PDFs and in patient reports and imaging.
Role
Built a pgvector RAG chatbot and a dual-mode multimodal medical assistant.
Outcome
Semantic retrieval-augmented answering, plus patient/expert modes with database-backed appointment scheduling.

[05] — Research

Peer-reviewed, before the agents.

PublicationSecond authorSept. 2026

First-Principles Study of Transition-Metal Doping in Monolayer Black and Blue Phosphorene: A Promising Pathway for High-Performance Electrode Design for Electric Double-Layer Supercapacitors

A. S. Reddy, A. K. BT, A. Tiwari, S. Das, A. Bhattacharjee, S. Kanungo

ACS Applied Energy Materials, vol. 9, no. 18, pp. 12434–12461, Sept. 2026

Read paper ↗doi:10.1021/acsaem.6c01956

[06] — FAQ

Frequently asked questions.

Who is Arun Karthik BT?

Arun Karthik BT is an AI Software Engineer at 10xScale.ai in Hyderabad. He builds production LLM systems — multi-agent orchestration, RAG and hybrid retrieval, and real-time voice agents.

What does Arun work with?

LangGraph, AgentFlow, CrewAI and MCP/FastMCP for agents; Gemini and Vertex AI for models; Qdrant and pgvector for retrieval; and FastAPI, PostgreSQL, Redis and TaskIQ for the backend. Languages are Python and TypeScript.

What has he built recently?

A legal-intelligence platform that parses 1,000+ court and registry PDFs into PostgreSQL, an AI tutoring platform with low-latency LiveKit voice and adaptive multi-agent routing, a database agent on LangGraph with MCP tools, and a pgvector RAG chatbot over 200+ PDFs.

How does he approach cost and latency?

By replacing LLM calls with deterministic parsers wherever accuracy allows, keeping Gemini structured output as a fallback, and using voice activity detection to avoid unnecessary speech processing.

Has he published research?

Yes. He is second author on a peer-reviewed paper in ACS Applied Energy Materials (vol. 9, no. 18, 2026) on transition-metal doping in monolayer phosphorene for supercapacitor electrodes.

How can I get in touch?

Email arunkarthikbalakrishnan@gmail.com, connect on LinkedIn at linkedin.com/in/arun-karthik-0aa085426, browse code on GitHub at github.com/arunkarthik-error404, or use the contact form on this site.

[07] — Contact

Let's talk

Open to backend and platform engineering roles, freelance projects, and interesting product ideas. The fastest way to reach me is email — I reply within a day.