Data Scientist | ML Engineer & LLM’s | RAG · Fine-Tuning · AI Agents | MS Data Science @ UMD

github

Jeevan Sai Reddy C.  • Peerlist

Built an end-to-end Clinical Decision Support System that transforms raw, unstructured medical notes into actionable intelligence. By combining domain-specific BERT architectures with advanced Generative AI, this pipeline acts as an automated clinical assistant capable of diagnosing, routing, and recommending treatments based on complex patient histories. 1. Processed 4,000+ MTSamples transcriptions across 13 medical specialties into structured intelligence. 2. Fine-tuned BioClinicalBERT (on i2b2/n2c2 corpora) for Named Entity Recognition (NER) to extract medical problems, tests, and administered treatments with an F1 score > 0.87. 3. Deployed a PubMedBERT-based multi-class classifier achieving 93% accuracy in routing clinical notes to the correct specialty. 4. Orchestrated Google's Gemma3-4B via local Ollama REST APIs, enforcing strict JSON schemas to perform advanced clinical reasoning and infer comprehensive, step-by-step treatment plans. 5. Built an interactive Streamlit UI, implementing proactive GPU memory management (VRAM flushing) to run this massive multi-stage transformer pipeline smoothly on local hardware.

Generative AI Clinical Decision Support Pipel

1. Fine-tuned Llama-3-8B into a domain-specific data science code assistant using supervised fine-tuning (SFT); built the full training-data pipeline, curating, cleaning, and ChatML-formatting a 50K-example instruction corpus. 2. Built a QLoRA training pipeline (Hugging Face PEFT, 4-bit quantization) that cut GPU memory use by ∼75%, making full training runs feasible on a single consumer GPU. 3. Built a held-out evaluation harness benchmarking fine-tuned outputs against base Llama-3-8B on SQL and pandas code-generation tasks, scoring with ROUGE-L and BLEU. 4. Reduced catastrophic forgetting by mixing general-instruction regularization data into the training set, preserving base-model performance on non-domain tasks.

DataAdapt: Fine-Tuned Data Science Code Assis

Architected a highly optimized, enterprise-grade Retrieval-Augmented Generation (RAG) system capable of parsing and reasoning over complex, 200+ page research documents. By moving beyond naive text splitting and implementing semantic-aware retrieval, this system dramatically improved context precision and generative faithfulness. 1. Engineered a multi-stage ingestion pipeline using LlamaParse to extract and preserve complex Markdown tables and layouts from massive PDFs. 2. Replaced rigid token-window splitting with dynamic Semantic Chunking (via Gemma-3 embeddings) to partition documents by contextual shifts. 3. Slashed vector search latency and storage by 3x using Matryoshka Representation Learning (MRL), truncating dense vectors to 256 dimensions with no semantic loss. 4. Deployed a Dockerized Weaviate v4 vector database supporting Hybrid Search, Multi-Query Expansion, and Cross-Encoder Reranking. 5. Integrated DeepSeek-R1 (via Ollama) as the generative reasoning engine to synthesize reranked context into heavily grounded, source-cited responses. 6. Validated architecture via RAGAS, achieving 0.98 Context Recall, 0.93 Context Precision, and 0.72 Faithfulness.

DeepRAG: An Optimized RAG System for Research

1. Designed a data-driven framework to optimize inventory policies (s, S policies) and minimize total supply chain costs, including holding, procurement, and stockout penalties. 2. Implemented Gradient Boosting (XGBoost/LightGBM) and Prophet models to generate robust, multi-horizon demand forecasts, mitigating the bullwhip effect. 3. Developed a comprehensive cost function and used Monte Carlo simulation to assess the impact of fluctuating lead times and demand volatility on operational expenditure. 4. Determined optimal reorder points and order quantities to meet a target service level while significantly reducing holding and shortage costs.

Supply Chain and Inventory Optimization

1. Developed a validated GARCH(1, 1) model on 14 years of daily log returns (2010–2024) to forecast conditional volatility. 2. Discovered and quantified high volatility clustering (persistence alpha1 + beta1 = 0.9736) using the ACF/PACF of squared returns (ARCH effects test). 3. Determined the Long-Run Volatility (sigma_LR = 1.66%) and forecasted immediate risk, revealing the asset was in a low-volatility state (sigma_t+1 = 1.33%) with a slow mean-reversion trend. 4. Validated model robustness using diagnostic checks on Standardized Residuals, confirming the model was correctly specified and the errors were white noise.

