Skip to main content
Pakon Poomson
Public engineering evidence

Projects built beyond the prototype.

Six GitHub-backed case studies covering retrieval, multimodal AI, validated document analysis, forecasting, explainable ML, and public-data systems.

AI Engineering

Agentic RAG, multimodal extraction, and validated document-analysis products.

Case studyPython + FastAPI + LangGraph

Customer Support RAG Triage Agent

Support teams need consistent ticket triage, relevant evidence, and grounded draft responses without relying on a generic chatbot.

Customer support RAG triage dashboard and workflow concept.

What I built

  • System: A retrieval-grounded support triage system with observable agent steps, safe fallback behavior, and measurable evaluation.
  • Implementation: Fixed support workflow with typed state instead of an open-ended chatbot loop.

What it proves

  • Skills: RAG architecture, LangGraph orchestration, vector retrieval, provider routing, evaluation design, FastAPI, React, Docker, and deployment troubleshooting.
  • Core stack: Python, FastAPI, LangGraph, Qdrant
RAGLangGraphQdrantEvaluation
DetailsGitHub
Demo availableNext.js 16 + React 19 + TypeScript

Receipt AI Expense Tracker

Manual expense entry is slow and error-prone when receipts mix Thai and English text, inconsistent layouts, and Buddhist Era dates.

Receipt expense tracker dashboard with totals, categories, and receipt history.

What I built

  • System: A multimodal receipt-to-expense workflow with safe provider routing, human review, local persistence, and analytics.
  • Implementation: Image-capability filtering prevents receipt images from being sent to text-only providers.

What it proves

  • Skills: Multimodal AI product design, Thai/English normalization, schema enforcement, local-first architecture, Next.js, IndexedDB, and testable provider fallback.
  • Core stack: Next.js 16, React 19, TypeScript, Gemini
Multimodal AILocal-firstThai/EnglishAnalytics
DetailsGitHubDemo
Demo availableReact + Vite + FastAPI

AI Resume Matcher

Candidates need an evidence-based way to compare a PDF resume with a job description and identify actionable gaps.

Structured AI resume analysis report with score, strengths, gaps, and recommendations.

What I built

  • System: A validated document-analysis pipeline that converts a PDF resume and job description into a structured career report.
  • Implementation: Validates extension, MIME type, size, extractable text, and job-description length.

What it proves

  • Skills: Production-minded AI APIs, provider abstraction, Pydantic contracts, secure file handling, React reporting, automated testing, and Vercel delivery.
  • Core stack: React, Vite, FastAPI, Python
Document AIFastAPIProvider routingCareer tooling
DetailsGitHubDemo

Data & ML

Reproducible forecasting, explainable classification, and public-data intelligence.

Case studyPython + pandas + statsmodels

Climate CO2 Forecasting ML

Atmospheric CO2 forecasting needs leakage-safe time-series evaluation rather than a visually impressive model trained on future information.

CO2 Forecast Lab dashboard showing model comparison and atmospheric trend.

What I built

  • System: An end-to-end forecasting and anomaly-analysis system with reproducible data preparation, model comparison, API serving, and dashboard evidence.
  • Implementation: Chronological train, validation, and test splits prevent future leakage.

What it proves

  • Skills: Time-series methodology, leakage prevention, honest model evaluation, statistical forecasting, PyTorch experimentation, FastAPI, React, and reproducibility.
  • Core stack: Python, pandas, statsmodels, scikit-learn
ForecastingTime seriesEvaluationAnomaly detection
DetailsGitHub
Frontend showcasePython + scikit-learn + PyTorch

Explainable Cancer Diagnosis ML

A medical ML portfolio project must show error costs, explainability, and strict limitations instead of presenting accuracy alone.

Explainable cancer diagnosis dashboard with measured model results and disclaimer.

What I built

  • System: An explainable tabular-ML workflow that connects model comparison, safety metrics, SHAP evidence, inference contracts, and a reviewer-facing dashboard.
  • Implementation: Shared stratified split and validation-driven model selection.

What it proves

  • Skills: Explainable ML, model governance, classification metrics, error analysis, scikit-learn, PyTorch, SHAP, FastAPI, React, and responsible communication.
  • Core stack: Python, scikit-learn, PyTorch, SHAP
Explainable MLSHAPClassificationResponsible AI
DetailsGitHubDemo
Demo availableNext.js + TypeScript + FastAPI

Thai Procurement Intelligence

Thai public procurement records are difficult to search, compare, and explain when data is multilingual, high-volume, and inconsistently structured.

Thai procurement intelligence home dashboard with records, budget metrics, and province counts.

What I built

  • System: A bilingual public-data intelligence platform spanning ingestion, normalized storage, retrieval, analytics, and evidence-backed Q&A.
  • Implementation: Validation, normalization, deduplication, and import counters for source ingestion.

What it proves

  • Skills: Data engineering, public-data product design, PostgreSQL/pgvector, FastAPI, Next.js, bilingual UX, LLM integration, and transparent data limitations.
  • Core stack: Next.js, TypeScript, FastAPI, Python
Data engineeringHybrid searchBilingualPublic data
DetailsGitHubDemo