Peerapon Wechsuwanmanee

• Led development of a multi-agent AI system that won the 'Technology Love' category at the 2025 company hackathon, which involved hundreds of developers. The system automatically finds internal documentation and REWE-specific infrastructure fixes for common software issues. Internal estimates put manual lookups and cross-team questions at ~5% of weekly developer time.
• Led the move to production for the centralized store-employee assistant's multimodal hybrid search. It replaced a third-party keyword tool and cut document lookups from hours to under 5 minutes. Designed its Google ADK A2A orchestrator with MCP, Kubernetes, and LangSmith to route queries between document-search and task-system agents. Rollout to thousands of German markets is planned by the end of 2026, and its natural-language task-creation extension placed 3rd at Google Agent Factory 2026.
• Defined the shared GenAI/ML architecture for production deployment and monitoring.
• Built a few-shot RAG service with LangChain, Qdrant, FastAPI, and Snowflake to generate personalized coupon copy from aggregated master-data offers. It replaced weekly manual copywriting after the move away from a third-party loyalty platform, while editor corrections supplied curated examples for later generations.
• Built a multi-agent GenAI system with FastAPI, Snowflake, and pgvector that matches weekly offers to seasonal recipes for Easter, Christmas, and summer barbecues. The system gives marketers an explanation for each match and saves hours of manual work each week.
• Took a multi-source recipe data collection service from a 3-hour AI-generated prototype to the production pipeline that now powers Penny's seasonal recipe-matching system.
• Introduced spec-driven development practices beyond the immediate team through sessions attended by nearly a thousand internal developers across Europe.
• Co-founded the Machine Learning Guild, a cross-team community of 10 to 20 ML practitioners. Regular talks and hands-on sessions help teams standardize ML workflows and reuse existing accelerators instead of rebuilding them. New engineers onboard faster through the community.

• Integrated a legacy high-performance database engine with Apache Spark by building the required Scala components.

• Built a reusable AWS MLOps accelerator with Terraform-provisioned SageMaker training and inference templates plus MLflow tracking. It cut the time to build proofs of concept across client engagements.
• Applied the accelerator to an LLM-based PDF summarization project with LangChain, from training to serving.
• Applied the accelerator to a PyTorch computer vision system that detects surface defects in steel production-line images.
• Built a PySpark predictive-maintenance pipeline for an automotive client, from data extraction and feature engineering to a Streamlit decision dashboard used to reduce maintenance costs.
• Extended and optimized a complex Scala/Spark ETL pipeline for automotive IoT analysis. The changes improved throughput and reliability when processing terabytes of sensor data.
• Delivered a long-stalled ETL project in half the estimated time by implementing missing domain equations and reworking the data-processing workflow.

• Led a team of 4 to 5 engineers for a Speech-to-Text product. The team developed speaker diarization, speaker recognition, and speech recognition models.
• Built an ASR pipeline for a low-resource language, from data acquisition through fine-tuning to deployment. It reached an 8% character error rate on the in-domain Thai test set, compared with 11% for Google Speech Recognition.

• Developed multiscale finite-element models of damage and formability in advanced high-strength and multiphase steels, including dual-phase steels. The work combined micromechanical ABAQUS simulations through the Python API with experimental material characterization.
• Authored 9 peer-reviewed journal and conference papers (2014-2021) spanning surface-roughness effects on cold formability, hybrid damage mechanics, hole-expansion and edge-crack sensitivity of multiphase steels, and micromechanical plasticity-and-damage initiation (NUMIFORM, IDDRG, Materials & Design, Computational Materials Science).
• Developed multiscale finite-element tools and multi-objective optimization methods for abrasion-resistant and wear-resistant steels within a university and industry consortium funded by the European Commission's Research Fund for Coal and Steel.
• Applied SVR, ANN, and LSTM models to predict ferritic-steel flow curves. The models treated the stress-strain response as time-series data.
• Coordinated international research projects across institutions and maintained the MATLAB, Python, and ABAQUS toolchain used for the institute's damage-modelling work.

• Built Python-driven 2D and 3D computer-aided engineering models of multiphase steel microstructures in ABAQUS. The work supported later research and contributed to 2 peer-reviewed journal publications.

