Preksha Barjatya (AI engineer, Indore)Curious by default.Careful by design.

Choose where to start. Everything is one scroll away.

Show me what you've built.Three projects, from resume scoring to cited research papers.
How do you keep it honest?A case study on checking citations in code, not by the model.
Where have you worked?Two internships, one in AI and one in data.

Every path leads back to the work.

IST --:--:--

PythonFastAPILangChainLangGraphRAGGroq APIDockerKubernetesAWSPandasPythonFastAPILangChainLangGraphRAGGroq APIDockerKubernetesAWSPandas

I turn language models intodependable,production-ready tools.

Illustrated portrait of Preksha Barjatya wearing round sunglasses

I'm Preksha Barjatya, a B.Tech CSE (AI & ML) student at Acropolis Institute of Technology and Research, Indore (affiliated to RGPV Bhopal), graduating in 2027.

I build Retrieval-Augmented Generation (RAG) applications, FastAPI-based AI solutions, data pipelines, and SQL-driven analytics systems, turning language models into dependable, production-ready tools.

Currently interning as an AI Engineer at Santerra Hygiene Pvt. Ltd., building agentic automation workflows, after previously working as a Data Analyst Intern at Think AI Corporation.

Degree

B.Tech CSE (AI & ML)

CGPA

7.94 / 10 (2023 - 2027)

Internships

2 (AI & Data)

Focus

Agentic AI (RAG, LLMs)

Skills

The core skills behind an AI engineer, building up from programming fundamentals to the AI/ML work that defines the role.

Programming

(3)

Python
JavaScript
SQL

Python Concepts

(9)

OOP
Functions
Modules & Packages
File Handling
Exception Handling
Decorators
Generators
Context Managers
Type Hinting

Backend Engineering

(4)

FastAPI
REST APIs
API Design
Pydantic

AI/ML & Generative AI

Core

LangChain
LangGraph
RAG
Document Embeddings
Vector Search
Prompt Engineering
Generative AI
Groq API

Cloud & DevOps

(6)

Git
Docker
Docker Compose
Kubernetes
AWS (ECR)
AWS (ECS)

Data & BI

(8)

Pandas
NumPy
Matplotlib
EDA
Feature Engineering
Data Pipelines
Power BI
Tableau

Experience

(Jul 2026 - Present)

AI Engineer Intern

Santerra Hygiene Pvt. Ltd.

(01)
Build AI-powered automation workflows and practical AI applications using Python, prompt engineering, and model integration to streamline a startup's business operations.
(02)
Support business data and finance-related processes, applying automation to improve efficiency and scalability of daily operational workflows.

(Apr 2026 - Jul 2026)

Data Analyst Intern

Think AI Corporation

(01)
Migrated legacy reports to Power BI and Tableau, building interactive dashboards to visualize key business metrics and support decision-making.
(02)
Performed data analysis/EDA using Python and SQL, and designed data pipelines to clean and transform multi-source data into BI-ready formats while gathering requirements with cross-functional teams.

SelectedWork

Resume-to-JD Gap Analyzer

(01)

LangGraph
Tool Calling
Pydantic
Gap Analysis

A multi-agent pipeline that extracts structured profiles from resumes and job descriptions, matches skills and experience, and generates explainable gap analysis with actionable recommendations. Validates both documents, performs semantic matching, identifies missing qualifications. Instrumented to track: analysis runtime, token usage, skill matches, gaps found. Deploy live to gather metrics on real resumes.

Live Site GitHub

Python

LangGraph

Tool Calling

Semantic Matching

Pydantic

FastAPI

Research & Document Intelligence Platform

(02)

LangGraph
Multi-Agent
RAG
FastAPI
PostgreSQL

A unified platform with two capabilities: a seven-agent LangGraph workflow that turns an assignment brief into a cited research paper (every citation verified against Crossref in code, human approves topic and thesis, runs checkpointed in Postgres for crash recovery), and a FastAPI-powered RAG service for document Q&A with pgvector storage and LLM answers grounded in retrieved passages with source citations. Research workflow: [X] min avg runtime, [X] tokens/run, [X] citations verified. Q&A evaluation: ragas score [X] on [X] questions, Dockerized.

Live Site GitHub

Research & Document Intelligence Platform screenshot 1

Python

FastAPI

PostgreSQL

Docker

LangGraph

RAG

pgvector

LangChain

Hackathons& Open Source

Hacktoberfest 2026 · DEV Open-Source AI Challenge, Week 1: Touch Grass

Dharohar धरोहर

An AI audio guide that tells you the story of the places you walk into, while your phone stays in your pocket.

Tap “Walk where I am” and it finds the heritage around you, writes a short grounded script for every stop with open-weight Gemma, voices it with ElevenLabs, then turns the screen black. GPS geofences play each story as you arrive, every stop is cached for dead zones, and it speaks English plus nine Indian languages.

