Pongpanoch Chongpatiyutt

Hi, I'm Pongpanoch Chongpatiyutt.

AI Engineer

I build RAG workflows that turn messy sources into usable evidence and fail clearly when the data is not enough.

AboutExperienceProject

I'm an AI Engineer focused on RAG systems that can be inspected, debugged, and improved after the first demo.

I started with machine learning through university coursework, then moved toward the engineering problems that decide whether AI products hold up: extraction quality, chunking, metadata, retrieval grounding, and honest fallback behavior.

At Saifa AI, I worked across PDF and web ingestion, Qdrant indexing, RabbitMQ and FastAPI worker flows, evaluation logs, replay payloads, and failure classification. The work taught me to treat an answer as only as strong as the evidence path behind it.

Because most company work is NDA-bound, this portfolio shows the same patterns in an open project: source ingestion, grounded chat, citations, request IDs, rate limits, and tests around the cases where RAG usually breaks.

Experience

AI Engineering Intern, Saifa AI (Sep 2025 - Present)

  • Worked on RAG pipelines for PDF and website sources, including extraction cleanup, chunking choices, metadata design, Qdrant indexing, and retrieval filters.
  • Built evaluation and replay workflows with JSONL logs, payload snapshots, and extraction/retrieval comparisons so regressions were easier to reproduce.
  • Improved RabbitMQ and FastAPI worker flows from message intake to reply output, with safer fallback paths when upstream data or model calls failed.
  • Added guardrails around payload validation, duplicate-event handling, structured logs, and failure classification to make incidents easier to trace.

Tooling

PythonFastAPIRabbitMQQdrantLangChainOpenAI APIAnthropic APIGemini APIRAGVector SearchJSONL Evaluation LogsPrompt EngineeringDockerGit

Student Assistant, TU Berlin - Faculty V (Aug 2023 - May 2025)

  • Managed CNC-based fabrication workflows for open-source hardware projects, from CAD/CAM preparation to machine operation.
  • Supported machine setup, handoff checks, and repeatable fabrication runs for student and research teams.
  • Helped operate makerspace infrastructure where documentation and careful handoffs mattered as much as the tools.
  • Produced practical setup notes so other people could repeat technical work without guessing the missing steps.

Tooling

Fusion 360CAD/CAMCNC ProgrammingWorkshop Tooling3D PrintingTechnical Documentation

Project: Reusable RAG Delivery Foundation

This is the open version of the AI engineering work I want to be judged on: ingest sources, retrieve the right evidence, answer with citations, and handle the cases where the evidence is not good enough.

The goal is not another chat box. The goal is a reusable base that lets a team test business value sooner while still keeping the evidence trail visible for technical review.

Result: less time spent rebuilding ingest and retrieval plumbing, more time spent tuning the corpus, testing edge cases, and deciding whether the use case is worth shipping.

Problem

RAG projects often stall before business validation because ingestion, chunking, retrieval, and citation behavior all need careful setup.

System

A Next.js RAG workflow with PDF/URL ingestion, vector search, session namespaces, grounded chat, and source-level citation traces.

Benefit

The demo makes the product value visible, while the architecture, failure states, and quality controls give engineers something concrete to inspect.

Evidence

Proof of Work

A quick read on what I personally built and what I can explain in an interview.

What I built

PDF and URL ingestion, text cleanup, chunking, OpenAI embeddings, Pinecone namespaces, grounded chat, and citation rendering.

What I made reliable

Request IDs, rate limits, no-evidence responses, stream fallback behavior, clear API errors, and regression-oriented tests.

What I can discuss

Retrieval misses, chunk-size tradeoffs, citation metadata, replay/evaluation logs, API boundaries, and when a RAG answer should refuse.

Core Stack

Next.jsTypeScriptOpenAI APIPineconeBrowserlesspuppeteer-corepdf-parseReact MarkdownTailwind CSS

System Design

Architecture at a glance: one Next.js app handles ingest and chat APIs, indexes evidence in Pinecone, and returns citation-grounded answers with built-in guardrails.

Pong AI Demo

Upload PDF or ingest URL, then chat with grounded responses.

Status

Idle
Uploading
Scraping
Ingesting
Ready

Active Sources

No sources ingested yet.

Pong Assistant

Grounded answers with traceable evidence

Idle

Ask about uploaded files and URL content once ingestion is ready.

Press Enter to send, Shift+Enter for newline.

Verify important details before making decisions.

Companion Tools I Built

The demo is the baseline layer. These companion tools show how I think about RAG quality once the first version is working: measure it, replay failures, and tune with evidence.

RAG Evaluation Harness

  • Runs golden-set queries with expected evidence and explicit pass/fail criteria.
  • Tracks citation coverage, groundedness, hit-rate, latency, and per-query cost across versions.

Stack

PythonJSONL TracesPandasJupyter NotebooksRegression Baselines

Chunking and Retrieval Comparator

  • Compares chunk size/overlap, top-k, thresholds, and reranking strategies on the same corpus.
  • Produces side-by-side retrieval and answer quality reports to guide tuning decisions.

Stack

LangChainRecursiveCharacterTextSplitterOpenAI EmbeddingsPinecone NamespacesRetrieval Parameter SweepsPandasJupyter Notebooks

Edge-Case Replay and Failure Analyzer

  • Replays problematic queries and classifies misses: no-evidence, wrong-source, and partial-answer.
  • Links each failure to trace artifacts so prompt, retrieval, and ingestion fixes are faster.

Stack

FastAPIRabbitMQStructured LoggingReplay PayloadsGit

Contact

For AI/ML/Software opportunities or project collaboration, feel free to reach out.

p.chongpatiyutt@gmail.com
LinkedInGitHub