Assembling home orbit

Dev Gaglani

AI engineer, Mumbai. I simulate fly brains, map the Moon and make agents argue with each other. This is the small universe where all of it lives.

Home Orbit: Start here

What changed because I was there

Told by outcome rather than by date. The résumé version is one click away on LinkedIn.

500+ stocks, traded live. Software Development Intern · SPJIMR, under Prof. Shalini Talwar · June 2026 – now. A full stock exchange for the classroom: 500+ NSE/BSE tickers, AI-driven market events, professor and student dashboards for multi-round sessions. 100+ people trade at once and a Monte Carlo engine prices every tick in under 16 ms.

Number plates, read from live video. Computer Vision Intern, Founder's Office · QuantumFour AI · June – August 2026. Built the ANPR pipeline end to end: YOLOv8 detection, adaptive preprocessing, OCR, post-processing. Then, instead of hand-tuning it, wrote a LangChain agent that benchmarks every preprocessing chain, OCR model and Vision LLM against each other and ships the most accurate one.

200+ people in one room. Vice-Chairperson · National Innovation & Startup Policy Council, SPIT · Sept 2025 – Sept 2026. Led 30+ people across website, marketing and operations to run a startup symposium with 5+ speakers, then a national case competition with 180+ participants.

Pro-bono consulting that shipped. Partner · Sardar Patel Consulting Group, SPIT · Sept 2025 – now. Helped build a health agency's website and ran a national case-study competition with 100+ participants.

The Fly Room: Neuroscience

Wordle vs. a Fruit Fly

Can a fruit fly's brain learn to play Wordle? I wired up the real thing to find out: 3,127 neurons and 179,516 synapses from the FlyWire map of a fruit-fly brain, simulated spike by spike.

The connectome is frozen and used as a spiking reservoir. Only a small readout learns, by reward, the way dopamine teaches the fly. No LLM, no lookup table, and you can watch the brain fire while it plays.

Source on GitHub

A worm that walks (idea, not built yet)

C. elegans has 302 neurons and every connection is mapped. Wire that connectome to a simulated body and see whether it crawls without being told how.

What are you looking at? (idea, not built yet)

Train a decoder on open EEG recordings to guess which image a person was shown, then find out how little signal it actually needs.

KrushiMitra

A farmer's platform where a CNN reads a photo of a leaf, names the crop disease and suggests the remedy. A Random Forest predicts yield, and a WhatsApp bot sends weather alerts and fertiliser advice.

A language model for DNA (idea, not built yet)

The genome is the oldest codebase there is: three billion characters, no comments, and it compiles into you. Train a small transformer on genomes instead of text, then ask which single-letter mutations it finds surprising and compare that with known disease variants.

Shackleton: Space

LunarGPS

Google Maps for the Moon. The lunar south pole has water ice in craters that never see the Sun. I wanted to drive there, so I turned NASA's elevation data into something you can search and steer.

Drive a 3D rover across 370,000 km² of real south-pole terrain, search craters and ice by their official names, and get a route around slopes a rover can't climb. Every place card cites its dataset. The rover you are driving now is a small tribute to it.

Source on GitHub

Land it (idea, not built yet)

A reinforcement-learning agent that learns to set a lander down on the same real south-pole terrain LunarGPS already maps, with fuel as the budget.

A planet in the noise (idea, not built yet)

Search public Kepler and TESS light curves for the tiny, regular dip a planet makes when it crosses its star.

Descent: AI & systems

Veridion

Point it at a GitHub repo and ten scanner agents audit it in parallel: a security score, SOC2 and GDPR mapping on a knowledge graph, AI-written fixes, and a PDF report in your inbox. Ships with a one-click Chrome extension.

Open the productSource

AI SAR Generator

Retrieval-augmented generation, run by a team of agents. It pulls the relevant transactions, KYC records and regulatory guidance from a vector store, then drafts a FinCEN-compliant Suspicious Activity Report with an audit trail back to every source and an XGBoost risk score for the analyst who signs off.

Source

Cortex-X

Specialist AI doctors debate a patient case asynchronously, in the patient's own language, until they agree. A human clinician stays in the loop to validate the call.

Source

Market Simulator

A stock exchange for the classroom: 500 NSE/BSE companies, a Monte Carlo pricing engine, and a news engine that moves the market mid-session. Is a price just a belief with a number on it? This is where I go to test that.

Source

ExamShield

An exam system that refuses to lose an answer. Offline-first incremental sync, a server-authoritative timer, tamper-proof audit logs, and WebSockets that resume cleanly after a refresh, a crash, or a dead network.

Source

LeetTrack

Keeps a friend group honest about the daily grind: streaks, leaderboards and live stats for the whole squad.

Open itSource

A language model, from nothing (idea, not built yet)

Write and train a small GPT with no libraries doing the thinking, then open it up and look for the neurons that learned grammar.

Does the debate help? (idea, not built yet)

A benchmark that runs the same questions through one model and through a panel of agents, and reports honestly when the panel is just slower.

The Plot: About me

Dev Gaglani

I am a third-year Computer Engineering student at SPIT, Mumbai, with a FinTech minor on the side. I like questions that sound slightly unreasonable: can a fly's brain play Wordle, where would you park a rover on the Moon, can a group of agents reason better than one.

Most of what I build is an attempt to answer one of those properly, with real data, and then make it something other people can poke at. This island is the part of the universe that is just me. Add a block before you leave.

Education

B.Tech, Computer Engineering: 8.86 CGPA Sardar Patel Institute of Technology · 2024 – 2028

Minor, FinTech Engineering: 9.00 CGPA L&T EduTech · 2024 – 2028

Grade 12: 87.67% MHT-CET state rank 932, 99.6 percentile · 2024

Grade 10, IGCSE: 97.8% School rank 1 · 2022

Rooms I got into

EY Techathon 6.0: Top 0.03% Semi-finalist, top 57 of 185,000+ participants

IEEE HackFusion (International): Top 2.5% Finalist, top 50 of 2,000+ teams

NMIMS National Hackathon: Top 10 Of 615 participants

ITM Summerhacks 2.0: Top 10 National level

Smart India Hackathon: Top 50 Of 500+ teams

Codeforces: 1292 Max rating, 150+ problems solved

What I reach for

Courses that stuck: Machine Learning Specialization (Stanford), Real-Time Video AI Applications (NVIDIA), Algorithmic Trading (IIT Bombay).

Languages: Python, C/C++, TypeScript, JavaScript, SQL

AI / ML: PyTorch, scikit-learn, OpenCV, YOLOv8, LangChain, LangGraph, RAG, XGBoost, Vision LLMs

Web: FastAPI, Flask, React, Next.js, Node.js, Three.js

Data & ops: PostgreSQL, Neo4j, ChromaDB, Redis, Docker, Linux

dev.gaglani1@gmail.com