Skip to content

An animated diagram of a multi-agent pipeline: a document token travels between glowing agent nodes and is stamped approved at the end.

AI ENGINEER · AGENTIC & LLM SYSTEMS · MUMBAI, IN

I orchestrate AI agents that get real work done.

I build multi-agent systems, RAG pipelines, and LLM-driven document workflows — then ship them end to end with FastAPI, PostgreSQL, and Next.js. Currently engineering a healthcare-credentialing platform at Vitco.

View my workGet in touch
GitHubLinkedInEmailDownload résumé

Scroll

01 / ABOUT

From dataset to deployment, I own the whole loop.

I'm Suraj — an AI engineer who's happiest where language models meet production reality. At Vitco I work on a 3-person team building a healthcare-credentialing platform, where I turned a flaky LLM-classification step into a deterministic, field-driven routing layer that cut agent runtime ~25% and killed redundant token spend. Before that I trained a voice-clone detector to 92.5% on a dataset I built by hand, published an IEEE paper on transformer-based sentiment analysis, and shipped agentic research tools with LangChain and RAG.

I care about systems that are fast, correct under load, and actually maintainable — and about the craft of shipping them. Outside the terminal: origami, piano, chess, and the occasional beach cleanup back home in Mumbai.

Role
Jr. AI Engineer, Vitco
Focus
Agentic/LLM systems
Base
Mumbai, IN · remote-first
Open to
Senior/remote AI roles, intl.

02 / EXPERIENCE

The pipeline in production

Junior AI Engineer · Vitco Software Solutions

Backend · 3-person team · Healthcare credentialing platform

MAY 2025 — PRESENT/Mumbai

  1. Deterministic routing layer

    ~25% faster

    Replaced LLM classification in a CrewAI multi-agent document pipeline with a declarative, field-driven routing layer built from deterministic tool composition. Cut agent runtime ~25% and eliminated redundant LLM calls and their token cost.

    Before — LLM classifier with retries · After — deterministic field-driven router

  2. Hardened the backend under load

    0 batch failures

    Fixed connection-pool exhaustion and transaction-rollback bugs that were breaking batch document uploads under concurrency in the FastAPI/PostgreSQL backend.

  3. Pre-scan before the agents run

    fewer LLM calls

    Added a pre-scan step that extracts expiry and reference data from structured fields before invoking agents, so the expensive agentic work only runs when it needs to.

  4. Re-run without re-ingestion

    idempotent reruns

    Designed a re-run endpoint to re-execute agents on a document without a full re-ingestion pipeline — faster iteration and cheaper recovery.

03 / SELECTED WORK

Deployed agents

Sample input

Classifier confidence92.5%

Verdict: Human voice

01

AI-Generated Voice Classifier

Can you trust a voice? This model catches AI clones 9 times out of 10.

A deep-learning classifier that detects AI-generated / cloned voices at 92.5% accuracy, trained on a self-curated 5,000-sample dataset (3,000 AI-generated) with Librosa feature extraction and a TensorFlow model.

  • TensorFlow
  • Librosa
  • Python
  • NumPy
  • 92.5%accuracy
  • 5,000samples
  • 3,000AI-gen
View on GitHub

02

AutoResearcher

Give it a topic. It researches, reasons, and writes the paper.

An agentic research system that autonomously researches a topic, retrieves and summarizes sources, and generates structured papers via RAG, streaming output through a FastAPI backend.

  • LangChain
  • GPT-4o
  • FastAPI
  • RAG
  • Streaming
  • Autonomous multi-step
  • Streaming output
View on GitHub

Prompt

> remind me to send the Q3 report to Priya on Friday at 5pm

03

TaskBot

Natural language in, structured actions out.

A task-automation assistant built on Next.js and OpenAI Structured Outputs, turning freeform requests into reliable, schema-validated actions.

  • Next.js
  • TypeScript
  • OpenAI Structured Outputs
  • Schema-validated actions
View on GitHub

HOW IT WORKS

Inside the credentialing pipeline

  1. 01

    Document arrives

    A credentialing document lands in the platform — a licence, a certificate, an attestation. Batch uploads hit the FastAPI backend under real concurrency.

  2. 02

    Pre-scan extracts fields

    Before any agent runs, a pre-scan pulls expiry dates and reference data straight from structured fields. Cheap, deterministic, and it answers most of the question on its own.

  3. 03

    Deterministic router

    Instead of asking an LLM to classify the document, a declarative field-driven router composes the right tools. No guesswork, no redundant tokens, reproducible every run.

  4. 04

    Agents run where needed

    The expensive agentic work only fires on what the router couldn't settle. Result: the same output, ~25% less agent runtime, and a re-run endpoint that skips re-ingestion entirely.

04 / WRITING

Published research

IEEE Xplore · 2023

Automated Tweet Sentiment Analysis Using Machine Learning

An XLNet-based transformer pipeline combined with a sentiment lexicon for multi-class tweet sentiment classification and trend forecasting.

  • XLNet
  • Transformers
  • NLP
  • IEEE
Read on IEEE Xplore

Certifications

  • AWS Certified Cloud Practitioner
  • Generative AI with LLMs — DeepLearning.AI

05 / TOOLKIT

The toolkit

AI / LLM

  • CrewAI
  • LangChain
  • RAG
  • TensorFlow
  • OpenCV
  • scikit-learn

Backend

  • Python
  • FastAPI
  • PostgreSQL
  • SQLAlchemy
  • REST

Cloud

  • AWS Bedrock
  • AWS Textract
  • AWS S3

Frontend

  • Next.js
  • TypeScript
  • React
  • Tailwind
  • React Native

Also

  • JavaScript
  • C++
  • SQL
  • Git
  • Pandas
  • NumPy

Beyond code

Piano · Origami · Badminton · Chess

Hackathons: 4 · Finalist — Rakuten Code the Future, TetraFlip

Languages: English · Hindi · Gujarati

06 / CONTACT

> open a new run

Let's build something that thinks.

Open to senior, remote-first AI-engineering roles worldwide.

Email me
Download résumé
GitHub
LinkedIn
Email
surajshah.bombay@gmail.com
Phone
+91 98929 30800
Based in
Mumbai, Maharashtra, India