Srija Reddy Kethireddy

Session-Aware Recommendation Agent for User Interest Drift

Northwestern University · 2026

• Built a self-adaptive meta-controller on KuaiRand-Pure (1.44M interactions) with closed-loop Observe → Diagnose → Adjust → Recommend → Evaluate pipeline.
• Meta-agent outperformed all fixed baselines: Hit@5 0.3088 vs 0.3047, unique top-1 items 854 vs 615.
• Trained multi-behavior user response simulator and verified DDPG whole-session RL baseline in KuaiSim.

NL2SQL: Natural Language to SQL Generation

Northwestern University · May – Jun 2026

• Fine-tuned T5-Base (220M params) on WikiSQL (56K examples) achieving 73.76% validation exact match.
• Built FAISS-indexed RAG with Sentence-BERT for few-shot example injection; ran Hyperband search via Optuna (12 trials).
• Deployed as Streamlit chatbot — users upload CSV/SQLite, schema auto-parsed, SQL generated and executed on live data.

AI-Assisted Medical Trainee Application Evaluation

Feinberg School of Medicine, Northwestern · Apr 2026 – Present

• Building residency application screening pipeline: PDF parsing → structured fact extraction → rubric-aligned Excel scorecards.
• Privacy-first hybrid LLM scoring via local Ollama models — no applicant data leaves the institution.
• Deterministic Python scoring layered on LLM outputs for auditable, consistent decisions.

MarineGAN: Deep Convolutional GAN for Image Generation

Northwestern University · Apr – May 2026

• Trained DCGAN on 3,983 marine animal images (128×128); Generator 3.5M params, Discriminator 2.8M params.
• Grid search over 9 hyperparameter combos tracked via Weights & Biases — optimal: lr_d=0.0002, noise=0.1, 80 epochs.
• Documented failure modes: mode collapse after epoch 100, discriminator over-regularization, early stopping misapplication.

Constraint-Aware Multi-Agent Meal Planning System

Northwestern University · Jan – Mar 2026

• Safety-critical multi-agent system for dietary meal planning under strict constraints — zero unsafe recommendations across all test cases.
• Ingredient-level allergen detection, automated recipe repair, and re-validation after each repair cycle.
• Transparent reasoning traces; agent triggers automatic repair on constraint violation.

BookTunes: Emotion-Driven Music Generation

Northwestern University · Sep – Dec 2025

• Dual-path neural network (CNN + acoustic fusion, 1.2M params) for cross-modal emotion classification from text.
• Text-to-music retrieval pipeline using sentence transformers and FAISS to blend emotionally compatible tracks.

NBA Injury Risk Modelling

Northwestern University · Sep – Dec 2025

• Integrated 780,000+ player-game records and 18,000 injury events across 23 NBA seasons; engineered 34 features.
• XGBoost with SHAP interpretability — age, cumulative minutes, and prior injury history as primary risk drivers.

• Building agentic AI workflows for a procurement automation startup targeting SMBs.
• Integrating with ERPs, Excel, and email to automate inventory analysis, purchase order generation, and supplier communication.
• Reporting directly to the CTO.

• Built a document analysis pipeline processing 10,000+ aviation XML files; applied OCR to detect and label aircraft part figures with 95% accuracy, reducing manual review by 20+ hours/week.
• Automated PDF annotation and XML data mapping end-to-end with validation and quality checks across heterogeneous document formats.

• Fine-tuned CodeT5 on 1M+ code snippets with Hugging Face Transformers; deployed as a production REST API with 40% gains in retrieval and clustering efficiency.
• Designed and shipped a production RAG application using Python, Neo4j, and REST APIs — cut query response latency by 60%.
• Automated GitHub dependency parsing across 8 languages; collaborated cross-functionally with product and engineering teams.

• Trained an ANN-based binary classifier on 100,000 seismic samples achieving 90% classification accuracy using TensorFlow/Keras.
• Leveraged AWS for large-scale dataset management and scalable experiment pipelines.