Skip to content
IntroWorkResearchPathContact
01 —

AryanGurav

I build agent systems, then find out what breaks once someone is actually using them. Browser agents, a knowledge graph laid over systems that were never meant to talk, retrieval that checks its own work. Senior engineer at TheAgentic.AI. Three papers on GANs before that.

Open to interesting problems

Agents that hold up outside a demo.

Six of them below

Mumbai, India · 19.07°N 72.87°EMove the cursor — the ink answers · click to rippleScroll — the type comes apart
02

Selected work

01 / 06
01 / 06Open source

TheAgentic Browser

A browser agent that reads the page instead of following a script. Change a class name and a CSS-selector scraper dies; this one looks again and keeps going.

My role

I built the agent framework and the orchestration layer, plus the context pipeline that turns a raw DOM tree into something small enough for a model to hold at once.

View on GitHub
PydanticAILLM OrchestrationDOMPython

How it works

  1. 01Agent coordination and lifecycle on PydanticAI.
  2. 02A DOM layer that flattens raw HTML into structured state a model can read.
  3. 03Context handling that keeps agent state alive across multi-step flows.
  4. 04Action chosen at each step from the page and the task objective.
  5. 05Recovery paths for the page states nobody planned for.

Outcome

Open source, and the data-collection layer under several client builds.

02 / 06Sole architect

OrgMind

One hierarchical knowledge graph laid over systems that were never built to talk to each other, so an agent can ask a question that crosses all of them.

My role

Sole architect. Graph schema, ingestion, the Graph RAG layer and the API, all of it from an empty repo.

Neo4jKnowledge GraphsGraph RAGFastAPI

How it works

  1. 01A hierarchical Neo4j schema for entities, relationships and workflows.
  2. 02Ingestion pipelines that normalise disparate enterprise systems.
  3. 03Graph RAG: traversal over the graph feeding the retrieval step.
  4. 04A FastAPI layer so any agent can query it.

Outcome

Now the data backbone our agentic workflows run on. The queries it answers are the ones that stayed impossible while every system kept its own store.

03 / 06Co-architect

CortexON

A general-purpose agent framework. Plan, act, reflect, and bolt on new tools without opening up the core loop.

My role

Co-architected it, wrote more of it than anyone else, and mentored the junior devs building on top.

View on GitHub
PythonFastAPIAgent ArchitectureLLMs

How it works

  1. 01A core loop with planning, execution and reflection phases.
  2. 02Tool interfaces modular enough to extend without touching core logic.
  3. 03Memory and context management for sessions that run long.
  4. 04Review practice and mentorship for the juniors on the project.

Outcome

Open source, and the base layer for several internal and client deployments since.

04 / 06Contributor

Deepresearch V3

A search engine that splits a question into sub-questions, retrieves against each one, and returns structured research with its sources still attached.

My role

I worked the retrieval pipeline and the reasoning layer, mostly on multi-hop questions that need more than one pass to answer.

SearchRAGLLM ReasoningPython

How it works

  1. 01Hybrid retrieval: dense embeddings alongside sparse keyword matching.
  2. 02Reasoning chains that decompose a query into sub-questions.
  3. 03Source attribution and citation tracking on every output.
  4. 04Parallel retrieval and caching to hold latency down.

Outcome

Runs on our internal research workflows.

05 / 06Research · Springer

DeepCAS

A GAN that pairs self-attention with conditional generation and label smoothing, aimed at satellite imagery good enough to train on.

My role

Co-author. Architecture design, the attention mechanism, and the experimental evaluation.

Read the paper
Self-AttentionGANsLabel SmoothingDeep Learning

How it works

  1. 01A conditional generator with self-attention for long-range spatial structure.
  2. 02Label smoothing to steady training and cut mode collapse.
  3. 03Evaluation on EuroSAT by FID, IS and classification accuracy.
  4. 04Fidelity and diversity gains over the baseline GANs.

Outcome

Published in Springer Nature IJIT (2025). Beat the baselines we tested against on EuroSAT generation.

06 / 06Research · IEEE

GAN Satellite Classification

Four GAN variants put against each other on EuroSAT to find out which ones actually move land-use classification, and which only look like they do.

My role

Lead researcher. Designed the experiments, implemented the four variants, trained the classifiers, wrote the paper.

Read the paper
GANsCNNPyTorchRemote Sensing

How it works

  1. 01Implemented InfoGAN, CGAN, CDCGAN and ACGAN generators.
  2. 02Grew EuroSAT by 50% across every land-use category.
  3. 03Trained CNNs on original against augmented sets to measure the delta.
  4. 04Scored on accuracy, precision, recall and FID.

Outcome

CDCGAN took accuracy, ACGAN took precision and recall. IEEE ICESC 2024, cited by 4.

03

Research

3 papers · IEEE, Springer
2024

Comparing the Effectiveness of GAN-Based Methods for Enhancing EuroSAT Satellite Imagery Classification

Four GAN variants grew EuroSAT by 50%, then CNNs trained on the result to measure what the extra data bought. CDCGAN took accuracy; ACGAN took precision and recall.

First authorIEEE ICESCCited by 4
2025

DeepCAS: A Novel GAN Architecture for High-Fidelity Satellite Image Generation

Deep Conditional Attentional Smoothing. Self-attention for the long-range structure, conditional generation for the class, label smoothing to keep training from collapsing.

Co-authorSpringer IJITJournal article
2024

Generative AI in Marketing to Build Efficiencies through Content Creation and Beyond

What LLMs and vision-language models do to marketing and content workflows once they land inside a global capability centre.

Co-authorWelingkar ICSLTConference
04

The path so far

Drag →
Feb 2025 — Present

Senior AI Engineer / Forward Deployed Engineer

TheAgentic.AI · Mumbai

  • Built the agent framework, orchestration and DOM context pipeline behind TheAgentic Browser.
  • Co-architected CortexON. Highest code contributor, and mentor to the juniors on it.
  • Sole architect of OrgMind: hierarchical knowledge graphs and Graph RAG for queries that cross systems.
  • Led 4 client projects end to end: grants automation, automotive, marketing, sports analytics. Contributed to 3 more across pharma, crypto and fundraising.
Dec 2024 — Feb 2025

AI Intern

TheAgentic.AI · Mumbai

  • Wrote the first automated navigation and scraping passes for TheAgentic Browser.
  • Built a multi-agent coordination framework on PydanticAI.
  • Built the agent architectures that had to hold up once a client was on them.
Jun 2024

Generative AI Research Intern

Thadomal Shahani Engineering College · Mumbai

  • Researched generative models for marketing applications.
  • Published "Generative AI in Marketing" at the 5th International Conference on Sustainability, Leadership and Technology.

0+

Client projects led

0

Open-source systems

0

Papers published

PythonC/C++JavaJavaScriptSQLFastAPIReact.jsNext.jsPydanticAINeo4jPostgreSQLLLM OrchestrationAgent ArchitecturesKnowledge GraphsGraph RAGContext EngineeringDeep LearningGANsPandasGit
05

Say hello

Agents, graphs, retrieval, or an idea you can't stop thinking about. I read everything. Fair warning: I will ask what the eval looks like.

© 2026 Aryan GuravPress I ink pass · D dark · type agent · ↑↑↓↓←→←→baBuilt with too much coffee