Skip to content
View Amanatar's full-sized avatar

Highlights

  • Pro

Block or report Amanatar

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Amanatar/README.md

Aman Atar

Kaggle Expert LinkedIn MScFE IIT Hyderabad

Hey, I'm Aman.

I work on quantitative finance models and production ML systems -- mostly focused on what happens when clean math meets noisy market regimes, jump risk, and strict latency budgets.

Currently doing my MSc in Financial Engineering at WorldQuant University (plus AI/ML coursework with IIT Hyderabad). Outside research, I test ideas on Kaggle as a Kaggle Notebooks Expert -- currently ranked #30 worldwide out of 58,700+ ranked notebooks practitioners (peak #28, 40 medals: 4 Gold, 5 Silver, 31 Bronze).

Featured Systems & Code (Original Flagship Repositories)

System Focus & Implementation Details Stack
auditable-regime-aware-agentic-forecasting
TRACE-Fin		 Timestamp-restricted regime-aware return forecasting & conformal risk control. Replay across 1,772 out-of-sample trading days (2020-2026, SPY, QQQ, ^NSEI, TLT, GLD; net of 10 bps one-way costs) delivers 1.189 Net Sharpe (vs. 0.963 Buy & Hold and -1.075 Uncalibrated Point Forecast), 1.680 Sortino, 44.9% lower Max Drawdown (11.76% vs. 21.34%), and 4.19x turnover (vs. 121.86x). Regime-normalized ACI lifts 90% stress coverage to 93.60% in High-Volatility (vs. 84.14% static conformal; 89.81% overall) with 0 / 1,772 causal violations (100.0% leak detection). Python, scikit-learn, PyArrow, Conformal Prediction, Regime Detection
neural-bates-dml
Sobolev Pricing Engine
GitHub ↗ · Live Workbench ↗		 Twin-network Sobolev Differential ML surrogate for 10-D Bates SVJD option pricing & calibration. Evaluated out-of-sample (N_test=2,048 at N=8,192): reduces Delta RMSE by 36.0% (0.0150 -> 0.0096), Vega RMSE by 23.8%, and Price RMSE by 30.2% (0.00107 -> 0.00074) vs. Price-Only MLP (1.544 ms 128-batch CPU, 0.51 MB checkpoint). Achieves 7.6x-15.5x faster 5-regime surface calibration (0.43 s - 0.66 s PyTorch L-BFGS vs. 5.00-7.90 s FFT; 25.6-39.3 bps stress IV RMSE) and 6.07 s live SPY calibration (97.9% arbitrage pass rate). PyTorch, Sobolev DML, Carr-Madan FFT, Bates SVJD, L-BFGS
FinFaithful-RAG
Grounded Financial RAG & Benchmark		 SEC EDGAR 10-K grounded retrieval & verification engine. Benchmarks 4 retrieval pipelines across a 480-evaluation live benchmark over 120 audited Form 10-K questions (AAPL, MSFT, NVDA, JPM, GOOGL x 4 tiers). Hybrid RRF + Cross-Encoder lifts HitRate@5 to 72.2% (vs. 54.4% BM25, 44.4% FAISS), doubling Cross-Company comparative recall from 33.3% to 66.7%. Grounded Faithfulness increases from 2.66 / 5.0 (ungrounded baseline, 0.0% coverage) to 4.12 / 5.0 (89.2% citation validity) with 100.0% adversarial trap abstention (eliminating 26.7% baseline hallucinations). Python, FAISS, Sentence-Transformers, Cross-Encoder, BM25, Streamlit, LLM-as-a-Judge
IRL_Data_Science_-_GenAi_Internship
Production GenAI & MLOps		 Production GenAI & MLOps suite. Includes FinWiser (financial assistant with Gemini -> NVIDIA NIM LLaMA-3.1-70B 429 circuit-breaker failover on AWS EC2), enterprise RAG via Amazon Bedrock, and centralized MLflow tracking across 5 classification baselines. Python, AWS EC2, Bedrock RAG, MLflow, Flask, Docker
google-ai-agents-intensive
The Alpha Seeker		 The Alpha Seeker autonomous multi-agent trading loop built on Google ADK. Coordinates sequential, parallel, and feedback-loop agents for macro data ingestion, signal evaluation, and max drawdown (<15%) enforcement. Python, Google ADK, Multi-Agent LLMs, Technical Analysis
hypertension-risk-assessment
Clinical Risk Classifier		 4-stage clinical cardiovascular risk classifier & Flask inference service. Maps multi-factor telemetry into clinical risk stages (Normal, Stage 1, Stage 2, Crisis) with calibrated probability confidence and automated medical triage recommendations (pytest CI verified). Python, Flask, scikit-learn, Telemetry Triage

