> agent_router.initialize()

Arjun
Vankani

I design and ship multi-agent AI systems — LLM agents that route intent, retrieve knowledge, query databases, and act autonomously in production. 2+ years across Agentic AI, RAG, Computer Vision, and NLP, from research labs to live marketing platforms.

01

About

I'm an AI Engineer building production-grade agentic systems — from a multi-agent "Super Agent" that routes marketing queries across specialized sub-agents, to RAG pipelines backed by vector search, to research on person re-identification and haze removal for surveillance video. I care about systems that don't just generate text, but reliably retrieve, decide, and act.

Currently shipping agent infrastructure at OpinionAds in Ahmedabad, and previously trained 100+ learners in AI/ML as a Master Trainer for SAP's Code Unnati program.

2+
YEARS IN AI/ML
5
AGENTS SHIPPED
9.13
B.E. CGPA
1
BOOK CHAPTER
02

Capability modules

MOD_01 // AGENTIC AI

Agents & Orchestration

LangGraphMulti-Agent Systems Tool/Function CallingAgent Orchestration Guardrails & Eval
MOD_02 // LANGUAGE

LLMs & RAG

LLMsRAGLangChain Prompt EngineeringFine-tuning HuggingFaceOpenAI APIs
MOD_03 // RETRIEVAL

Embeddings & Vector Search

FAISSpgvector Semantic SearchEmbedding Models
MOD_04 // VISION

Computer Vision

CNNsYOLOv8CLIP Stable DiffusionOpenCVOCR
MOD_05 // INFRA

Cloud & Backend

AWS S3Aurora PostgreSQL DockerSpring BootFlask REST APIs
MOD_06 // CORE

Languages & Frameworks

PythonJavaSQL TensorFlowPyTorchScikit-Learn
03

Experience

AI Engineer@ OpinionAds
Aug 2025 — Present · Ahmedabad
  • Architected a multi-agent Super Agent that routes user requests across specialized AI agents based on intent, reducing manual query handling.
  • Built a RAG Agent on a pgvector-backed Aurora PostgreSQL pipeline for accurate, context-aware business responses.
  • Built a SQL Agent translating natural language into queries, letting marketing teams self-serve insights without SQL knowledge.
  • Designed evaluation & guardrail mechanisms to catch hallucinations and ensure reliable LLM output in production.
AI ML Master Trainer@ Edunet Foundation
Oct 2024 — Aug 2025 · Ahmedabad
  • Delivered AI/ML sessions under the SAP-sponsored Code Unnati Program, covering Power BI, SAP, ML and Deep Learning.
Data Associate Engineer@ Trestle
Mar 2024 — Nov 2024 · Ahmedabad
  • Built an OCR pipeline for automated data extraction from images.
  • Contributed to election analysis (CSDS/Lokniti) and food-security analysis (IFPRI, Sri Lanka) projects.
Research Associate@ Intelligent Surveillance Research Lab, DAIICT
Jan 2023 — Aug 2024 · Gandhinagar
  • Researched person re-identification under same-clothing conditions (GUJCOST project).
  • Developed haze removal techniques for a DST-funded research project.
NLP Engineer Intern@ Floatbot
Jan 2022 — Jun 2022 · Bhavnagar
  • Built NLP models for Indic languages and deployed via Docker + FastAPI for voice/chat bots.
04

Projects

~/projects 6 directories
00 semantic-image-search-clip/ text-to-image retrieval via CLIP embeddings independent 2025
+ CLIP-embedding pipeline indexes, ranks, and retrieves the most semantically relevant images.

Text-to-image retrieval system enabling search across both global datasets and local folders — queried via CLI.

CLIPEmbeddingsPython
01 comfyui-n8n-automation/ CSV-driven creative workflow orchestration independent 2025
+ Routes AI tasks from CSV files to the correct ComfyUI pipeline via n8n orchestration.

Automates image generation, video generation, upscaling, and segmentation through seamless HTTP-based integration.

n8nComfyUIHTTP
02 environmental-ai-toolkit/ multi-modal text + image analysis toolkit independent 2025
+ Single Gradio app unifying NLP and vision tasks for environmental analysis.

Sentiment analysis, NER, fill-in-the-blank, and QA alongside classification, detection, segmentation, and image generation.

HuggingFaceStable DiffusionGradio
03 llamaweb/ fine-tuned LLaMA + RAG over documents research 2024
+ LangChain document pipeline chunks, embeds, and enables context-aware Q&A.

Fine-tuned LLaMA for domain-specific knowledge retrieval, evaluated using BLEU score against reference answers.

LLaMALangChainFine-tuning
04 person-reidentification/ tracking + re-ID for surveillance video research 2023
+ Omni-Scale features outperformed alternative extraction methods in evaluation.

Object identification and tracking via CNNs and Deep SORT, with a FaceNet model for facial-feature accuracy.

CNNDeep SORTFaceNet
05 haze-removal/ U-Net dehazing benchmarked vs. 4 methods research 2023
+ U-Net outperformed DCP, COA, FFA, and AOD across light and heavy haze conditions.

Synthesized hazy images using beta and airlight parameters to validate removal quality before re-identification.

U-NetComputer Vision
05

Education & certifications

M.Tech, ICT (Machine Learning)

Dhirubhai Ambani Institute of ICT (DAIICT)
2022 — 2024 · Grade 8.16

B.E., Computer Science & Engineering

Government Engineering College, Bhavnagar (GTU)
2018 — 2022 · Grade 9.13
Generative AI with LLMs — AWS ML Specialization — Stanford Neural Networks & Deep Learning — DeepLearning.AI Applied ML in Python — U. Michigan Python for Data Science — IBM Data Analysis with R — Google