Economic Indicators Dashboard
Fifteen years of GDP growth and sector-level GVA data, filterable and visualized in real time.
- 15-year GDP trends with year-over-year comparisons
- Sector-wise GVA breakdown with interactive filtering
- Comparative analysis tools for economic research
JavaScript, Chart.js, D3.js, HTML5, CSS3
CytoGuard: Cellular Injury Detection in Cell Painting Screens
An open platform that flags injured wells in Cell Painting drug screens, names the likely injury mechanism, and shows the cells behind every call. Evaluated on 16,701 wells from IDR idr0133, with every compound scored by a model that never saw it.
- Macro-F1 of 0.462 on never-seen compounds, against 0.146 for the published Pycytominer baseline; AUROC 0.947 for injured vs healthy wells
- Raw images to injury calls with OME-Zarr, Cellpose-SAM and cp_measure; pixel-level scores match the published profiles at r = 0.95
- FastAPI job queue, plugin system and a web viewer with segmentation overlays; 20 tests and a one-command Snakemake workflow
Python, Cellpose-SAM, cp_measure, Pycytominer, LightGBM, FastAPI
PCVI Atlas: Heatwave Vulnerability Research
Public-health research pipeline testing whether U.S. communities with high prescriptions of heat-sensitizing medications face greater vulnerability during heatwaves, linking 8.79M Medicare Part D claims to summer heat exposure across Hospital Referral Regions.
- Four-stage pipeline: drug crosswalk (276 medications resolved to RxCUI/ATC), claims processing, evidence weighting, and causal analysis
- 6-year daily panel with conditional Poisson regression and 200-permutation placebo testing to separate correlation from causation
- Identified and corrected a prescriber-vs-patient-location denominator artifact by reattributing claims via hospital catchment areas
- Interactive Streamlit app for exploring results across NASA POWER, NOAA, and Census data sources
Python, Streamlit, Pandas, RxNorm/RxClass API, Dartmouth Atlas HRR
AMR-Climate Atlas: Antimicrobial Resistance Research
Causal panel testing whether a hotter year makes antibiotic resistance worse across 29 EU/EEA countries. Four government data sources are joined into a 2000-2023 panel, with a mandatory 200-permutation placebo test gating every result.
- Two-way fixed-effects regression across four pathogen-antibiotic combinations, cross-checked against independent genomic resistance data via BigQuery
- Live MCP server (Anthropic's official SDK) exposing live weather and genomic-resistance data, plus an eval-gated CI pipeline that refits every model on push
- Caught and documented a confounder column silently absorbed into the model's fixed effects, a bug with no error message
- Reported an honest mostly-null result: three of four combinations show nothing, the fourth a narrow, independently-corroborated exception
Python, Streamlit, MCP SDK, BigQuery, Fixed-Effects Panel Regression
burst-scheduler: Local-to-Cloud Burst Job Queue
A small job queue that runs jobs on local machines and, when a job would wait too long, sends it to cloud workers on Kubernetes or AWS Batch. Modeled on Princeton's Orchestrated Burst Compute and the HydroGEN stack.
- Policy engine weighs time saved against cloud cost, with deadlines, priorities, a spend-rate cap and a daily budget, and records its reason on every job
- JetStream dispatch with message de-duplication and redelivery, so a controller restart or worker crash loses no jobs
- React + TypeScript dashboard with live wait times, spend rate, backend usage and runtime policy controls
Python, FastAPI, NATS, Kubernetes, AWS Batch, React, Docker
Digital Wallet Platform (PayMe & AwiPay)
Enterprise-grade digital wallet application serving 100K+ users in Nepal. Secure transaction processing, spending controls, and a robust service layer architecture.
- Designed RESTful APIs for transaction processing
- Service layer encapsulating business logic & data access
- Security measures for financial data protection
- Transaction controls and configurable spending limits
Java, Spring Boot, Hibernate, MySQL, Maven
ML: House Price Prediction System
An ML system that predicts house prices from user preferences and market data, with a full pipeline from preprocessing to a Flask front end.
