M. Thufail Alwannabil Samas

Data and AI Professional

M. Thufail Alwannabil Samas

Machine Learning, Data Systems, and Applied AI.

Builds end-to-end data pipelines, classification and forecasting systems, FastAPI services, Docker-based inference, and local LLM workflows with human-in-the-loop controls.

View Work Download Resume GitHub LinkedIn
Surabaya, East Java, Indonesia Career focus: Machine Learning, Data Systems, and Applied AI

About

End-to-End Data, Machine Learning, and Applied AI

Data and AI professional experienced in building end-to-end data pipelines, classification and forecasting systems, and applied AI automation. Delivered Python/PostgreSQL monitoring pipelines, XGBoost and BiLSTM models, FastAPI services, Docker-based inference, and local LLM workflows with human-in-the-loop controls.

Professional experience spans operational DWDM monitoring, BMKG forecasting tools, and programming and AI education. Credentials include BNSP Data Scientist certification and a peer-reviewed XGBoost rainfall forecasting publication as first author.

Selected Work

Applied Data & AI Case Studies

Start with the problem, contribution, and outcome. Open the technical details when you want to inspect the implementation and validation choices.

Project context

Showing all 7 projects.

01

Data engineering & monitoring analytics | Python, PostgreSQL, HTML/CSS/JavaScript

DWDM Optical Sensor Monitoring with Change-Point and Degradation Detection

Problem
Operations teams need a consistent way to separate missing telemetry, short failures, persistent signal shifts, and gradual degradation across approximately 400 configured directional optical sensors whose active subset varies over time.
My Contribution
Built the incremental Python/PostgreSQL pipeline, data-quality and failure rules, change-point and degradation analysis, paired-sensor evidence, local API, priority workflow, and browser-based monitoring interface as a Junior Programmer at PT Aplikanusa Lintasarta.
Technical details
Data & Context
Approximately 400 configured directional optical sensors combine native 5-minute PRTG power telemetry with ENIMS metadata for sensors, links, directions, endpoints, and locations. The active subset can fall below or rise above 300 as sensor availability changes.
Approach
Built an incremental workflow from metadata synchronization and telemetry collection through failure grouping, PELT change-point detection, Theil-Sen and Mann-Kendall degradation analysis, paired-sensor evidence, priority ranking, and local dashboard delivery.
Validation & Controls
Statistical analysis uses only clean observations. Missing data, downtime, 0 coverage, and low-signal failures remain visible to operations but are excluded from change and trend calculations instead of being filled with invented values.
Illustrative DWDM optical signal trace with a change point, degradation trend, failure interval, missing interval, and monitoring output
Illustrative, non-confidential view of signal events, missing intervals, and monitoring output used by the workflow.
~400Configured Sensors
5 minTelemetry Cadence
3Operational Views

Project OutputThe pipeline turns raw telemetry into prioritized, traceable investigation views: failure events, persistent downward shifts, gradual degradation, paired-sensor evidence, and dedicated operations, event, and management screens.

Evidence & ScopeThe analysis helps teams decide what to investigate first but does not claim a physical root cause. Internal telemetry, source code, credentials, infrastructure, database contents, and operational screenshots remain confidential.

03

AI automation & workflow engineering | Python/FastAPI, n8n, PostgreSQL/pgvector, Ollama, Docker

AI Service Request Automation

Problem
Employees often submit short bilingual IT messages instead of complete ticket fields. Service teams must repeatedly infer the issue, catalog owner, priority, missing question, policy, and next action before useful work can begin.
My Contribution
Designed and implemented 1 end-to-end automation spanning web and REST intake, an 8-category IT catalog, local AI analysis and retrieval, deterministic controls, human authority, durable n8n orchestration, Service Desk delivery, notification, retry, and recovery.
Technical details
Data & Context
The controlled local system handles 5 backend request types across 8 employee IT categories using fictional English and Indonesian cases, policies, identities, permissions, approvals, and downstream records.
Approach
FastAPI and PostgreSQL own permissions, state, validation, approvals, idempotency, and audit evidence. n8n coordinates authenticated workflow entries, while local Ollama models provide structured analysis, catalog matching, and grounded policy retrieval behind deterministic and human controls.
Validation & Controls
The frozen final gate completed 40/40 cases. Classification, required fields, retrieval, and citations reached 100%; route and semantic task success reached 96.67%. Of 30 semantic requests, 27 progressed without service-agent triage review and 3 were safely held for review.
AI Service Request Automation workflow from employee request through local AI triage, rule validation, human authority, and Service Desk delivery, supported by Ollama, FastAPI, PostgreSQL with pgvector, n8n, and Docker
Short employee messages become catalog-matched service work while deterministic rules and accountable people retain authority.
40/40Final Cases Completed
100%Classification & Fields
100%Retrieval & Citations
96.67%Route & Task Success