• Developed an automation workflow for multi-body simulation of an electrical drive in MD ADAMS.
• Built thermal network models for the window-lift and wiper drives in ABAQUS.
• Automated the optimization workflow for both projects with optiSLang.

Personal Portfolio & Automated CV Generator

• Built and deployed a personal portfolio and blog as a fully static Astro site with React interactive islands, Tailwind CSS, and TypeScript, shipped to Cloudflare Pages.
• Architected a single shared JSON data contract - extracted from a YAML skeleton - that drives both the homepage and an automated Typst-based CV generator, keeping all content in sync across outputs.
• Added an MDX-driven blog with RSS feed and auto-generated sitemap, build-time image optimization, and a Vitest + Testing Library suite.

Semantic Product Search Engine

• Researched and evaluated search engine options for a second-hand e-commerce platform, selecting Meilisearch for its hybrid semantic/lexical search, bilingual Thai/English support, and cost profile at scale - saving months of in-house search development.
• Designed and deployed a real-time PostgreSQL → Meilisearch sync pipeline using logical replication and MeiliBridge, enabling sub-second product index updates without changes to the host application.
• Tuned search relevance through ranking rules, typo tolerance, embeddings, and query strategies (type-ahead, faceted filtering, brand relaxation) ahead of launch.
• Contributed compatibility patches to the open-source MeiliBridge Rust codebase to unblock production adoption.

Motorbike Condition Evaluation

• Built a multi-modal ML regression pipeline fusing 39-dimensional MFCC audio features (extracted from inspection videos via librosa), Thai-language BERT embeddings (WangchanBERTa), and categorical inspection inputs to predict a 0-100 motorcycle condition score.
• Benchmarked 5 regression algorithms (ElasticNet, Random Forest, Gradient Boosting, SVR, XGBoost) with scikit-learn RandomizedSearchCV hyperparameter tuning; serialized the best-performing pipeline for cloud-hosted inference.
• Engineered an audio feature extraction pipeline that converts MP4 inspection videos to log-mel spectrograms, computes MFCC + delta + delta-delta coefficients, and normalizes across axes to produce a stable 39-feature vector.
• Deployed a production FastAPI + Streamlit application with object-storage video storage, PostgreSQL result logging, and Docker — submitting a video returns a scored risk assessment in real time.

Credit Risk Assessment

• Performed exploratory analysis and built a reproducible data-cleaning and transformation pipeline on a raw credit dataset.
• Ran model selection across candidate classifiers with per-segment personalized feature selection to sharpen credit-default discrimination.
• Delivered an interactive Streamlit decisioning UI to score applicants and inspect the drivers behind each prediction.

Credit Default Prediction

• Built a full-lifecycle credit-default MLOps system: EDA, automated data-cleaning pipeline, MLflow experiment tracking, blue-green model selection, and Prefect-orchestrated automatic retraining.
• Ran a profit-optimized threshold simulation achieving a 30% profit increase on out-of-sample data versus a baseline cut-off policy.
• Deployed a production-grade, scalable FastAPI inference service serving real-time default predictions, with retraining triggered automatically as new repayment data arrived.

Customer Feedback Classification

• Reduced case resolution time by 18% by building an automated Thai-language text classifier that routes incoming feedback to the correct service team.
• Implemented a RAG pipeline using OpenAI embeddings and ChromaDB to retrieve similar labeled examples as few-shot context, improving classification accuracy without retraining.
• Packaged the system as a Dockerized FastAPI service, enabling new training examples to be added at runtime via CSV upload without redeployment.

Automotive License Plate Recognition

• Built a MATLAB computer-vision system that reads automotive license plates, developing the recognition pipeline almost from scratch in the era before deep-learning frameworks were available.

Built a multi-agent extension for REWE's centralized store-employee chatbot. It creates tasks from natural-language requests through a FastMCP tool on a decoupled Google ADK A2A architecture.

Built a multi-agent AI system that finds internal documentation and REWE-specific infrastructure fixes for common software issues.

Built a deep learning model for detecting anomalies in electrical drives from real-time data.

Built ML models to study microstructure, fracture toughness, and early-stage fatigue in nuclear power plant steels using non-destructive evaluation measurements. Worked with an international team using an agile process.

