Aman Atar · Maharashtra, IN (IST Active)

“The hidden harmony is stronger than the obvious.” — Heraclitus, Fragment B54

Looking for quiet structure inside noisy data — across quantitative finance, empirical machine learning, and production AI.

I'm Aman Atar. I started in commerce and computer applications (BCA) in Solapur and worked my way into quantitative financial engineering — learning early that clean equations only matter if they survive regime shifts, transaction costs, and a fresh Linux machine.

Kaggle Notebooks Expert #30 Worldwide · Top 0.05% (40 Medals)
WorldQuant University MSc in Financial Engineering (MScFE)
WorldQuant BRAIN Gold Level · Alpha Research
GitHub Engineering 10 / 10 Original Repos · Linux CI Passing
Drag toward 100% to strip random noise and reveal the latent curve.
01 · Thinking Principles

Three Notes on Modeling Reality Honestly

Click any note to expand. No buzzwords — just the working rules that guide how I write code and test hypotheses.

Kierkegaard noticed something every quantitative researcher eventually learns the hard way: looking backward makes chaos look neat and predictable. In a Jupyter notebook, you already have the entire historical CSV sitting in memory. If a rolling average, scaler, or target label accidentally peeks even one day ahead (t + 1), your backtest looks effortless.

Real markets don't hand you tomorrow's row. Treating strict timestamp cutoffs as a non-negotiable rule — where the model at time t is completely blind to anything after t — is the difference between honest research and fooling yourself.

The central idea in Marcus Aurelius's Meditations is separating what belongs to your judgment from external events you don't control. That maps cleanly onto engineering. You cannot stop volatility from spiking, correlations from breaking, or an external LLM API from throwing HTTP 429 rate-limit errors under load.

What you can control is how your system behaves when calm weather ends: modeling jump risk explicitly (Bates SVJD rather than smooth Gaussian assumptions alone), outputting calibrated uncertainty intervals instead of naked point guesses, and wiring fallback circuit breakers before shipping to users.

In Thinking, Fast and Slow, Kahneman shows how easily humans mistake a fluent, confident story for a true one. Large language models have the exact same flaw: ask an ungrounded model for a revenue number from an SEC 10-K filing and it will state a fabricated figure in polished prose (scoring 1.67 / 5.0 on faithfulness in my baseline tests).

Whether in financial Q&A, competition post-processing, or option calibration, fluency means nothing without verification — grounding answers in retrieved primary filings, checking analytical sensitivities, and testing code on a clean machine from scratch.

02 · Interactive Quant Intuition Lab

Visualizing Causal Boundaries & Volatility Surfaces

Hands-on numerical experiments: contrast honest point-in-time filtration against lookahead leakage, then inspect how jump-diffusion dynamics shape implied volatility smiles.

Hover over canvas to inspect step-by-step point-in-time metrics
In-Sample Fit (Before Cutoff) MAE 0.62 · Sharpe 1.45
Live Out-of-Sample (Post Cutoff) MAE 1.18 · Sharpe 1.24
90% Interval Coverage (OOS) 91% inside band
Validation Takeaway Realistic fit; band widens with shock
03 · Verified Systems & Benchmarks

Selected Engineering Work

100% original repositories — dependency-pinned and tested automatically on fresh Ubuntu 24.04 GitHub Actions runners.

01 · Quant & Time-Series GitHub ↗

TRACE-Fin: Point-in-Time Regime-Aware Forecasting

Return-forecasting research framework built around strict timestamp cutoff checks (zero lookahead leakage), regime-shift detection, transaction-cost utility, and calibrated 90% (α = 0.10) conformal prediction intervals.

PythonPyTorchscikit-learnConformal PredictionUbuntu CI Pass
02 · Option Pricing & Stochastic Models

Neural Bates DML: Sobolev Option Pricing Surrogate

Twin-network PyTorch surrogate for the 10-parameter Bates stochastic volatility jump-diffusion (SVJD) model. Supervised on analytical prices and Greeks via Sobolev loss, achieving 1.58× faster SPY calibration (6.07 s vs. 9.61 s FFT), 10.3× per-surface speedup (3.26 ms CPU), 340 bps IV RMSE, and 40.1% lower Delta error.

PyTorchDifferential MLCarr-Madan FFTBates SVJD1.58× Calib Speedup10.3× CPU Surface340 bps IV RMSE
03 · Financial NLP & Evaluation GitHub ↗

FinFaithful-RAG: SEC 10-K Retrieval & Faithfulness Benchmark

End-to-end retrieval and LLM-as-a-Judge evaluation pipeline over SEC EDGAR 10-K filings. Compares TF-IDF, dense FAISS (~8 ms vector lookup), and hybrid retrieval via NVIDIA NIM (Llama 3.1 70B), lifting answer faithfulness from 1.67 / 5.0 (raw LLM) to 5.00 / 5.0.

