AI Product Engineer · India

Arpan Gupta

I build production RAG systems

LLM systems on AWS that hold up in production: retrieval that finds the right document, agents with guardrails, and evaluations that prove the numbers moved.

See what I build Get in touch
scroll
0%
RAG answer accuracy, up from 82.4% on a 191-question eval
0%
Fewer tokens per answer (66-72%) with no accuracy loss
0+
Automated tests on the agentic platform
0
Confidently wrong document picks across 88 live queries
01

About

I'm an AI Product Engineer at ODEK Appcraft India, where I take LLM features from idea to production for clients in travel, insurance and staffing. I hold an MSc in Artificial Intelligence & Robotics from the University of Glasgow and started my career as a backend engineer at Verizon and Infogain.

I care about the unglamorous parts that make AI products trustworthy: evaluation sets, routing, guardrails, cost per call and latency. I'm an AI-first developer: most of what I ship today is built with Claude Code and MCP in the loop.

whoami.json
{
  "role": "AI Product Engineer",
  "focus": ["RAG", "agents", "evals"],
  "stack": ["Bedrock", "Claude", "Python", "TypeScript"],
  "education": "MSc AI & Robotics, Glasgow",
  "based_in": "India",
  "open_to": "Applied AI roles"
}
02

What I build

how a question flows through my RAG service

?Questionstaff portal, per-client config
⚙Query stagefix typos, split, route
⌕Retrievevector search per part
⇅Reranktrim to what matters
✦AnswerClaude, grounded
✓EvaluateLLM-as-judge, recall
Production RAG

Retrieval that finds the right document

A multi-tenant RAG microservice live on client staff portals in travel and insurance, with per-client, per-use-case prompt and model configs.

  • Accuracy 82.4% → 95.6%, tokens down 66-72%
  • LLM query stage that corrects, splits and routes questions
  • Off-topic questions refused 24 of 24 times
Retrieval R&D

Benchmarks, not guesses

Compared S3 Vectors, OpenSearch and pgvector hybrid search, chunk sizes and reranker pools with recall and LLM-as-judge metrics.

  • Picked the lowest-cost backend at equal quality
  • Ingestion-time LLM classifier: new knowledge bases with zero code changes
  • Evaluated self-hosted gpt-oss-20b on vLLM
Agentic Systems

Agents with guardrails

A multi-tenant platform for multi-agent workflows on Orkes Conductor with Bedrock tool calling and MCP tool servers.

  • Database-driven guardrails: kill switch, RBAC, rate limits
  • Per-call cost tracking on every billable LLM call
  • LLM failure triage and daily AI-generated ops reports
Conversational AI

Assistants people can talk to

A voice-enabled booking assistant prototype for a luxury travel operator, built on Bedrock Agents with Lambda tools and structured JSON turns.

  • Gathers trip needs in natural conversation
  • Recommends lodges, off-peak dates and offers
  • Rich cards for quotes and provisional bookings
03

Experience

  1. Sep 2025 - Present

    AI Product Engineer · ODEK Appcraft India

    • Built and optimised a multi-tenant RAG microservice (TypeScript, Bedrock Knowledge Bases, S3 Vectors, Claude) used by travel and insurance clients.
    • Raised answer accuracy from 82.4% to 95.6% and cut tokens by 66-72% with an LLM query-routing stage.
    • Building a multi-agent platform on Orkes Conductor with MCP servers, guardrails and per-call cost tracking (700+ tests).
    • Delivered staff and freelancer portals from zero to production on Zoho Catalyst and Zoho CRM with AI-first development.
    BedrockClaudeRAGOrkesMCPTypeScript
  2. Jun 2024 - Aug 2025

    Freelance Web Developer · Self-employed

    • Designed and built responsive websites for restaurants and cafes in Glasgow, with seasonal and festival campaign updates.
    • Integrated table booking, payment gateways and social media links to improve visibility and customer access.
    • Provided ongoing maintenance and feature updates based on client feedback.
    JavaScriptResponsive designBooking & paymentsClient delivery
  3. Sep 2022 - May 2024

    MSc Artificial Intelligence & Robotics · University of Glasgow

    • Graduated with a 2:1. Coursework in Machine Learning, Deep Learning, Conversational Systems, HCI and Robotics.
    PyTorchDeep LearningROSDialogue Systems
  4. Apr 2022 - Sep 2022

    Software Development Engineer · Verizon Data Services

    • Built Django features and UIs for a device forum platform, speeding up backend response with caching and query optimisation.
    • Migrated key functionality to AWS, cutting infrastructure costs and improving server efficiency by 25%.
    • Led sprint planning with product owners, completing 90% of prioritised user stories.
    PythonDjangoAWS
  5. Aug 2021 - Mar 2022

    Software Engineer · Infogain

    • Developed Python and Django features for a US electrical equipment manufacturer.
    • Raised test coverage by 21% and automated 100+ regression cases with Selenium.
    PythonpytestSelenium
  6. Jul 2017 - Jul 2021

    B.Tech Computer Science and Engineering · Bharati Vidyapeeth College of Engineering

    • Co-authored a paper on stress detection using facial landmarking techniques.
04

Projects

Deep Learning · PyTorch

Colon Cancer Cell Classification

A CNN with batch normalisation and data augmentation that classifies histopathology images of cell nuclei into 4 types: cancer, immune, connective and normal epithelial.

view on GitHub →
Conversational AI · MSc

Sportify: Football Chatbot

A task-oriented dialogue system for Premier League football on Dialogflow CX, pulling live data through webhooks running on Google Cloud Functions.

view on GitHub →
Robotics · MSc

Baxter Robot Playing Chess

A simulated Baxter robot that picks and places chess pieces by the rules of the game, built with ROS, Gazebo and Python.

05

Skills

Generative AI & LLMs

Amazon BedrockClaude Sonnet & HaikuNovaTitan EmbeddingsPrompt engineeringStructured outputsTool callingModel routingvLLMPyTorch

Agents & Orchestration

Multi-agent workflowsMCP serversOrkes ConductorLangGraphBedrock AgentsGuardrailsHuman-in-the-loop

RAG & Evaluation

Bedrock Knowledge BasesVector & hybrid searchChunkingRerankingQuery routingLLM-as-judgeHallucination trackingCost optimisation

Engineering

PythonTypeScriptNode.jsDjangoAngularPostgreSQLpgvectorSQL ServerRedisAWSDockerJenkinsZoho Catalyst & CRM
06

Beyond code

My time in Scotland turned me into a hiker: the Old Man of Storr, the Cairngorms and Goatfell on Arran so far, with Ben Nevis next on the list. I also discovered kayaking on Loch Lomond, and I'm hooked.

07

Let's build something intelligent.

I'm open to Applied AI and AI product engineering roles. The fastest way to reach me is email.

itsapn.gupta@gmail.com LinkedIn GitHub