Skip to content
View SiemonCha's full-sized avatar

Block or report SiemonCha

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
SiemonCha/README.md

Sansiri Charoenpong (Siemon)

ML & Backend Engineer — I build Python services around machine learning models and ship them behind APIs with tests and CI.

BSc Computer Science, University of Exeter (2026). Based in Bangkok, open to relocation across Asia. Open to ML engineering, backend, and data roles.


Currently

  • Drilling SQL — window functions, CTEs, query plans, and index behaviour. Solutions and notes in sql-practice.
  • Extending the LLM evaluation work from my dissertation: measuring how much scoring output moves when the prompt is paraphrased but the meaning isn't.
  • Building out backend fundamentals around the services below — rate limiting, caching tiers, graceful degradation on model timeout.

Selected work

ImageTrust-AI — real vs. AI-generated image detection, 94.5% accuracy, with a second head predicting the generator family (GAN / diffusion / other). ResNet18 + Grad-CAM, served via FastAPI. → Live demo on Hugging Face Spaces

pcb-defect-detection — YOLOv8 6-class defect detector. Quantised with ONNX Runtime INT8 for a 3× inference speedup (36ms → 12ms), containerised behind FastAPI with GitHub Actions CI/CD.

thai-sentiment — PhayaThaiBERT fine-tuned on a 21k-message Thai corpus to 82% accuracy with class-weighted training. FastAPI service at 85% test coverage; 1,000+ inference requests served on a public Hugging Face deployment.

ECM3401-LLM-Essay-Scoring — BSc dissertation, first-class. 3,600 CEFR classifications across 18 conditions (GPT-4o-mini, Phi-3-mini). Found that aggregate accuracy hid a systematic bias — 90% of B2 essays graded one band low — invisible until the data was split by segment.


Experience

AI Engineering Intern, SCG Digital (Siam Cement Group) · Jul–Aug 2025 Cut repeated-request latency from 2s to 200ms with Redis caching and async processing. Built the seven-category metrics aggregation layer consolidating evaluation output from 50+ model endpoints into one reporting view used by the internal red team.


Stack

Python · SQL · PyTorch · Transformers · YOLOv8 · ONNX Runtime · FastAPI · Docker · PostgreSQL · pytest · GitHub Actions


siemon.cha@gmail.com · LinkedIn

Pinned Loading

  1. ECM3401-LLM-Essay-Scoring ECM3401-LLM-Essay-Scoring Public

    Measuring semantic robustness in LLM-based CEFR essay scoring through systematic prompt paraphrasing. University of Exeter Year 3 Computer Science research project.

    Python 1 1

  2. ImageTrust-AI ImageTrust-AI Public

    94.5% accuracy AI image detector with generator type classification (GAN/Diffusion/Other). ResNet18, Grad-CAM, FastAPI. Live demo on HuggingFace.

    Python

  3. pcb-defect-detection pcb-defect-detection Public

    Production PCB defect detection: 99.5% mAP, 3x inference speedup (ONNX+INT8), FastAPI deployment. YOLOv8 + complete MLOps pipeline.

    Python

  4. thai-sentiment thai-sentiment Public

    Production Thai NLP: 82% accuracy, 1000+ requests served. BERT fine-tuning, REST API, HuggingFace hosting, ONNX optimization.

    Python