Arun Raghav S · BNY India Co-founder, Crayonz.ai & IITianVibes (IIT Jodhpur 2025)
AI Engineer. Ships production LLM agents.
Channel agnostic
Voice, chat, WhatsApp, IG DM, MCP servers. Whatever channel the user is already on.Math-heavy,football-obsessed, allergic to slop.
Selected work
Propzing (prop8t.ai)
Multi-channel real-estate AI for Dubai
A Dubai real-estate B2B needed a single AI system serving agents across voice calls, WhatsApp, and web chat over 1.6M+ DLD transaction records, with production reliability.
Voice agent on OpenAI Realtime + Exotel telephony; Node/WebSocket on Cloud Run
40-tool MCP + generative apparel studio + IG DM agent
College-branded apparel commerce needed to meet users where they already are: inside their LLM chat and inside their Instagram DMs, with a working design studio, real payments, and print-on-demand fulfilment behind it.
40-tool MCP server on Cloudflare Workers with OAuth 2.0, turning Claude, ChatGPT, Perplexity, and Mistral into a full apparel design studio
Fabric.js + Gemini generative design studio with displacement-mapped mockups on real garment templates
IG DM commerce agent on LangGraph + Gemini 2.5 Flash with 23 tools, including a JEE exam solver with SymPy verification
Cashfree checkout + Qikink POD fulfilment across the storefront and both agents
Financial data workflows across a large capital markets team needed automation without losing the auditability and precision required at a top-tier bank.
Production AI data agents automating analysis and manipulation of financial datasets for analyst teams
Designed and implemented data models and schemas for the Pershing clearing & custody team, with contracts adopted across downstream consumers
Led model migration of legacy financial data models onto a modernised schema, validated end-to-end against production volumes with zero downtime
Markets · Corporate Treasury
Analyst
India
Pune
current
Aug 2025
Python
SQL
Financial data modeling
Otsuka Corporation (Tokyo)
Multi-modal RAG + autonomous multi-agent game
Two research questions in one summer internship: could a vision-language RAG pipeline meaningfully outperform a text-only RAG on mixed-media documents, and could bounded LLM agents play a full social-deduction game without collapsing under token cost?
Multi-modal Advanced RAG pipeline extracting text, tables, and image summaries from PDFs via a vision-language model; custom multi-vector retriever
Autonomous multi-agent game on Phaser JS + LangChain, with bounded reasoning loops so a group of LLM players could run a Werewolf-style deduction match
Short-form vertical video (Instagram Reels, YouTube Shorts) usually means motion-designer hours per clip. HyperFrames is the "author it as HTML, render it as MP4" pipeline that ships reels on repeat without opening a video editor.
HTML-composition + timing DSL that treats data-* attributes as the animation timeline
Deterministic HTML → MP4 renderer with beat-synced audio, blueprints for scenes and transitions
8+ shipped reels across brand and content workflows, including week-round-up, elevator, listicle, and typewriter compositions
Compare six face-detection approaches head-to-head on the same dataset, then feed the best detector into a face-recognition and clustering pipeline strong enough to work on real photos.
Benchmark of Viola-Jones, HOG, SSD, YOLO, Faster R-CNN, and MTCNN on the same face-detection task
End-to-end pipeline with MTCNN + FaceNet embeddings for recognition, deep-CNN embeddings for clustering