Mastra agent: SerpApi facts → Gemma script → ElevenLabs voice, held to 13 evals
Caught Gemma inventing a balcony that doesn't exist; fixed by grounding every script in search results
Open-weight model swapped in one line when a provider retired the deployed version

Gemma
Mastra
ElevenLabs
SerpApi
Render
Open source

Live App GitHub Write-up on DEV

Dharohar banner: the app over a photo of Kanch Mandir, Indore

A research paper writer where code checks the citations, not the model

Drafting a paper with an LLM fails in two ways that matter more than fluency. It invents or misattributes citations, and it drifts from the brief. I built Research Paper Agents so that code catches both, and so a person approves the direction before any drafting effort is spent.

Live demo Code Full write-up

7
agent roles in one LangGraph workflow: 67
automated tests, including real Postgres: 3 to 4 min
and about 35k to 39k tokens per paper: ~6
live runs used to find the problems below

How it works

Seven steps run along a fixed graph. Two of them wait for a person to approve.

(01)
Brief analyst
Reads word limit, sections and citation style
(02)
Topic strategist
Proposes a topic
You approve this
(03)
Source scout
Searches OpenAlex and Semantic Scholar, keeps only papers whose DOI resolves in Crossref
(04)
Thesis writer
Proposes a thesis
You approve this
(05)
Outliner
Plans sections and assigns sources to each
(06)
Drafter
Writes each section from its own sources
(07)
Critic
Chooses between rewrite and find more sources, up to 4 rounds

Decisions that shaped it

Fixed graph, not an LLM supervisor

Routing is plain code. It is cheaper, easy to test, and fits free-tier rate limits. The only dynamic choices are the two human approvals and the critic's rewrite or find-more-sources call.

Citations verified in code

A source is kept only if its DOI resolves in Crossref and the title and year match. This is the one guarantee that should not depend on model behavior.

Reference list generated by code

The writer can only cite ids of verified sources, and code builds the bibliography, so the model cannot invent a reference entry.

Human approval at topic and thesis

These are the cheapest points to correct direction, before any drafting tokens are spent.

Postgres as run store and queue

One datastore holds runs and jobs. Progress is checkpointed after every step, so a run resumes after a crash or deploy.

What live runs taught me

Real runs against Groq, OpenAlex and Crossref turned up problems that tests with fake data could not.

A prompt went over the provider's token cap

Seen: A thesis request was 8,779 tokens against a limit of 8,000.

Fix: Cap sources at 15, trim abstracts, and show the critic only the body and the sources it cites.

Off-topic sources in the pool

Seen: 42 verified sources included an edge-computing paper for a last-mile delivery query.

Fix: Rank by relevance first, verify in rank order, and stop when the pool is full.

The length limit was missed

Seen: The first draft was 2,086 words against a 1,500 maximum, because the count included a 369-word reference list.

Fix: Count the body only, on the text the reader actually sees.

An invented statistic

Seen: The critic flagged a "12% fuel reduction" that no source supported.

Fix: The writer may not state numbers missing from the abstracts, and the critic flags any that slip through.

A section came back with one word

Seen: Draft 3 of one run had an Analysis section of a single word.

Fix: Retry degenerate output and treat any section under 40 words as a hard issue.

What I did not measure

Paper quality. There is no golden set and no human rating of argument quality.
Whether each sentence is faithful to its source. Every reference is matched to a Crossref record, but sentence-level faithfulness is judged only by an LLM reviewer, which is not proof.
A finished live run approved by the reviewer. The live runs that finished were all flagged for human review, mostly for word count, and the daily token quota ran out before the later length fixes could be confirmed.

Where "lost in the middle" fits

Liu et al. (2024) report that language models use information at the start and end of a long prompt better than information buried in the middle. Several choices in this project happen to agree with that: keeping only the 15 most relevant sources, ranking before prompting, and giving each section only its own sources.

Those choices were driven by token limits and observed failures, not by the paper, and this project did not test its claim. The prompts are also short, a few thousand tokens, far below the lengths where the effect was reported.

A small test is planned but not run yet, because it needs model quota. It would move the most relevant sources between first, middle and last position across 6 to 8 real assignments and compare citation validity, section coverage and reviewer approval. The expected result is little or no difference at these lengths, and a null result would be reported as one.

Certifications

(Certified)

Agentic AI Certified Foundations Associate

Issued Sep 2026
Expires Sep 2028
ID 103539377AAI26OFA

CSE (AI & ML) Dept., Acropolis Institute of Technology and Research, with Navigate Labs(Certified)

Generative AI, RAG, Multimodal & Agentic AI

OpenEDG Python Institute(Certified)

Programming with Python Professional Certificate

Let's worktogether

I'm actively looking for opportunities to grow as an AI engineer. If you have a role or project in mind, let's connect.

prekshabarjatya2105@gmail.com GitHub LinkedIn

(+91) 9993098023 (New Palasiya, Indore, India, 452001)