Models

Three systems,
one thesis.

Markets need pricing. Lenders need scoring. Operators need forecasting. Built all three — each one shipped, not just notebooked.

FILE 01 / DEC 2025 – JAN 2026
LiveRisk — Portfolio Risk Intelligence
Market Risk
  • Engineered a Monte Carlo risk engine simulating 10,000 correlated geometric Brownian motion paths via Cholesky decomposition, computing 95% VaR and CVaR with FinBERT sentiment-adjusted multipliers across 12 years of market data.
  • Shipped as a full-stack platform (FastAPI + Next.js) with five historical stress-test scenarios — 2008 crash, COVID, rate shock, dot-com, and the WSB short-squeeze — plus a 60-day LSTM portfolio forecast.
  • Built to answer the question every risk desk asks: "what happens to this book if the world breaks again?"
10,000
Simulated GBM Paths
5
Historical Stress Scenarios
60-day
LSTM Forecast Horizon
PythonFastAPINext.jsFinBERTLSTMMonte Carlo / CholeskyVaR / CVaR
FILE 02 / JAN – MAR 2026
CreditIQ — AI Financial Intelligence Platform
Credit Risk
  • Built a 5-layer deep neural network credit classifier trained on 10,000 records across 14 features, reaching 82.5% AUC-ROC, with SHAP explainability layered in for ECOA regulatory compliance and model interpretability.
  • Engineered a FinBERT NLP sentiment pipeline scoring live financial news at 87% accuracy, with a ChromaDB vector store and RAG layer for real-time market intelligence retrieval and report generation.
  • Developed a GradientBoosting IPO return predictor across nine signals — burn multiple, NRR, P/S ratio — fed by a live yfinance pipeline, surfacing high-return patterns from historical IPO data.
0.825
AUC-ROC, Credit Classifier
87%
Sentiment Pipeline Accuracy
14 / 9
Credit Features / IPO Signals
TensorFlowSHAPFinBERTChromaDB / RAGGradientBoostingyfinance
FILE 03 / JAN – FEB 2026
SolarSense — Predictive Maintenance Platform
Operational Risk · Finalist
  • Ran exploratory data analysis across 50+ operational parameters to detect degradation patterns, then deployed an XGBoost fault-prediction model with SHAP explainability to surface root-cause insights for maintenance teams.
  • Translated model output into recommendations that cut maintenance costs by 35% and lifted energy output by 8% — the kind of operating-leverage number that gets a model funded.
  • Presented findings at the AI for Sustainability Hackathon 2026 (Canadian University Dubai) — Finalist.
-35%
Maintenance Cost
+8%
Energy Output
50+
Parameters Analysed
XGBoostSHAPEDAPython