MuhammadNuril Huda

Founder · Aureum

A conversational personal-finance copilot. It reads balances and market data, then talks through allocation and stock analysis without ever holding anyone's money. I designed the AI orchestration, the analytics, the market intelligence pipelines, and the infrastructure under all of it. The model was never the hard part. Making the numbers trustworthy enough to act on was, which is why the arithmetic lives in code and not in the prompt.

AI orchestration · Financial analytics · Market intelligence · Cloud infrastructure

AI Engineer · Ayclip

Long video in, short clips out. I built and deployed the containerised Python service behind it: face tracking with 9:16 auto-reframing, semantic highlight detection, speech-to-text with word-level subtitle sync, and async processing so one enormous upload never blocks the queue.

Python · FastAPI · Celery · YOLOv11 · MediaPipe · ByteTrack · Whisper · FFmpeg · Docker · AWS ECS/S3

AI Evaluation Consultant & Code Review Specialist · Alignerr

I write adversarial data science and machine learning problems used to test frontier models. Each one ships with a deterministic grader, private held-out fixtures, and a reference solution pinned to a fixed score anchor. A task only counts if the model fails all five scored attempts, so the problems have to resist shortcuts and survive review. Before this I reviewed Python and ML code for logic errors and edge cases.

Python · Benchmark design · Deterministic graders · Rubric evaluation · RLHF/SFT · QA workflows

Freelance Data Scientist & AI Engineer

Machine learning and analysis for individuals and small teams across finance, healthcare, and consumer work. Three that I can describe: gold price forecasting with Random Forest against regression baselines, compared on held-out error before delivery; an Indonesian sentiment system over comments from four beauty brands, weighing a curated lexicon against fine-tuned IndoBERTweet and few-shot LoRA, which flagged the posts where a competitor was pulling stronger positive sentiment; and a hospital booking assistant on FastAPI, Docker, and LangChain RAG that let patients pick a doctor and confirm a slot without an administrator in the loop. Client names stay private.

Python · scikit-learn · IndoBERTweet · LoRA · Hugging Face · LangChain · FastAPI · Docker

AI Trainer & LLM Evaluator · Outlier

Scored RLHF and SFT output against task rubrics across general, technical, and computer science domains, and documented where the reasoning, the facts, or the instruction-following broke down.

RLHF · SFT · Rubric evaluation · Annotation QA

Data Scientist Intern · Data Glacier

Travel demand forecasting on PySpark and GCP, joining flight records with taxi and weather data. I checked the series for seasonal structure before picking a model, using an Augmented Dickey-Fuller test, seasonal decomposition, and ACF/PACF. SARIMA and LSTM then cut held-out MAPE by 15%, and redesigning the schema and storage layout dropped measured pipeline runtime by 90%.

80M+

Flight records processed · PySpark · GCP · SARIMA · LSTM

Data Analyst Apprentice · GoTo Impact Foundation

Defined the learning metrics and wrote the SQL behind BigQuery and Looker Studio dashboards. Program owners used them to spot participants who were falling behind, find the strongest ones, and judge how classes were performing.

SQL · BigQuery · Looker Studio · Dashboard design

AI Engineer Apprentice · Orbit Future Academy

Led a team of four building a deepfake classifier with MTCNN face extraction and InceptionResNetV1 transfer learning, plus the modular preprocessing and evaluation workflow around it.

TensorFlow · CNN · MTCNN · InceptionResNetV1 · Transfer learning

DeciSense

A Python assistant for the repetitive start of any data science job: dataset intake, validation, profiling, model recommendation, evaluation, and reporting. Deterministic checks run first and the language model only speaks once there is something real to explain.

Python · Data validation · Profiling · ML planning · LLM explanations