vmore2
Update TheraMind to Nature pub, SOSTOS, impact, multi-LLM ensemble details
855c577
|
Raw History Blame Contribute Delete
6.91 kB
# Vrushket More - Projects
## TheraMind (Featured Research Project)
**Gen-AI Multi-LLM Ensemble Healthcare Pipeline**
- **Type**: Research Project, First Author Publication
- **Tech Stack**: Python, LangChain, Multi-LLM Ensemble (GPT, Gemini, Mixtral, Claude), LoRA-BERT, RAG, FAISS
- **Publication**: npj Precision Oncology (Nature, 2026) — https://www.nature.com/articles/s41698-025-01265-1
- **Description**: Built a Gen-AI multi-LLM ensemble pipeline processing 1M+ PubMed & clinical documents for NSCLC drug repurposing. Independently converged on the multi-agent ensemble architecture later validated by Google's Co-Scientist (Nature, May 2026).
- **Key Achievements**:
- Solved clinical hallucination (60-70% → 95-98%+ recall) using parallel multi-LLM ensemble with n×n majority-vote hitting 99.7% specificity at single-classifier latency
- Built hierarchical RAG + two-tower LoRA-BERT: ANN on summaries → 1,000 candidates; full-doc re-rank → top 100; favorability-aware BERT (rank-16 LoRA, 100K oncologist labels) → final 20, near-perfect top-12 recall, trained in <2 hrs on a single GPU
- Rankings directly influenced $500K in pharma validation studies
- Work acquired by NASA-affiliated firm SOSTOS
- **Impact**: First author in npj Precision Oncology (Nature, 2026); led conceptualization, methodology, software, and writing
## RuntimeX
**AI Agent Harness**
- **Type**: Agent Framework / LLMOps
- **Tech Stack**: Python, LiteLLM, ReAct Agents, SQLite
- **Description**: A production-grade AI agent harness solving the "black box" problem in LLM workflows. Built for total visibility, it features deterministic replay, time-travel forks, and a self-healing memory system.
- **Key Achievements**: 100% deterministic, O(1) replay cost, zero token waste.
- **GitHub**: github.com/vmore2/runtimex
## Cadence Audiobook Generator
**Generative AI Pipeline**
- **Type**: Generative AI / NLP
- **Tech Stack**: LangGraph, Gemini 2.5 Pro, OpenAI TTS, Streamlit
- **Description**: End-to-end Generative AI pipeline that autonomously transforms raw ebooks into studio-quality audiobooks. Leverages Gemini 2.5 Pro for contextual dialogue attribution and OpenAI for dynamic, emotion-driven voice synthesis.
- **Key Achievements**: 100% automated, 5+ emotions mapped natively.
- **GitHub**: github.com/vmore2/Cadence-Audiobook-Generator.git
## AgentRank
**Novel Temporal Embeddings**
- **Type**: AI/ML Model
- **Tech Stack**: Python, Embeddings, NLP
- **Description**: First embedding model that knows *when* memories happened.
- **Key Achievements**: 0.65 MRR (+21% vs ada-002), 99.6% Recall@5 (+8% vs MiniLM).
- **GitHub**: github.com/vmore2/AgentRank-base
## CogniHive
**Multi-Agent Memory System**
- **Type**: AI/ML System
- **Tech Stack**: Python, Multi-Agent Systems, RAG
- **Description**: Transactive memory for AI teams. Routes queries to the right expert with "who knows what" tracking.
- **Key Achievements**: 6+ integrations, 15x token savings.
- **GitHub**: github.com/vmore2/CogniHive
## QSVAPS
**Quantum Superposition Verification for Agent Plan Safety**
- **Type**: Quantum Computing / AI
- **Tech Stack**: Qiskit, Quantum Computing, Python, OpenAI
- **Description**: Uses Grover's quantum search algorithm to verify AI agent plans, finding constraint violations.
- **Key Achievements**: O(√N) speedup over classical verification.
- **GitHub**: github.com/vmore2/Quantum-AI
## FinFlow
**Financial Intelligence Pipeline**
- **Type**: Data Science / Data Engineering
- **Tech Stack**: Apache Airflow, Python, ETL, Astronomer
- **Description**: Automated financial intelligence pipeline fetching real-time stock & crypto prices, calculating technical indicators, and generating trading signals.
- **GitHub**: github.com/vmore2/Finflow
## Peeker AI Analytics Platform
**Customer Analytics Dashboard**
- **Type**: Professional Internship Project
- **Tech Stack**: Python, ETL, Analytics, Dashboard Development
- **Company**: Peeker AI (Startup)
- **Description**: AI-powered customer analytics platform with intelligent ETL pipelines and personalization features
- **Key Achievements**:
- Expanded paying customer base by 50%
- Improved personalization accuracy from 80% to 93%
- Built automated ETL pipeline for LinkedIn and email data processing
- **Live Demo**: peeker.ai
## Self-Driving Cars Model
**Computer Vision for Autonomous Vehicles**
- **Type**: Academic Project
- **Tech Stack**: Python, OpenCV, SVM, Computer Vision
- **Description**: Real-time lane detection and vehicle classification system for autonomous driving
- **Key Achievements**:
- 98% precision in vehicle detection
- 97% recall for accurate classification
- Real-time processing of 720p video streams
- Heat-map based metrics for pattern identification
- **Features**: Lane curvature detection, vehicle offset tracking, temporal smoothing
## DermWise
**AI-Powered Dermatology Diagnostic Assistant**
- **Type**: Healthcare AI Project
- **Tech Stack**: Deep Learning, Computer Vision, Healthcare AI
- **Description**: AI diagnostic assistant leveraging deep learning for skin condition analysis and classification
- **GitHub**: github.com/vrushketmore/DermWise
## ClimateWise
**Climate Analytics Platform**
- **Type**: Data Science Project
- **Tech Stack**: Python, Data Analysis, Visualization
- **Description**: Data-driven climate analytics platform with predictive modeling for environmental pattern analysis
- **GitHub**: github.com/vrushketmore/ClimateWise
## Clash of Trails
**Trail Recommendation System**
- **Type**: Data Science/Web App
- **Tech Stack**: Recommendation Systems, Data Science, Analytics
- **Description**: Interactive trail recommendation system using data science to match hikers with their perfect outdoor adventures
- **GitHub**: github.com/vrushketmore/Clash-of-Trails
## English to French Translation Model
**Neural Machine Translation**
- **Type**: NLP/HuggingFace Model
- **Tech Stack**: Transformers, NLP, MarianMT, Fine-tuning
- **Description**: Fine-tuned MarianMT model for English to French translation, trained on KDE4 dataset
- **HuggingFace**: huggingface.co/vrushketmore/marian-fine-tuned-kde4-en-to-fr
## Cross-lingual Summarizer
**Multilingual Text Summarization**
- **Type**: NLP/HuggingFace Model
- **Tech Stack**: mT5, Summarization, Multilingual NLP
- **Description**: MT5-small model fine-tuned for English to Spanish summarization on Amazon reviews dataset
- **HuggingFace**: huggingface.co/vrushketmore/mt5-small-finetuned-amazon-en-es
## Power BI Dashboards
**Data Visualization Projects**
- **Type**: Business Intelligence
- **Tech Stack**: Power BI, Data Visualization, Analytics
- **Description**: Interactive business intelligence dashboards including Data Professional Survey analysis
- **Features**: Salary insights, job satisfaction metrics, programming language preferences
- **GitHub**: github.com/vmore2/Power-bi