I'm Abira Wisnunggal, a Data Science student working toward AI Engineering. My focus is the part of AI work that happens after the model runs: getting the right context to it, and wrapping the whole thing in a service that holds up.
Most of my projects sit between LLM applications, retrieval, OCR, and document understanding on one side, and backend services and data layers on the other. The part I care about is whether the full pipeline answers correctly on real input, not whether a notebook runs once.
Alongside that I'm learning the operational side: MLOps, LLMOps, system design, and model evaluation. This is still something I'm working through, and most of what I know so far comes from building the projects below.
Build the model. Engineer the system. Make the intelligence useful.
|
AI Engineering Putting models behind APIs, then wiring in retrieval, data, and evaluation. |
LLM Applications Using a language model inside an application rather than as the whole product. |
|
Retrieval-Augmented Generation Chunking, embeddings, and retrieval quality. |
OCR & Document AI Turning PDFs and images into text a model can retrieve from. |
|
Machine Learning & Deep Learning Regression and classification, neural networks, and computer vision. |
MLOps & LLMOps Currently learning: experiment tracking, evaluation, and reproducibility. |
The interesting part isn't a single model. It's the path from raw information to something you can actually call.
model → retrieval → orchestration → backend → data → deployment
LLM · RAG · Embeddings · Vector Databases · Text-to-SQL · OCR · Document AI · LLM Evaluation
LangChain · LangGraph · LlamaIndex · FastAPI
Secondary stack
Web / Application:
TypeScript · Node.js · Express.js · Next.js · Tailwind CSS · Prisma
Cloud / Deployment:
AWS · Google Cloud · Vercel
Development:
Git · GitHub · Bitbucket
Databases / Storage:
MongoDB · pgvector
A few repositories, and the problem each one was trying to solve.
|
A chatbot that routes each question to whichever backend can answer it: document retrieval, Text-to-SQL, or OCR, and renders charts when the answer is numeric. Longer jobs run on async workers. Problem → approach: retrieval, SQL, and document parsing lived in three separate places. Intent routing picks one per request, and async workers keep the slow paths off the request thread. Python · FastAPI · RabbitMQ · Redis · Docker
|
Detects vehicles in video and tracks them across frames, so the same vehicle keeps an ID instead of being counted again on the next frame. Problem → approach: manual traffic observation. YOLOv8 for detection, DeepSORT to hold an identity across frames. Python · YOLOv8 · DeepSORT · OpenCV
|
|
Predicts rainfall from weather-station data, covering EDA, cleaning, feature engineering, model training, tuning, and evaluation. Problem → approach: raw observations taken through cleaning and features to a classification model I could actually measure. Python · scikit-learn · Machine Learning
|
A budgeting workspace that tracks transactions and forecasts cashflow, with an assistant that answers questions about the numbers. Problem → approach: budgets, transactions, and forecasts in one place, with an assistant for questions the charts don't answer directly. TypeScript · Next.js · AI Application
|
The direction I'm working toward:
model → retrieval → orchestration → backend → data → deployment
I want to keep the data science foundation underneath it. That means working across LLM and RAG systems, OCR and vision, and the API and data layers that hold them together.
AI Engineering · Machine Learning · LLM · RAG · OCR · Document AI · Data Science