Prakriti Sharma

• –Built an LLM-powered spare-parts cost estimation system, reducing manual quote preparation time by ~60% by mining SAP, SharePoint, and Teams purchasing data with Python, LangChain, and the OpenAI/Gemini APIs.
• –Improved aftermarket pricing accuracy by ~30% by developing an LLM-assisted historical cost modeling pipeline using Python, Scikit-learn, Pandas, and SAP data extraction scripts.
• –Accelerated same-day quote generation for sales teams by building an AI-driven data retrieval and recommendation API and React dashboard, using FastAPI/Node.js, Python, LangChain, PostgreSQL, and LLM embeddings.

• –Engineered multi-level risk blacklisting and whitelisting APIs in Java and Go, expanding system coverage and contributing 7% of product revenue within 5 months.
• –Designed and deployed Success Rate RESTful APIs with ReactJS/TypeScript dashboards, improving transaction visibility and driving 15% revenue growth over 2 quarters.
• –Refactored Java payment gateway modules with zero downtime and wrote Python automation scripts covering 100% of regression testing, improving release velocity.
• –Integrated commonauthsvc into CI/CD pipelines using Jenkins, automating 700+ test cases/day.
• –Built 3 levels of blacklisting APIs in Java, expanding the team's scope by up to 25%.
• –Led a cross-functional team of 10 to deliver a React Native cross-platform app, increasing customer engagement by 40% in 6 months.
• –Developed RESTful APIs for Aggregation Rules in Java and ran performance testing with JMeter, boosting AutoCollect revenue by 9.8% and Cashgram by 17.6%.
• –Achieved 100% automation coverage using Appium integrated with BrowserStack for cross-platform testing.

• –Developed robust prototypes addressing server consistency challenges, using Kafka for event streaming and Zookeeper for distributed coordination, resolving 78% of major system crashes.

Master of Science in Computer Science, CGPA: 3.5/4.0

Bachelor of Technology in Computer Science Engineering, CGPA: 8.56/10.0

Sludge-RAG — Organizational Context Layer for RAGgithub.com/prakriti31/sludge-RAG

A RAG system that surfaces not just what a policy says, but why it exists. Sludge-RAG mines the Slack threads and email chains around a document, runs a multi-turn Claude pipeline to extract the organizational rationale behind each decision, and attaches it as metadata to every indexed chunk — so a query like "why is retention 90 days" comes back with the GDPR and audit context, not just a dry PDF excerpt. A dual-pane "Vibe-Check" playground puts plain vector search side by side with rationale-augmented search so the difference is visible, not just claimed.

AgentHub — Terminal-First Coding Agent Managergithub.com/prakriti31/AgentHub

A SaaS dashboard for developers running multiple AI coding agents — Claude Code, Copilot, Cursor, Codex — in one place instead of juggling terminals and logs. Tracks sessions per agent with token counts, cost, duration, and status, streams a terminal-style activity log, and pushes Slack and email alerts when a session finishes or fails. Auth is JWT plus Google OAuth with per-user @alias handles for team visibility, rate-limited and request-logged on the backend.

TrackSplitAI — Multi-Tenant Expense Splitting SaaSgithub.com/prakriti31/SplitTrack

A multi-tenant expense tracking and splitting SaaS with AI categorization, a conversational assistant, and real subscription billing — not a toy CRUD app. Seven FastAPI microservices talk over a RabbitMQ event bus with dead-letter queues and per-consumer dedup, Gemini categorizes expenses and answers questions, and Stripe Checkout handles billing with signature-verified, idempotent webhooks. Runs on Kubernetes (EKS) via Terraform-provisioned infra, autoscales on CPU and queue depth, and ships through a GitHub Actions pipeline with coverage gates, Trivy scanning, and Prometheus/Grafana/Loki observability.

Instagram Analytics — Real-Time Scraping Pipelinegithub.com/prakriti31/Big_Data_Instagram

An Instagram analytics pipeline that scrapes profiles via Apify, queues the work through Kafka, and pushes results to a React dashboard over WebSockets in real time. FastAPI handles the API layer, a Python consumer service does the actual scraping and writes to MongoDB or DynamoDB depending on deployment, and rate limiting plus per-query performance logging keep the scraping layer honest at scale.

ARIA — Adherence Risk Intelligence Agentgithub.com/prakriti31/ARIA

A between-visit clinical intelligence platform for hypertension management that gives clinicians an 8-minute pre-visit briefing instead of a blank chart. Ingests patient data via FHIR R4, then runs a three-layer nightly pipeline — a deterministic rule engine (gap, therapeutic inertia, adherence-BP correlation, deterioration, and variability detectors) feeds a weighted risk score, which feeds a Claude-generated 3-sentence clinical narrative validated against faithfulness and safety guardrails before it's ever shown. Also runs a deterministic drug-interaction checker, a tool-using clinical chatbot, and a patient-facing PWA for home BP submission — all driven by a 30-second poll worker, no Celery, no Redis queue.

PartSelect Chat Agentgithub.com/prakriti31/partselect-ai-assistant

A scoped AI chat assistant for PartSelect's refrigerator and dishwasher catalog — finds parts by symptom or model number, verifies compatibility against a specific model number (never guesses), and walks customers through installation and troubleshooting with interactive, localStorage-persisted checklists. Built on Claude tool-use for part search, compatibility checks, and order support, with rich inline product cards and hard scope enforcement so it refuses anything outside the parts catalog.

Image Shadow Removal Using Non Shadow Mapping — IJRAR, 2023www.ijrar.org/papers/IJRAR23B3740.pdf

A shadow removal algorithm that detects shadows and lifts them by matching each shadow region to its most similar non-shadow counterpart, then transferring brightness while preserving texture. Detection combines gradient, texture (Texton), and spatial-distance features from mean-shift segmentation with Y-channel and HSI-space brightness cues clustered via k-means. Removal uses HSV histogram matching between each shadow region and its best-matched non-shadow region, followed by boundary smoothing across the seams. Built and evaluated in MATLAB.