Project OutputDelivered a reproducible Employee Service Desk and authorized Service Operations workspace that turns short IT messages into traceable catalog work, asks at most 1 focused follow-up, and preserves approval, escalation, resolution, retry, and recovery evidence.

Evidence & ScopeThe frozen final gate used 30 untouched bilingual semantic cases and 10 deterministic workflow controls. 90% progressed without service-agent triage review; all services and models cleaned up, and hosted or paid AI calls were 0. Results are controlled local evidence with fictional data, not production or measured business impact.

View repository
04

Data Science | Python, XGBoost, scikit-learn

Rainfall Forecasting with XGBoost and Grid Search

Problem
Rainfall timing matters to salt-production planning in Sumenep, while daily observations are zero-heavy, highly skewed, and difficult to evaluate with random splits. The study forecasts next-day rainfall and examines both amount error and BMKG intensity behavior.
My Contribution
Designed the forecasting workflow from sentinel-code handling and development-only EDA through 23 prior-only features, 3 XGBoost Grid Searches, annual TimeSeriesSplit validation, and multi-regime evaluation, and authored the linked peer-reviewed publication as 1st author.
Technical details
Data & Context
Used 1,461 daily BMKG observation dates from the Trunojoyo area in Sumenep Regency. Codes 8888 and 9999 were treated as unknown values; all 366 final dates received forecasts, with 308 observed targets available for evaluation.
Approach
Compared default and tuned 2-stage XGBoost, direct Tweedie XGBoost, and their development-weighted ensemble through 5,184 temporal CV fits. Development evidence selected a 0.50 wet threshold and a 40% 2-stage plus 60% Tweedie amount forecast.
Validation & Results
Used 2 expanding full-year TimeSeriesSplit checks before a reused final calendar year. The selected ensemble recorded RMSE 9.2521 mm, 5.9% total bias, 0.7679 wet/dry balanced accuracy, 75% BMKG category accuracy, and 0.4429 BMKG Macro-F1.
Actual rainfall where observed and the selected daily XGBoost ensemble forecast across the reused 366-day final calendar period
Actual and selected ensemble forecasts across all 366 final dates; error metrics use the 308 dates with observed rainfall.
9.2521 mmReused-Final RMSE
0.7679Wet/Dry Balanced Accuracy
75%BMKG Category Match

Project OutputThe selected experiment produces 366 next-day forecasts with wet/dry alerts and BMKG intensity labels, supported by development-selection tables, calibration, error-regime diagnostics, feature importance, and a reproducible local evidence bundle.

Evidence & ScopeControlled evidence comes from 2 annual development checks and 1 reused final year. Correct-category counts were 4/15 for moderate and 0/3 for heavy rainfall, which remains the main model limitation. Original observations remain private; the displayed metrics come from the documented forecasting experiment.

View repository Read publication
05

Clinical classification | Python, XGBoost, scikit-learn

Thyroid Cancer Recurrence Classification with XGBoost, Information Gain, and Grid Search

