class AryanJain:
def __init__(self):
self.name = "Aryan Jain"
self.institution = "IIT Madras — BTech Chemical Engineering (2024–2028)"
self.cgpa = 8.84
self.jee_rank = "99.87 %ile in JEE Mains 2024 (15L+ candidates)"
self.research = [
"Stochastic Robotics Lab, IISc (Advisor: Shishir Kolathaya)",
"ValenceAI — LLM × CFD Research (Advisor: S. Abhinav Raman)",
]
self.papers = [
"NeurIPS 2026 (under review) — RL Environment for ML Integrity Auditing",
"SurgeLLM @ ACL 2026 — When LLMs Orchestrate but Do Not Compute",
]
self.ip = "Provisional patent: streaming multi-agent treasury orchestration"
self.interests = ["Robotics", "Mechanistic Interpretability", "Quantitative Systems", "RL"]
def say_hi(self):
print("Let's build something that matters. 🚀")
me = AryanJain()
me.say_hi()|
Research Intern — Robotics × AI
|
Undergraduate Research Project
|
| Type | Work | Venue |
|---|---|---|
| 📝 Paper | "An Interactive RL Environment on ML Experiment Integrity Auditing" — First Author | NeurIPS 2026 (under review) |
| 📝 Paper | "When LLMs Orchestrate but Do Not Compute" — First Author | SurgeLLM @ ACL 2026 |
| 🔒 IP | Provisional patent: streaming-native multi-agent treasury orchestration for asset allocation & cashflow forecasting | Filed |
| Achievement | Details |
|---|---|
| 🥇 Inter IIT Tech Meet 14 | Built a production-grade streaming treasury system; Sharpe ratio 7.78, ~92% RMSE reduction, 85% drawdown reduction |
| 🚀 OpenEnv Hackathon (Meta × PyTorch × HuggingFace) | Selected in top teams for offline Grand Finale, Bangalore, from 70,000+ registered developers |
| 📐 JEE Mains 2024 | 99.87 percentile out of 15+ lakh candidates |
| ✅ JEE Advanced 2024 | Qualified out of ~2.5 lakh candidates |
🔍 CLIPSCOPE — Mechanistic Interpretability for Foundation Models
Built a mechanistic interpretability pipeline for Vision-Language Models, discovering latent concepts in CLIP via sparse feature learning.
- Trained a 6.3M-parameter Sparse Autoencoder on 160K CLIP ViT-B/16 activations → 89% explained variance
- Demonstrated feature steering: interventions on learned latent directions produced up to 24.9% probability shifts
- Implemented Anthropic-inspired SAE: Top-K sparsity, ghost-gradient feature recovery & decoder normalization
PyTorch Transformers Sparse Autoencoders CLIP Mechanistic Interpretability
📈 Options Pricing Engine
Full derivatives pricing library spanning closed-form, stochastic, simulation, and numerical PDE methods.
- Heston pricing via Albrecher little-trap & Laguerre-quadrature Fourier inversion using live market data
- Implied volatility via Newton-Raphson/Brent; synthetic Heston parameters recovered within 10% error
- Constructed live SPY vol surfaces from 859 contracts across 5 expiries; analyzed term-structure dynamics
Python NumPy SciPy Black-Scholes Heston Monte Carlo Finite Differences
📊 Limit Order Book Simulator & RL Market Maker
Event-driven LOB simulator with price-time-priority matching and a learned market-making agent.
- PPO agent from scratch (GAE + clipped objective) inside a custom OpenAI Gym environment
- Implemented Avellaneda-Stoikov framework from stochastic-control theory with inventory-aware quoting
- Microstructure analytics: adverse-selection cost, fill-rate, inventory utilization & Kyle Lambda
- Order-flow generation via Hawkes process
Python PPO Reinforcement Learning Market Microstructure Gym
📉 Statistical Arbitrage Research Platform
Pairs-trading research platform with rigorous statistical methodology and realistic execution.
- Engle-Granger & Johansen cointegration with Kalman-filter dynamic hedge ratios
- Walk-forward pipeline with zero-lookahead, block-bootstrap Sharpe intervals & transaction-cost-aware execution
- Evaluated on 992 NSE trading days: 60 trades, OOS Sharpe 0.226, avg holding period 2.1 days
- LOB adapter for realistic fill simulation and adverse-selection analysis
Python Kalman Filter Cointegration Pairs Trading NSE
🤖 Project GRASP — Intelligent Robotic Arm (iBot Club, IITM)
Deploying an intelligent robotic arm to autonomously operate via visual scene understanding and natural language.
- Conditioned ACT's CVAE encoder on ResNet18 visual obs → +40% task success over baseline
- In-context learning for real-robot control: formatted teleoperation demos as LLM few-shot prompts
- Fixed gradient collapse in VLA models via cross-attention pooling → 85% reduction in cross-instruction correlation
- Open-vocabulary detection + geometric grasp planning with OpenCV & SayCAN → 90% success in sim
PyTorch ACT ResNet18 SayCAN OpenCV LLM Robotics
🧠 Project NEUROSPIKE — Spiking Neural Networks (BT Club × Prof. Gopalakrishnan)
Biologically plausible alternative to backpropagation using Predictive Coding networks.
- 2-layer hierarchical Predictive Coding network in Brian2 modelling Free Energy Principle update dynamics
- 96.9% test accuracy, 0.9999 cosine similarity vs. standard backprop
- On 5-task split-MNIST: PC-native update rule forgot 68% less than naive backprop fine-tuning
- Demonstrated generative inference ("dreaming") in < 300 steps by clamping output neurons to MNIST targets
Brian2 Spiking Neural Networks Predictive Coding Neuromorphic AI
🏦 Inter IIT Tech Meet 14 — Streaming Treasury System (Pathway)
Production-grade real-time treasury orchestration system for banks.
- Autoregressive CashFlow forecasting with Gaussian HMM regime detection → ~92% RMSE reduction
- Fundamental Analysis Agent → overall Sharpe ratio of 7.78 with 85% drawdown reduction
- Real-time yield forecasting via Nelson-Siegel yield-curve modeling on NSE G-Sec data over Kafka
- Bond pricing from first principles: YTM, duration, convexity & dirty-price
- 18 custom Pathway operators across bonds, equities, and forex feeds
Python Kafka Pathway HMM Nelson-Siegel RL LLM
🎮 OpenEnv Hackathon — RL Environment for LLM Project Managers (Meta × PyTorch × HuggingFace)
First-of-its-kind RL environment training LLM project managers under deception and long-horizon software crises.
- Trained LoRA-GRPO policies on Qwen-1.5B evaluating strategic cross-verification under falsified observations
- Adaptive multi-agent dynamics: deceptive LLM teammates, social testimony graphs & crisis escalation
RL LoRA GRPO Qwen Multi-Agent LLM
| Role | Org | Highlights |
|---|---|---|
| Coordinator, Head of CV & RL | iBot Club, IITM | Mentored 500+ students; led 4-week CV bootcamp; designed RL locomotion project with 45-member team |
| Open-Source Contributor | Neural-LAM | Probabilistic forecasting — Graph-EFM integration; ensemble calibration diagnostics |
| Teaching Assistant | GN1002, IITM | Mentored 15-20 freshmen on personal development & time management |
| Saathi Peer Mentor | IITM | Guided freshmen through academics and institute life |
I'm always open to collaborating on research in robotics, interpretability, RL, or quant systems — feel free to reach out!