FAISSSentence-TransformersSEC EDGARNVIDIA NIMLLM-as-a-Judge
04 · Production GenAI & MLOps Internship Work GitHub ↗

FinWiser Cloud Failover, Bedrock RAG & MLflow Suite

Built as Data Science with GenAI Intern at Innomatics Research Labs: deployed FinWiser on AWS EC2 (Ubuntu + Gunicorn) with live circuit-breaker failover from Gemini API (HTTP 429) to NVIDIA NIM, plus Amazon Bedrock RAG and 5-model MLflow sentiment tracking. Also shipped clinical & mental-health ML apps during internships as Machine Learning Engineer Intern at TheSmartBridge and AI & Cloud Technology Intern at Edunet Foundation.

AWS EC2Amazon BedrockMLflowFlaskStreamlit
05 · Multi-Agent Workflows GitHub ↗

The Alpha Seeker: Multi-Agent Trading Workflow (Google ADK)

Built during the 5-Day AI Agents Intensive with Google: coordinates sequential and loop agents for market signal ingestion, thesis critique, and drawdown-gated risk checks.

Google ADKMulti-Agent SystemsPythonRisk Gating
06 · Competitive ML & Alpha Research Kaggle Profile ↗

Kaggle Notebooks Expert (#30 Worldwide) & WorldQuant BRAIN Gold

Ranked #30 of 58,650+ globally on Kaggle (Peak #28 · 4 Gold, 5 Silver, 31 Bronze medals · 2,800+ upvotes) alongside Gold Level in the WorldQuant BRAIN alpha challenge. Selected Gold-medal notebooks:

  • ans-verifys — Reasoning verification harness (Gold · 749+ upvotes) Empirical Pattern: Test-time reasoning verification & self-consistency voting
  • improved-metric-hack-last-call — Metric-aware threshold optimization (Gold · 196+ upvotes) Empirical Pattern: Metric-aligned threshold search & post-processing optimization
  • nvidia-nemotron-sfttrainer-training — Nemotron SFT pipeline (Gold · 176+ upvotes) Empirical Pattern: TRL supervised fine-tuning & memory-efficient LoRA
  • biohub-geometric-fusion — Geometric feature fusion (Gold · 172+ upvotes) Empirical Pattern: Non-linear geometric feature fusion across multimodal embeddings
04 · Honest Status Ledger

What I'm Doing Right Now (November 2026)

No inflated roadmap — just what is actually on my desk this month and what I'm aiming for next.

Studying
MSc in Financial Engineering (MScFE) — WorldQuant University

Working through graduate coursework in derivatives pricing, financial econometrics, stochastic modeling, and collaborative group projects, alongside AI & drone navigation coursework with IIT Hyderabad.

Competing
Kaggle Competitions & Pushing for the 10th Silver+ Medal

Currently holding 4 Gold and 5 Silver medals in Kaggle Notebooks (#30 Worldwide) — one Silver medal short of the 10 Silver+ threshold for Kaggle Notebooks Master — while competing in tabular fraud and scientific ML benchmarks.

Researching
Cross-Sectional Equity Alphas & Multi-Timeframe Pipelines

Testing low-turnover alpha signals on WorldQuant BRAIN (Gold level) and experimenting with time-series foundation models + LightGBM stacking for Indian (NSE) and US equities.

Open To
Quant Research, ML Engineering & Applied AI Roles

Looking for teams where mathematical honesty, leak-free validation, and shipping clean Linux-tested code matter more than buzzwords.

05 · Off the Clock

Bookshelf & Quiet Interests

Books I actually keep on my desk and problems I enjoy working through when no deadline is attached.

Philosophy & Stoicism

Meditations — Marcus Aurelius

A grounded reminder to stay calm when experiments fail, focus strictly on your own craft, and drop ego when data proves your favorite hypothesis wrong.

Judgment & Bias

Thinking, Fast and Slow — Daniel Kahneman

Why human intuition misreads randomness, overfits small samples, and confuses a coherent story with statistical probability.

History & Execution

Sapiens & The Hard Thing About Hard Things

Harari on how shared systems and markets evolved; Ben Horowitz on staying clear-headed when real-world execution gets messy.

Pure Problem Solving

Project Euler in Python ↗

Solving number-theory and combinatorics problems cleanly without brute-force shortcuts — keeping algorithmic thinking sharp outside frameworks.