Xavier Puspus
AI Architecture and Engineering
About
I work end-to-end on AI systems and the tooling around them: retrieval and evaluation, agent memory, satellite remote sensing, cloud infrastructure described in natural language, civic data, and image forensics. Comfortable from data ingest through backend services to a polished frontend. Mostly Python and TypeScript. Physicist by training.
Technical stack
- Languages
- Python, TypeScript, JavaScript, SQL
- AI / ML
- LangGraph, MCP, OpenAI SDK, Anthropic SDK, scikit-learn, PyTorch, Transformers, CLIP
- Retrieval
- Vector stores, contextual retrieval, RAPTOR, PageIndex, knowledge graphs, hybrid, rerankers
- Geospatial
- Google Earth Engine, Sentinel-1 InSAR, Sentinel-2, CLIP-based detection, MapLibre, GeoJSON
- Eval / stats
- Benchmark harness design, agent-memory evaluation, bootstrap CIs, permutation tests, calibration
- Backend
- FastAPI, Flask, Postgres
- Frontend
- Astro, Next.js, React, Tailwind, ECharts
- Infra
- AWS, Netlify, Docker, GitHub Actions
- Domains
- Retrieval, agent memory, satellite remote sensing, civic data, fraud detection, image forensics, data visualization, NLP, fairness, RAG
Featured projects
- cloudwright Python · ★ 31
AI-powered cloud architecture - describe infrastructure in natural language, get Terraform, cost estimates, and compliance reports
- kb-arena Python · ★ 7
Benchmark 9 retrieval architectures (vector, contextual, QnA, knowledge graph, hybrid, RAPTOR, PageIndex, BM25, rerank) on your own docs. Automated hyperparameter search with bootstrap CIs and significance tests.
- ai-workflow-benchmark Python · ★ 10
Benchmark harness measuring AI coding tool+workflow performance, not just model capability. 100 tasks, sigmoid scoring, 12 capability dimensions, gap analysis.
- ph-civic-data-mcp Python · ★ 3
The first multi-source MCP server for Philippine civic data: earthquakes, weather, typhoons, procurement, population, poverty
- solar-map-ph Python · ★ 0
Open-source rooftop solar detection from public satellite imagery. CLIP-ViT-L + logistic regression, Platt-calibrated, deterministic Docker build. Coverage: Greater Metro Manila. Free homeowner roof-lookup tool, CC-BY-4.0 data, MIT code.
- ghostwatch Python · ★ 3
Satellite verification of government infrastructure , see it from space
Selected open source
Evaluation and benchmarks
- memory-arena 20 agent-memory strategies through one eval, one judge, one model, with reproducible result stamping. Python
Satellite and geospatial
- leaves-ph Metro Manila tree canopy measured from Sentinel-2, per LGU, 2019 to 2026. Python
- sinkmap-ph Land subsidence for seven Philippine cities from Sentinel-1 radar, 2016 to 2025. Python
- floodwatch-ph Two-track flood mapping from Sentinel-1 SAR and a recurrence model. Python
- ghostwatch Satellite verification of government infrastructure. Python
Civic and Philippine open data
- dataviz-ph Animated bubble charts joining public spend, GDP, and poverty, 2014 to 2024. Python
- spend-watch-ph Monitoring government spending from public records. TypeScript
- flood-guard-ph Flood-control project monitoring. Python
- AIforSEA Submission to the Grab Safety Challenge. Jupyter Notebook
Developer tooling
- pip-skill Turn any installed Python package into a coding-assistant skill from its real function signatures. Offline, no API key. Python
AI safety and forensics
- parity-fairness Measure bias from data and machine learning models. Python
- fraud-detection-ids Error level analysis and noise analysis on government IDs. Python
Document AI and chat tooling
- ChatPDF Web app to chat with any PDF. Python
- ChatWEB Web app to chat with parseable websites. Python
- ChatCSV Web app to chat with your CSV. Python
- auto-question-generation Generate questions and possible answers from input text or URL. Jupyter Notebook
- ZeroShotClassification Text classification with minimal data, custom labels and context. Jupyter Notebook
- gpt2_text_completion Web application for text completion using GPT2. Python
On-device and offline AI
- airgap Air-gapped AI customer support. On-device LLM, offline-first, config-driven, 7 industry templates. TypeScript
Generative tools and agents
- ai-leetcode-solver Solve LeetCode problems from screenshots via GPT. JavaScript
- ai-instagram-caption-generator Upload an image, generate Instagram captions. JavaScript
- ai-company-search GPT-driven search for company metadata. JavaScript
Domain ML and applied modelling
- aedes Mosquito-borne disease modelling for project AEDES. Jupyter Notebook
- pads Precision agriculture and data science in the Philippines. Jupyter Notebook
- Earthquake-Analysis M6+ earthquake count analysis, bar-chart-race prep, and forecasting attempts. Jupyter Notebook
- news-image-classifier-webapp Image-based news classifier (fake / good / bad / sports / science). Python
- NLP-transfer-learning Custom named-entity recognition via transfer learning. Jupyter Notebook
- pricing-webapp Pricing model deployed as a web application. Jupyter Notebook
- review-rating-predictor-webapp Predict review rating from text. Jupyter Notebook
- dow-jones-sentiment-predictor-webapp Stock direction prediction from news article sentiment. Jupyter Notebook
- census-income-webapp Income-bracket prediction deployed as a web app. Jupyter Notebook
- advertising-webapp Sales forecast model deployed as a web app. Jupyter Notebook
Full list of 55 repositories at /projects .
Education
- BS Applied Physics, University of the Philippines Diliman
- MS Applied Business Analytics, University of Asia and the Pacific