Live viralness prediction and strategy platform: type a topic in plain English and get data-backed recommendations on where to post, when to post, and what format to use across 50 communities in 10 industry categories.
- •ML pipeline collecting 500K+ Reddit posts, filtered into a 96K-row model-ready dataset with NLP (VADER sentiment, NRC emotions, readability) and behavioral feature engineering
- •XGBoost virality-index model with temporal train/test split, SHAP explainability, and 20+ experiments tracked in MLflow
- •Dataset versioned on HuggingFace Hub with a DuckDB lookup layer powering real Table 1/2 queries behind the app
- •Deployed Streamlit app with keyword parse layer, confidence tiering, and LLM-assisted guidance for posting strategy
- •Three expertise modes: newbie (plain-English advice), experienced (data tables), expert (full model details and feature importance)
- Python 🐍
- XGBoost
- HuggingFace 🤗
- DuckDB
- MLflow
- LangChain
- FastAPI
- Streamlit
- GitHub Actions
role: ML Tech Lead & PM
Predictive-maintenance demo for a small-molecule API final step (R-101 → CR-201 → FD-301 → ML-401). 100% synthetic data — OT sensors land in Postgres, Databricks trains health/quality models, and a Streamlit app surfaces early warning before residual-solvent failures.
- •Three-tier early warning computed from data: multivariate SPC (Hotelling T²) flags at batch 16 — 5 batches ahead of conventional alarms and 19 before the first quality failure at batch 35
- •PCA/T² equipment-health model (2 PCs, 93.4% baseline variance) with Western Electric 2-point run rule; Ridge quality model for residual-solvent drivers
- •Feature pipeline over a 40-batch campaign: 345K+ tag_history rows → 26-column batch_features table (FD-301 vacuum, cake ΔP, filtration time, LOD, plus upstream/downstream context)
- •CR-201 crystallizer tab: mechanistic cooling-crystallization simulator (numpy-only, no DB) with live Batch Designer, design space / PAR / NOR (~81% PAR-box reliability), and Sobol sensitivity
- •6-tab Streamlit dashboard (Equipment Health, Failure Drivers, Alarms, CAPA, Data Integration, CR-201) backed by Postgres; 25 pytest checks passing
- •Published on figshare (DOI 10.6084/m9.figshare.33511504) — a full in-silico ICH Q8/Q11 characterization of CR-201: Sobol sensitivity, a grid-resolved design space, an audit of PLS and gradient-boosted surrogates against the generating mechanism, and a 2 L → 2000 L scale-up analysis
- Python 🐍
- Streamlit
- Postgres
- Databricks
- PCA / Hotelling T²
- Ridge
- numpy
- pytest
role: Solo build
Autonomous agentic pipeline that eliminates manual job hunting across target company career pages.
- •Scrapes 60+ company career pages daily across Greenhouse, Lever, and Ashby ATS formats
- •Claude-powered role fit evaluation with multi-tier keyword filtering
- •SQLite deduplication and caching layer preventing duplicate alerts
- •SMTP morning email with new listings and match scores
- •Streamlit dashboard to add companies, track history, and generate tailored resumes from job descriptions
- Python 🐍
- Claude API
- GitHub Actions
- SQLite
- Streamlit
- FastAPI
role: Solo build
Provider-facing patient management dashboard with full CRUD, real-time data sync, client-side search and filtering, and Zod schema validation. Built with React, TypeScript, Firebase, Tailwind, and shadcn/ui.
- •Full CRUD patient management workflow for providers
- •Real-time data sync backed by Firebase
- •Client-side search and filtering across patient records
- •Zod schema validation on all form inputs
- React
- TypeScript
- Firebase
- Zod
- Tailwind
- shadcn/ui
role: Solo build