Problem
Build a reproducible recurrence classifier while testing whether a smaller set of informative features can preserve performance and reduce unnecessary model input.
My Contribution
Developed the Python V1 workflow for data audit, mixed-type preprocessing, fold-local Information Gain selection, checkpointed XGBoost search, and multi-metric evaluation for this collaborative academic project.
Technical details
Data & Context
The public UCI file contains 383 differentiated thyroid cancer rows with clinical, pathological, treatment, response, and recurrence variables. V1 removes 19 exact full-row repetitions and retains 364 rows.
Approach
Handled numeric, binary, ordinal, and nominal features explicitly, then combined entropy-based Information Gain selection with XGBoost and a checkpointed Grid Search path.
Validation & Controls
The completed V1 design uses shuffled, non-nested 10-fold K-Fold evaluation and repeats Information Gain selection inside each training fold. It achieved 97.25% pooled out-of-fold accuracy, 95.37% F1, 98.05% specificity, and 98.6% ROC-AUC.
Top 10 encoded features using their original column names, ranked by mean Information Gain across 10 training folds
Top 10 encoded features by mean Information Gain across 10 training folds.
97.25%OOF Accuracy
95.37%F1 Score
98.6%ROC-AUC

Project OutputThe completed V1 workflow transforms 383 source rows into a 364-row modeling set after removing 19 exact full-row repetitions, then combines fold-local feature selection with exhaustive XGBoost search and multi-metric out-of-fold evaluation.

Evidence & ScopeThe reported metrics belong to a shuffled, non-nested 10-fold academic evaluation whose selection criterion was accuracy. They support a reproducible classification study, not clinical diagnosis, patient care, or deployment claims.

View repository
06

Deep learning & time-series forecasting | Python, TensorFlow/Keras, NumPy, Pandas

Rainfall Forecasting with BiLSTM and Grid Search

Problem
Next-day rainfall depends on changing weather patterns across several recent days, making a sequence-based model more appropriate than treating each observation as an isolated row.
My Contribution
Developed the Python forecasting workflow during the BMKG internship, using NumPy and Pandas for 7-day sequence preparation and TensorFlow/Keras for 3-layer BiLSTM training, focused Grid Search, chronological evaluation, and result reporting.
Technical details
Data & Context
Daily BMKG observations from the Trunojoyo observation area in Sumenep Regency provide 4 weather variables and historical rainfall for each input sequence.
Approach
Built 7-day multivariate sequences, fitted Min-Max normalization on the training period, and trained a 3-layer stacked BiLSTM while tuning units, batch size, and the learning-rate drop period.
Validation & Controls
Evaluated 27 configurations on an 80:20 chronological split so every holdout observation followed the training period. The selected 3-layer model recorded MAAPE of 0.8073 and RMSE of 10.2734 mm; impossible negative predictions were clipped to 0.
Observed rainfall and next-day BiLSTM predictions over the chronological test period
Observed rainfall and next-day predictions over the chronological test period.
0.8073MAAPE
10.2734 mmRMSE
27Configurations

Project OutputThe workflow turns 7 days of multivariate weather history into next-day rainfall estimates, preserving the selected model, holdout predictions, evaluation summaries, and actual-versus-predicted visual for review.

Evidence & ScopeThe reported internship metrics use the original BMKG observations; the public repository provides a runnable same-schema sample. Because the same chronological holdout selects and reports the winning configuration, the result is model-selection evidence rather than an untouched final estimate.

View repository
07

Statistical time-series forecasting | Python, statsmodels, Pandas

Rice Price Forecasting with ARIMA and Walk-Forward Validation

Problem
Price planning benefits from a forecast that is easy to explain and tested in the same order that future monthly prices would actually become available.
My Contribution
Implemented the Python/statsmodels ARIMA workflow for this collaborative academic project, covering Pandas-based monthly preparation, stationarity checks, order selection, diagnostic screening, expanding-window evaluation, and forecast reporting.
Technical details
Data & Context
60 public monthly observations from January 2020 through December 2024 cover medium and premium rice prices from the Open Data Jawa Timur portal.
Approach
Modeled medium and premium prices separately with ARIMA, using ADF-based differencing, ACF and PACF inspection, candidate evaluation, convergence checks, and residual diagnostics.
Validation & Controls
Started with the first 48 observations, forecast the next month, added the newly available actual price, and repeated the process across the remaining 12 months. The selected models recorded 2.08% MAPE for medium rice and 1.82% for premium rice.
Observed and 1-step-ahead forecast prices for medium and premium rice during walk-forward evaluation
Observed and 1-step-ahead forecasts over the walk-forward test period.
2.08%Medium Rice MAPE
1.82%Premium Rice MAPE
12Walk-Forward Steps

Project OutputThe workflow turns 60 public monthly observations into explainable 1-step-ahead forecasts, selected ARIMA orders, diagnostic records, error summaries, and comparison visuals for both rice categories.