Financial Volatility Modeling and Risk Foreca

Designed a unified, algorithmic marketing analytics architecture that merges top-down market mix modeling with bottom-up multi-touch attribution. This engine allows enterprises to mathematically prove ROI across advertising channels and dynamically redistribute budgets to maximize Return on Ad Spend (ROAS). 1. Integrated macro-level Google Merchandise Store spend logs with micro-level Criteo touchpoint sequences (2.5GB+ dataset). 2. Architected a highly tuned Marketing Mix Model (MMM) using Meta's Robyn framework, leveraging geometric adstock transformations and Hill functions to quantify lagged ad effects and diminishing returns. 3. Developed robust Multi-Touch Attribution (MTA) models, benchmarking Markov Chains and Shapley values against traditional heuristic rules to properly allocate conversion credit. 4. Designed a mathematical Reconciliation Engine utilizing non-linear solvers to fuse MMM budget bounds with MTA crediting, maximizing global Marginal ROAS. 5. Evaluated model precision using NRMSE, MAPE, and algorithmic fairness metrics against naive last-click baselines. 6. Deployed a full-stack Streamlit dashboard to visualize ROI saturation curves, touchpoint heatmaps, and actionable budget reallocation strategies.

Algorithmic Marketing Attribution & ROAS Opti

1. Develop machine learning models to predict race outcomes by analyzing driver performance, constructor data, circuit characteristics, and weather conditions. 2. Analyze the influence of starting grid positions on race results, with a focus on their correlation with race victories. 3. Assess the consistency of top drivers and constructors, identifying key patterns that shape performance across seasons. 4 Study accident rates across circuits, highlighting high-risk tracks such as Monaco and Spa-Francorchamps due to their complex layouts

Formula 1 Championship (1950 - 2024)

1. Built and validated an ARIMA (AutoRegressive Integrated Moving Average) model to forecast a publicly available economic indicator with trend and seasonality. 2. Applied the Box-Jenkins methodology, utilizing ADF testing for stationarity and ACF/PACF plots to identify and achieve an optimal ARIMA specification. 3. Validated model integrity by refining parameters using AIC/BIC criteria and confirming the final residuals were White Noise via the Ljung-Box test and checks for Heteroscedasticity. 4. Achieved a measurable improvement in forecast accuracy (measured by RMSE) over baseline historical averaging methods.

Univariate Time Series Forecasting using ARIM

Collaborated in a four-person team to select and integrate the Image dataset (1k training, 200 validation, and 200 test images with various image sizes) for real-time analysis ◦ Designed and implemented advanced machine learning algorithms to detect and correct color cast issues in images, enhancing overall color fidelity and visual accuracy. ◦ Attained substantial improvements in image fidelity, as evidenced by quantitative metrics such as color accuracy and perceptual quality assessments, outperforming traditional correction methods.

Breaking through color cast: Enhanced image F

1. Built an agentic AutoML system that takes any raw CSV and runs end to end with zero manual code changes: LLM-based dataset profiling, automated preprocessing, and a model tournament across 10+ algorithms (XGBoost, LightGBM, linear baselines). 2. Prevented data leakage, overfitting, and underfitting via three guard gates in global preprocessing (including a two-phase correlation guard) and a Pydantic-validated schema layer that adapts the pipeline to unseen datasets. 3. Implemented LLM-driven diagnostics and reporting: LLaVA (multimodal) inspects OLS residual plots for non-linearity and Llama 3 interprets Macro-F1 deltas to generate the final model summary report. 4. Sandboxed all LLM-generated code with a self-correction loop (up to 3 retries) and a critic validator node that checks the winning model for leakage, overfitting, and underfitting, surfacing its findings in the user-facing summary. 5. Tested across 5 datasets (synthetic and real, spanning classification and regression) with zero code changes; the preprocessing gates held in every run and the critic validator never flagged leakage or overfitting.