- Data cleaning and preprocessing pipeline
- Feature engineering for accuracy improvements
- Interactive web interface for preference input
- Real-time prediction results with explanations
Python, scikit-learn, pandas, NumPy, Flask
CISO Budget Defense Calculator & Conference Paper
Published conference paper + interactive calculator for CISOs to compare Cold Site vs. Hot Site disaster recovery strategies. Translates technical metrics into financial impact.
- Real-time cost comparison with customizable RTO/RPO parameters
- Executive-ready report generation with charts and ROI analysis
- Risk quantification translating cybersecurity metrics into financial impact
- Evidence-based recommendations for disaster recovery investments
HTML5, CSS3, JavaScript, Chart.js, Bootstrap
Small Business Reality Check: ALE Tool
Risk-assessment tool for small business owners that translates cybersecurity threats into concrete Annual Loss Expectancy (ALE) metrics and investment decisions.
- Interactive ALE calculator with customizable threat scenarios
- Visual risk matrices: probability vs. impact
- Cost-benefit analysis for security controls & insurance
- Executive summary generation for stakeholders
HTML5, CSS3, JavaScript, D3.js, Bootstrap
Warranty Claims Management App
A warranty-claims process modeled end-to-end on the Salesforce Platform, from customer submission through technician resolution, built as one coherent system, not disconnected samples.
- Trigger framework, Batch/Queueable/Schedulable Apex, and a custom Error Log for exception tracking
- Multi-step LWC claim wizard, technician dashboard, and Lightning Message Service for cross-component comms
- Screen Flows, dollar-threshold Approval Processes, Platform Events, and Named Credentials/External Services integration
- Agentforce conversational agent that checks claim status and creates/updates claims via invocable Apex
- SFDX project structure with GitHub Actions CI running Apex tests & coverage checks on every push
Apex, LWC, Flow Builder, Agentforce, Platform Events, GitHub Actions
Harvest Market Platform
Multi-persona Salesforce platform built for the UNLV Management & Entrepreneurship Department, with a custom data model for real-time inventory and complex producer sales hierarchies.
- Custom data model for a multi-persona Producer/Consumer platform
- Dynamic Screen Flows guiding product listings & checkout
- Automated Order creation and Inventory Management
Screen Flow, Custom Objects, Sales Hierarchies
CI/CD Release Pipeline Automation
Automated CI/CD pipelines across Salesforce Education Cloud and Sales Cloud environments, streamlining Apex and LWC release cycles for the development team.
- Deployment time reduced by 87%
- ~150 engineering hours saved monthly
- Streamlined Apex & LWC release cycles end-to-end
- Built on Salesforce DX with automated org validation
CI/CD, Copado, GitHub Actions, Jenkins, Salesforce DX
Sales & Service Cloud Solutions @ UNLV OIT
Designed and delivered Salesforce Sales & Service Cloud solutions using Flow, Apex triggers/classes, and LWC to support departmental operations and student service workflows.
- Flow automation for departmental & student service workflows
- Apex triggers/classes and LWC for operational tooling
- Roles, profiles, permission sets, public groups & sharing rules administration
- Bulk data operations via Data Loader for secure, compliant access
Apex, LWC, Flow Builder, Data Loader
Qualtrics Form Automation at Scale
Manages hundreds of Qualtrics forms with scalable Flow automation and Custom Metadata, so departments reuse the same student application process every year without rebuilding it.
- Custom Metadata-driven Flow automation
- Profile-based access control across hundreds of forms
- Reusable annual application process, no redevelopment needed
Flow Builder, Custom Metadata Types, Qualtrics Integration
Org-to-Org Data Seeding & Migration Tool
A Bulk API tool that seeds data org-to-org, matching parent-child records automatically to preserve relationships and skip Data Loader's manual limits.
- Automated parent-child record matching across environments
- Sandbox data seeding time cut by 80%
- Bulk API integration bypassing manual Data Loader limits
Salesforce Bulk API, Apex, Data Migration
mini-swe-agent: Agent-Computer Interface Ablations on SWE-bench Lite
A small SWE-agent-style coding agent that reads a GitHub issue, explores the repo, edits code, runs tests and submits a patch, built to measure whether the tools an LLM is given matter as much as the model. Runs entirely on local Ollama models.
