AI engineer · Tangerang, Indonesia

I build computer vision and LLM systems, from the training loop to the server that keeps them running.

Kernel

In this headline: Your browser runs a 3×3 Sobel edge filter over this headline. Point at the words to move the window.Tap the words to move the window there.
Lead project: Semi-Supervised PPE Detection. Pseudo-labeling lifts YOLOv9 from 0.493 to 0.621 mAP@50 on the same 20% of labels. Read the case study
Status: Open to AI Engineer roles, internships, and research collaborations. kennyvws1@gmail.com

Work

Computer Vision2025Team course project · manuscript public

Semi-Supervised PPE Detection

Finds missing safety gear on construction sites, trained with hand labels on only 20% of the images.

Detects helmets and safety vests in construction-site footage, then marks each worker compliant or in violation. Trained on the public SH17 dataset. The 50-epoch training log and the manuscript are public.

Code and training log Manuscript

mAP@50 by epoch

Repo training log, 20% labeled split. Epoch 50: precision 0.704, recall 0.573, mAP@50-95 0.409. Point at the curve to read any epoch.

View as table

Per-epoch validation metrics
Epoch	mAP@50	mAP@50-95	Precision	Recall
1	0.373	0.220	0.545	0.368
2	0.372	0.221	0.593	0.361
3	0.389	0.231	0.569	0.379
4	0.392	0.232	0.498	0.376
5	0.422	0.255	0.569	0.399
6	0.422	0.257	0.549	0.416
7	0.445	0.267	0.626	0.428
8	0.478	0.292	0.603	0.447
9	0.470	0.289	0.594	0.442
10	0.502	0.311	0.623	0.470
11	0.504	0.313	0.577	0.507
12	0.513	0.321	0.606	0.493
13	0.523	0.328	0.661	0.494
14	0.528	0.332	0.666	0.485
15	0.536	0.336	0.642	0.502
16	0.538	0.344	0.681	0.498
17	0.553	0.345	0.673	0.515
18	0.544	0.349	0.727	0.480
19	0.558	0.355	0.667	0.510
20	0.545	0.353	0.674	0.520
21	0.547	0.351	0.662	0.517
22	0.572	0.367	0.703	0.535
23	0.563	0.361	0.668	0.530
24	0.571	0.371	0.666	0.533
25	0.567	0.363	0.665	0.543
26	0.578	0.372	0.681	0.547
27	0.587	0.381	0.669	0.545
28	0.581	0.381	0.677	0.555
29	0.590	0.382	0.746	0.523
30	0.597	0.390	0.702	0.545
31	0.593	0.386	0.729	0.537
32	0.600	0.392	0.734	0.554
33	0.594	0.392	0.699	0.556
34	0.601	0.393	0.678	0.568
35	0.587	0.384	0.724	0.539
36	0.606	0.401	0.689	0.573
37	0.606	0.398	0.685	0.575
38	0.605	0.395	0.715	0.562
39	0.606	0.395	0.685	0.582
40	0.621	0.407	0.751	0.559
41	0.605	0.394	0.678	0.576
42	0.597	0.397	0.739	0.533
43	0.606	0.403	0.718	0.560
44	0.611	0.403	0.725	0.557
45	0.613	0.408	0.739	0.558
46	0.611	0.408	0.717	0.558
47	0.611	0.406	0.742	0.553
48	0.613	0.408	0.722	0.564
49	0.614	0.408	0.728	0.562
50	0.614	0.409	0.704	0.573

+0.128

mAP@50 gained by pseudo-labeling at the best epoch, with the same hand labels.

Supervised10% labels0.371
Supervised20% labels0.493
Semi-supervised20% labels0.621

0.621 is epoch 40, the best epoch of the log above. The run ends at 0.614 on epoch 50.Ablation study · results/grafik_ablation_study.png

How it works

01Baseline
Train YOLOv9-c on the 20% labeled split.
02Pseudo-label
Run the baseline over the other 80% and keep detections above a confidence threshold.
03Retrain
Train again on labeled and pseudo-labeled data together.
04Comply
Match each PPE box to a person by IoU, then mark that worker compliant or in violation.

LLM / Agents2025Active

FinRAG-ID

Cited answers from Indonesian financial reports.

Ask an Indonesian annual report, prospectus or financial statement a question and get an answer that cites its pages. It runs local-first, and one Docker Compose command brings up Qdrant, Ollama, the API and the UI.

Code

Ingest
pdfplumber pages with metadata, split into sentence chunks.
Retrieve
Sastrawi BM25 and BGE-M3 dense search over Qdrant, fused by reciprocal rank fusion.
Answer
Llama 3.1 8B on Ollama, OpenRouter when Ollama is down, sources tagged [S1] to [Sn].
Evaluate
RAGAS faithfulness, answer relevancy and context precision, plus latency p50 and p95.

Built withPoint at a tool to see the stages it runs inTap a tool to see the stages it runs in

Also built

Data Systems2025Active

Bandar Tracker

Smart-money flow tracker for the Indonesia Stock Exchange.

Reads KSEI ownership data and broker summaries to show where large players are accumulating or distributing, the analysis IDX traders call bandarmologi. FastAPI with async SQLAlchemy behind a one-file Lightweight Charts front end.

FastAPI / Async SQLAlchemy / Pandas / Data Pipeline

Private repository

LLM / Agents2025Active

RAG Evaluation Harness

Measure retrieval before trusting a RAG answer.

A reproducible harness that scores RAG pipelines with RAGAS against local models on Ollama, and tracks faithfulness and context recall across config changes so regressions show up early.

LlamaIndex / RAGAS / Ollama / Evaluation

No public repo yet

ML Systems2024Complete

Tiny Transformer From Scratch

A 12M-parameter decoder-only model, built to learn the internals.

Trained on Indonesian Wikipedia, with tokenizer training, mixed precision, a KV cache and a minimal inference server.

PyTorch / Transformers / Tokenizers

No public repo yet

About

I'm Kenny, a computer science student at Universitas Multimedia Nusantara in Tangerang, working toward AI engineering. I like the unglamorous parts of machine learning: cleaning the data, fighting the training run, getting a model to behave on a real server instead of a demo notebook.

I build on a Mac and ship to a Linux homeserver with Docker, so everything has to work on both. If a tool is interesting and I can host it myself, I will.

Tools

Languages: Python
TypeScript
C++
SQL
Bash
ML / Deep Learning: PyTorch
YOLOv9
scikit-learn
OpenCV
Pandas
NumPy
LLM / Applied AI: Hugging Face
LangChain
LlamaIndex
Ollama
FastAPI
RAGAS
Infra / MLOps: Docker
Linux
GitHub Actions
Weights & Biases
PostgreSQL
Ubuntu

Clean the data. Fight the training run. Ship the server.

Hiring an AI engineer or intern, or looking for a research collaborator? Write to me.

kennyvws1@gmail.com