Sakshi Asati

Built a LangChain + GPT-4 documentation assistant with FAISS retrieval so clinicians review notes faster, while Streamlit dashboards, FastAPI services, Celery, and Redis coordinate authenticated workflows backed by PostgreSQL inside Docker.

• Clinical reviewers cut note synthesis time by 35% with contextual retrieval prompts.
• Role-based Streamlit workspaces and background Celery jobs deliver sub-2s response times.

Fine-tuned YOLOv8 on retail shelf images to detect and classify 34 product categories for automated inventory tracking. Built a FastAPI inference service and Streamlit dashboard to process images and return predictions in under 2 seconds, enabling near real-time stock checks during store operations.

• Automated inventory tracking across 34 product categories with computer vision.
• FastAPI inference service delivers predictions in under 2 seconds for real-time stock checks.

Delivered a full-stack CRM platform (MongoDB, Node.js, React) that centralizes support tickets and integrates an AI chatbot for FAQs and password resets, cutting agent workload by 30% and accelerating resolution times.

• Unified multi-channel tickets into a single dashboard with live SLA tracking.
• GenAI assistant deflected 30% of routine requests and trimmed first-response time by 18%.

Benchmarked ML ensembles to tame irregular demand, reaching 97% forecasting accuracy and clustering SKUs by behavior so planners cut overstock risk by 20% and tighten procurement decisions.

• Feature-store pipelines in Python keep demand signals fresh for nightly re-trains.
• Merchandising dashboards surfaced 20% inventory risk reduction for top 50 SKUs.

Analyzed 400+ food profiles with ANOVA, t-tests, and interpretable logistic regression to surface high-risk ingredients, uncovering a 0.31 correlation between dairy, wheat, shellfish and allergy severity to inform safer meal planning.

• Interactive Tableau dashboards let dieticians probe ingredient-level risk scores.
• Regulatory brief summarised findings for labeling compliance teams.

Fine-tuned DenseNet121 over 112K NIH chest radiographs with mixup augmentation and focal loss, reaching 91.7% macro AUC across 14 pathologies. Deployed triage-ready inference service with Grad-CAM explanations.

• Grad-CAM heatmaps gave radiologists transparent decision support within 1.5s.
• Firebase-hosted inference API processed 5k studies/week with automated quality checks.