Ankit Surana · AVP Data Engineer at NatWest

I make LLM answers you can check.

RAG, NL-to-SQL and agentic AI on top of ten years of event-driven data platforms. I care about the whole path, from the data model up to the number someone acts on.

  • 500M+transactions a day on the platform I work on
  • <5sto answer a customer's question about their spending
  • 10+ yrsbuilding data platforms, 3.5 of them in Edinburgh
  • 2 talkson retrieval and agents for upGrad in 2026
Ankit Surana
Open to conversations Bengaluru
01

Talks

What I have been explaining on stage this year. Each talk ships with code you can run on a laptop.

Conference talkSeptember 2026Bengaluru

Retrieval as a Guardrail: deterministic answers from natural language

NL-to-SQL assistants fail quietly: the query runs, the number is wrong. This talk puts retrieval before SQL generation and routes every question down one of five paths, so the assistant answers from a template, fills one slot, asks, explains, or refuses, instead of guessing.

upGrad AI Marathon Bengaluru

MasterclassOctober 2026Online

Build Your First AI Agent: from ML and LLMs to real-world workflows

Build a tool-using support agent, then put it on a budget: the big model labels data once, a small classifier learns from it, and each message stops at the cheapest helper that can answer. 1.9× cheaper and 1.6× faster on the test set.

upGrad AI masterclass

02

Work

Each one has a write-up, so it stands on its own whether or not the source is public. Open repos link out as well.

03

Field notes

One real problem per post. What broke, why I did not see it coming, and what fixed it.

04

About

Ten years of making sure the numbers turn up on time.

I am Ankit Surana, an AVP and data engineer at NatWest in Bengaluru. Before that I spent close to eight years at Infosys, three and a half of them based in Edinburgh, ending as a technical lead. Most of my career has gone into event-driven pipelines on AWS. Kafka, Lambda, Glue, Redshift, Athena. The platform I work on carries north of and has to answer .

Lately that has meant agents. I was on the team that shipped a personalised spending assistant to customers, one of the first agentic AI features any bank put in front of the public, alongside a retrieval pipeline that reads the documents nobody wants to read. Today my work is the infrastructure underneath LLM applications: RAG pipelines, NL-to-SQL engines, agents and the LLMOps that keeps them honest. I hold an MSc in AI and Machine Learning from Liverpool John Moores and an AWS Solutions Architect certificate, and this year I started speaking about it, for upGrad in Bengaluru and online.

What I keep returning to is ownership. I would rather understand a system from the data model up to the button someone clicks than hold one slice of it very well. That is why I build things on my own time, and why none of them are ever quite finished. The latest is NoByte, a growing set of browser tools and games that run with no server and no account, mine end to end from the maths to the pixels.

What that added up to

  1. 80%less manual turnaround on insights covering roughly 500 million daily transactions, after rebuilding the pipeline end to end in Apache Airflow.
  2. <5sto answer a customer's question about their own transactions, by architecting agentic AI services across AWS, Azure AI and PostgreSQL.
  3. 88%accuracy classifying financial transactions, reached by working out which predictive parameters carried signal and dropping the rest.
  4. 40%off code management processing time across 50+ repositories, by automating the move from Bitbucket to GitLab. That one earned a Risk Hero Award.
  5. 30%faster queries through the Teradata to Redshift migration, because the reporting SQL was rewritten rather than lifted across unchanged.
  6. 20%more output from a team of ten analysts, after mentoring them and changing how the work was split.

On the record

Data Hero Award, NatWestRecognised for data engineering work2025
Risk Hero Award, NatWestAutomating code management2024
Speaker, upGradAI Marathon Bengaluru, and an online AI masterclass2026
AWS Solutions ArchitectAssociate, and Cloud Practitioner2024
MSc, AI & Machine LearningLiverpool John Moores. Thesis on clustering chatbot conversations for golden datasets and drift monitoringJul 2026
Executive PG, ML & AIIIIT Bangalore. Academic Star of the cohortAug 2026
Google DeepMind HackathonSelected, 1 of 250 from 4,000+ applicants2026
Winner, LJMU group projectCash In Loop. Concept to working POC in three days2025
05

Skills

The tools I reach for without thinking about it.

Languages

PythonSQLJavaScriptDart

Streaming & big data

KafkaApache SparkAirflowHadoopHiveElasticsearch

Cloud

AWS LambdaGlueRedshiftAthenaKinesisS3EMRSageMakerECSAzure AI

GenAI & agents

RAGNL-to-SQLLangChainLangGraphAgentic systemsMCPLangfuseLLMOpsOpenAI APIGemini

Machine learning

scikit-learnTensorFlowPySparkAnomaly detectionForecastingOpenCV

Data

Data modellingWarehousingETLPostgreSQLMongoDBRedis

Platform

DockerKubernetesTerraformGitLab CIJenkins

On-device & backend

FlutterTFLiteFastAPIDjangoFlaskREST

Let us talk about the hard part.

If you are working on retrieval, agents or the data underneath them, want a talk, or just want to argue with a post, I read everything.

Usually replies within a day

Back