- Swappable agent-computer interface (windowed viewer vs full-file dump, ripgrep vs grep) behind a tool registry, scored in official SWE-bench Docker containers
- Ablation: the windowed interface converged about 2.5× faster (10.5 vs 26.0 steps) and reached the edit stage where the basic one never did
- Found and fixed a parser bug that silently discarded 37 consecutive correct tool calls; best attempt passed 9 of 10 tests
Python, Ollama, Docker, SWE-bench harness, pytest
Noisy Survey: Randomized-Response Anonymous Surveys
An anonymous survey app that adds noise to sensitive yes/no answers before they are stored, so no answer can be traced back to anyone, yet the true rate can still be recovered.
- Forced-response and Warner mechanisms with unbiased estimators, 95% confidence intervals and a per-question ε (local differential privacy)
- Stores only id, question and noisy answer; a test fails if an identifying column is ever added, and the access log is off
- Dashboard shows raw noisy results next to the corrected estimate
Python, FastAPI, SQLite, Jinja2, pytest
Zenith: Enterprise AI Security & Observability Gateway
An open-source security and observability gateway for the Model Context Protocol (MCP): one middleware proxy that replaces scattered, client-side safety assumptions with a single enforced checkpoint.
- Built a 6-step connection handler managing real-time JSON-RPC 2.0 streams over WebSockets/stdio
- Monitors AI execution times, token usage, and payloads in detail
- Decoupled middleware architecture built to scale AI-agentic workflows without latency/state bottlenecks
Python, WebSockets, JSON-RPC 2.0, MCP
Hyperion: AI-Native SRE Mesh
Autonomous "Air Traffic Control" for massive-scale distributed infrastructure. A zero-trust middleware gateway that validates every AI-generated action against RBAC, PII redaction, and Safety Intent Analysis before execution.
- Mitigates catastrophic production risk by gating AI actions before they touch live systems
- Real-time infrastructure dashboard over WebSockets with live node health grids in a single view
- Automated audit log scoring every proxied request and AI decision
JavaScript, WebSockets, RBAC, Claude
CAFE-RAG: Climate-Health Dataset Discovery Layer
Agentic, retrieval-augmented search over 1,300+ Harvard Dataverse climate-health datasets (CAFE, CIESIN, HELD), turning natural-language queries into ranked, DOI-linked results with joinable-dataset suggestions instead of manual metadata browsing.
- Seven-stage pipeline: harvest → structure → dual-index (FTS5 + Chroma) → knowledge graph → hybrid retrieval → evaluation → Streamlit UI
- Hybrid retrieval combining lexical (SQLite FTS5) and semantic (Chroma) search via Reciprocal Rank Fusion
- Local LLM (Ollama) query-intent extraction and metadata classification with fully offline inference
- Reproducible evaluation harness with synthetic gold queries: recall@10 of 0.773, MRR of 0.635
Python, SQLite (FTS5), ChromaDB, Ollama, Pydantic, Streamlit
Project Aegis: Enterprise MCP Security Gateway
WebSocket proxy sitting between AI agents and MCP servers, enforcing three gates (RBAC, AI-driven intent analysis, and PII redaction) on every JSON-RPC call, with a full audit trail and live observability dashboard.
- Agent-specific allowlists restrict which MCP methods each agent can invoke
- Local LLM (Ollama) evaluates destructive operations (delete/drop/write/merge) before execution; fail-closed if the model is unreachable
- PII scrubbed from payloads sent to the model while original data still reaches backend servers
- Supabase-backed audit log with an HTTP API and dashboard for allowed/denied/blocked decisions
Node.js, WebSockets, Supabase (PostgreSQL), Ollama
AI Knowledge Assistant (GraphRAG)
Retrieval-augmented generation assistant combining semantic document search, a knowledge graph layer, and webhook-based workflow automation for organizational knowledge bases.
- Vector-based semantic search over ingested docs via ChromaDB
- Entity extraction and relationship storage in Neo4j
- FastAPI chat endpoint returning source-attributed answers, with webhook automation hooks
Python, FastAPI, ChromaDB, Neo4j