Rishit
Belani

Quantitative Research Intern

• Replaced an incumbent XGBoost approach in FlexTrade's FlexTCA execution-analytics workflow with an Almgren–Chriss–based pre- and post-trade cost model across intraday and multi-day horizons, delivering 80% faster inference at comparable accuracy.
• Improved cost-forecast WinR² by 60% on high-cost tail executions with tail-lift model variants, validated by cross-validation on strictly pre-trade inputs to eliminate post-trade leakage.
• Engineered a compliance RAG pipeline over 30K+ pages of buy-side documents (hierarchical parsing, hybrid BM25/vector retrieval, cross-encoder reranking), reaching 94% critical-issue recall and cutting first-pass review time by 80%.
• Mentored undergraduate interns on order-fill probability modeling, guiding research design and validation.

Quantitative Researcher

• Developed a joint PD/LGD/EAD modeling framework using LSTM, XGBoost, MLP and Transformer models; achieved 0.96 ROC-AUC for default-risk discrimination and 21% fewer missed defaults at a fixed precision operating point in evaluation.
• Prevented regime-driven model decay with a dual feature-selection pipeline: mutual-information scoring for predictive power plus a concept-drift index, removing 96% of candidate features that degraded out-of-cycle.

Software Development Engineer

• Built an LLM-assisted maintenance and incident-resolution system for critical smart-building environments, using historical incident logs and procedural manuals and reducing mean time to resolution by 30% across recurring faults.
• Engineered a modular RAG architecture indexing 10K+ documents in a hybrid relational/vector store (retrieval P95 < 800 ms) while contributing to smart-building software for large-scale data-center and hospital deployments.

Data Science Intern · E-commerce

• Worked on customer churn prediction, customer segmentation, and retail-data processing to support e-commerce analytics and retention decisions.
• Published a related customer lifetime value demonstration using dummy/sample retail data: estimated six-month purchase frequency with BG/NBD, purchase value with Gamma-Gamma, and customer segments.

The linked CLTV notebook uses dummy/sample data, not actual StreetStyle customer records; it is a separate demonstration, not evidence of production deployment or measured business lift.

Quantitative research · Statistical ML · Python / LightGBM

+0.0131 mean walk-forward rank IC across 4,943 trading dates; 95% block-bootstrap CI [+0.0083, +0.0178].

• Developed cross-sectional momentum and short-horizon reversal signals from daily equity OHLCV data, using lagged liquidity to select up to 1,000 eligible names per day and explicit checks against time-series leakage.
• Compared an interpretable rank blend with conditioned Ridge and gradient boosting on matched out-of-sample dates. The boosted model improved raw rank IC from 0.0124 to 0.0173, but its tested portfolio gross Sharpe was lower (0.141 versus 0.566).
• Investigated overnight-versus-intraday signal concentration, turnover, risk exposures, and execution-cost budgets, demonstrating why predictive accuracy is not the same as tradable alpha.

AI engineering · Model evaluation · Python

• Built a four-agent system for CaseHOLD legal reasoning with a graph-based evidence retriever and a structured Plaintiff–Defense–Judge debate.
• Reported a 16-percentage-point accuracy improvement over a vector-RAG baseline (p = 0.025); evaluated architectures using controlled ablations, paired bootstrap intervals, McNemar’s test, and retrieval-quality metrics.
• Designed a Streamlit interface for inspecting agent reasoning and retrieved evidence, with tracking for token cost and latency.

Data science · Applied ML · NLP

• Developed a churn-scoring workflow using time-aware customer features, XGBoost, SHAP explanations, and automated business insights.
• Combined structured purchase histories with NLP-based complaint classification in a Streamlit dashboard with natural-language business Q&A.
• Evaluated on synthetic customer data; a system-design demonstration rather than evidence of real-world predictive lift.

Quantitative data science · Backtesting · Streamlit

• Co-developed an interactive strategy-research dashboard for FX, equities, commodities, and other instruments using dual-exponential-smoothing signals, historical market data, and position-aware backtesting.
• Implemented cost-aware trade evaluation, parameter grid searches, performance diagnostics, and an ML signal module with engineered features, Random Forest, and optional SHAP analysis.

Applied Data Science · Algorithmic Trading · Databases · Modern Data Structures

• Wrote problem sets and led instruction on backtesting, execution strategy construction and market microstructure for a graduate Applied Data Science and Algorithmic Trading course, supporting 100+ students across a five-TA team.
• Maintained related quantitative modeling notebooks exploring MLP, LSTM, and graph neural networks for cross-sectional equity ranking, walk-forward evaluation, and long/short research.

Related independent research shared in the context of my Algorithmic Trading teaching; the repository is not represented as official Columbia course material.

Columbia University

• Constructed a 24-month, multi-million-dollar AI roadmap and ROI model for a B2C e-commerce client.