Built an NLP model to predict English-language product ratings from Malaysian and Indonesian e-commerce sites. The entry placed in the top 5%. Also built a computer vision model for 42-class product image classification.

Improved predictive maintenance accuracy from turbocharger inspection data.

Built an ML model to predict whether a laser machine could extract a part successfully from its geometry and machine parameters.

Mechanical design and manufacturing engineer on Team Skuba, Kasetsart University's interdisciplinary autonomous-robot-soccer team, at the international RoboCup Small Size League.

Influence of surface roughness on cold formability in bending processes: a multiscale modelling approach with the hybrid damage mechanics model

International Journal of Material Forming, 14(6) (2021)Source

Numerical Evaluation of Surface Roughness Influences on Cold Formability of DP Steel

13th Intl. Conf. Numerical Methods in Industrial Forming Processes (NUMIFORM) (2019)Source

Modeling the Surface Roughness Influence on the Hole Expansion Ratio of Multiphase Steel

38th Intl. Deep Drawing Research Group (IDDRG) Annual Conference (2019)Source

Scalebridging Approaches to Assess the Edge Crack Sensitivity of Multiphase Steels

13th Intl. Conf. Numerical Methods in Industrial Forming Processes (NUMIFORM) (2019)

A Numerical Approach to Evaluate Roughness Effects on Localization and Damage in Sheet Materials

37th Intl. Deep Drawing Research Group (IDDRG) Annual Conference (2018)Source

Finite Element Model in Abrasion Analysis for Single-Asperity Scratch Test

Fracture Fatigue and Wear, Springer (2018)Source

Multi-objective Optimization of Multi-scale Finite Element Analysis for Wear Resistant Steel

U.S. National Congress on Computational Mechanics (2017)

Prediction of Plasticity and Damage Initiation Behaviour of C45E+N Steel by Micromechanical Modelling

Materials and Design (2017)Source

The Modelling Scheme to Evaluate the Influence of Microstructure Features on Microcrack Formation of DP-steel

Computational Materials Science (2014)Source

Development of the Modified Kinematics for a Wheeled Mobile Robot

Intl. Technical Conf. Circuits/Systems, Computers and Communications (2010)

Kinematics Simulation by using MSC.ADAMS for a 6-DOF Industrial Robot Arm

Thailand Simulation and Modeling Conference (2010)

• Designed and taught a weekend course on AI product development for ~40 IT and non-IT professionals, from model to deployed service.
• Topics: Python, Machine Learning, FastAPI, LLM API, Streamlit, Docker, Linux, AWS.

• Taught a weekend data engineering course for ~100 learners through lectures and hands-on workshops.
• Topics: Basic Data Engineering, Python, SQL, AWS, Git, Workflow Orchestration.
• Prepared and managed AWS cloud infrastructure for data engineering, data science, and data analytics sections.

• Organized and hosted the meetup at the Data Reply Munich office, inviting Seldon as a guest speaker.
• Presented a talk titled 'The Big Picture and Use Cases in Predictive Maintenance'.
• Published recap article on Medium (DataReply publication).

• Taught an intermediate Python course for Munich residents, newcomers, and refugees.
• Guided students through their final projects.

• Founded the Thai localization team and led the translation of 2 of the 9 chapters in the Hugging Face NLP course.

• Taught 450 online students through hands-on lessons in Python, NumPy, Pandas, feature engineering, and model deployment.
• Guided students through computer vision, NLP, and time-series analysis projects.

• Mentored career switchers on data science projects through weekly sessions.
• Donated all compensation to charitable institutions (UNICEF, Wikimedia, ProVeg).

Used SVR, ANN, and LSTM models to predict the stress-strain response of ferritic steel under different conditions. Each curve was treated as time-series data.

Developed multiscale computational tools to model abrasion in agricultural steel as part of a university and industry collaboration funded by the European Commission's Research Fund for Coal and Steel.

Studied and implemented an adaptive control scheme for medical devices and industrial machines.

Developed and optimized a lumped-mass modelling workflow to reduce elasticity in a delta robot.

Simulated the kinematics of a 6-DOF industrial robot arm.