Applied Engineering Repositories

  • practical-deep-learning-for-coders: Deep learning vision architectures, transfer learning pipelines, and tabular neural networks implemented via Fast.ai and PyTorch (8 notebooks & forward/backward pass verified).
  • student-mental-health-companion: Affective state computing assistant combining VADER sentiment analysis, custom emotional lexicon mapping, and dual-mode Gemini/offline dialogue generation with immediate crisis routing.
  • machine-learning-project: Catalog of 20 supervised, unsupervised, and deep learning implementations across clinical healthcare, finance, NLP, and signal processing (80.3% test accuracy on clinical cardiovascular telemetry).
  • Project_euler: Vectorized and dynamic programming solutions to Project Euler problems (1-26) in pure Python (100% mathematical match verified).

Competitive ML on Kaggle (Kaggle Expert)

  • Official Standing: Kaggle Notebooks Expert -- #30 Worldwide (Peak #28) out of 58,736+ ranked practitioners (Top 0.05%).
  • Medal Count: 40 Medals (4 Gold, 5 Silver, 31 Bronze) with 2,800+ community upvotes across 800+ published research notebooks.
  • Notable Gold Notebooks:
    • ans-verifys (749+ upvotes, Gold) -- Reasoning verification and answer integrity pipelines for competition LLM outputs.
    • improved-metric-hack-last-call (196+ upvotes, Gold) -- Post-processing and threshold optimization under non-standard competition loss metrics.
    • nvidia-nemotron-sfttrainer-training (176+ upvotes, Gold) -- Supervised fine-tuning pipeline for Nemotron models using HuggingFace TRL.
    • biohub-geometric-fusion (172+ upvotes, Gold) -- Graph and tabular geometric feature fusion architecture.

Technical Focus & Tools

  • Quant & Modeling: Python, PyTorch, LightGBM, XGBoost, CatBoost, Differential ML, Stochastic Calculus, Carr-Madan FFT
  • GenAI & Infrastructure: FAISS, Sentence-Transformers, Google ADK, Amazon Bedrock RAG, Gemini API, NVIDIA NIM, MLflow, FastAPI, Flask, Docker, AWS EC2
  • Data & Tools: NumPy, Pandas, SciPy, OpenCV, PostgreSQL, Linux / Bash, Git

Contact

Pinned Loading

  1. auditable-regime-aware-agentic-forecasting auditable-regime-aware-agentic-forecasting Public

    TRACE-Fin: Regime-aware return forecasting with strict point-in-time cutoffs and conformal prediction intervals.

    Python

  2. FinFaithful-RAG FinFaithful-RAG Public

    SEC 10-K retrieval & faithfulness benchmark comparing TF-IDF, dense FAISS, and hybrid RAG with an LLM judge.

    Python

  3. google-ai-agents-intensive google-ai-agents-intensive Public

    The Alpha Seeker: Autonomous multi-agent trading system built with Google Agent Development Kit (ADK) & Kaggle

    Jupyter Notebook

  4. hypertension-risk-assessment hypertension-risk-assessment Public

    End-to-end clinical hypertension risk prediction system with multi-stage ML classification and Flask UI

    Python

  5. neural-bates-dml neural-bates-dml Public

    Differential Machine Learning (Sobolev training) neural surrogate for Bates SVJD model calibration & FFT pricer

    Python

  6. student-mental-health-companion student-mental-health-companion Public

    Empathetic mental health conversational AI with VADER sentiment tone-matching and multi-turn state memory

    Python