Evidence & ScopeThe public dataset and runnable workflow are included in the repository. Candidate orders are ranked and reported on the same 12-month walk-forward period, so the selected result is not an untouched final estimate. The univariate models use past prices only.

View repository

Publication

Peer-Reviewed Rainfall Forecasting Study

1st Author | Peer-Reviewed Article

Prediksi Curah Hujan di Kabupaten Sumenep Menggunakan Metode Extreme Gradient Boosting (XGBoost) dan Algoritma Grid Search

M. Thufail Alwannabil Samas, Nurissaidah Ulinnuha, Moh. Hafiyusholeh

MIND (Multimedia Artificial Intelligent Networking Database) Journal, Vol. 11, No. 1, June 2026, pp. 30–43

Open DOI

Experience

Professional Experience

Junior Programmer

PT Aplikanusa Lintasarta

Developed a project-stage DWDM optical monitoring and assurance analytics system for approximately 400 configured directional optical sensors, integrating 5-minute PRTG telemetry with ENIMS network metadata through an incremental Python/PostgreSQL pipeline, failure detection, statistical change-point and degradation analysis, paired-sensor evidence, and 3 browser-based operational views.

Coding and AI Teacher

SMP GIKI 3 Surabaya

Taught Python fundamentals and introductory AI concepts to Grade 8 and 9 students over 5 months, using Teachable Machine and practical coding exercises while designing lesson materials, quizzes, and image-classification activities to strengthen programming logic, problem-solving, and foundational AI understanding.

Programming Teaching Assistant

Universitas Islam Negeri Sunan Ampel Surabaya (UINSA)

Guided undergraduate students across 4 semesters of Basic Programming and Programming courses in MATLAB and Python, providing hands-on support in debugging, data handling, computational problem-solving, and algorithmic thinking while developing introductory machine-learning materials covering Linear Regression, K-Nearest Neighbors, and K-Means clustering.

Data Science Intern

Stasiun Meteorologi Maritim Kelas II Tanjung Perak Surabaya (BMKG)

Developed an internal extreme-weather monitoring dashboard, processed and compared BMKG cloud-prediction imagery with Himawari IR observations using Python, and built a 3-layer BiLSTM rainfall forecasting workflow using 7-day sequences and an 80:20 chronological split, achieving MAAPE of 0.8073 and RMSE of 10.2734 mm.

Skills

Capabilities Across Data and Applied AI

Programming & Query Languages

Python, SQL, R, and MATLAB

Data Engineering & Databases

Pandas, NumPy, PostgreSQL, MySQL, pgvector, psycopg, data ingestion, data validation, data transformation, and incremental data pipelines

Machine Learning & Statistics

Scikit-learn, XGBoost, TensorFlow/Keras, statsmodels, ARIMA, feature engineering, statistical analysis, model evaluation, cross-validation, and time-series analysis and forecasting

AI Engineering & Automation

FastAPI, n8n, Ollama, RAG, embeddings, vector search, structured LLM outputs, REST APIs, human-in-the-loop workflows, idempotency, and retry and recovery workflows

Development, Delivery & Visualization

Docker, Docker Compose, Git, GitHub, GitHub Actions, CI/CD, unit testing with unittest, Jupyter Notebook, model serving, batch inference, Matplotlib, HTML, CSS, and JavaScript

Education

Bachelor of Mathematics (S.Mat.)

Universitas Islam Negeri Sunan Ampel Surabaya (UINSA)

August 2021 – January 2025 | GPA: 3.46/4.00 | Completed the 4-year bachelor's program in 7 semesters (3.5 years)

Thesis: Rainfall Forecasting in Sumenep Regency Using XGBoost and Grid Search.

Certification

Data Scientist

Badan Nasional Sertifikasi Profesi (BNSP)

Valid September 2024 – September 2027

Contact

Data & AI Opportunities

Open to Data Scientist, Data Analyst, Data Engineer, Machine Learning Engineer, AI Engineer, and AI Automation opportunities, including remote roles.

mthufailalwannabils@gmail.com
Surabaya, East Java, Indonesia
LinkedIn GitHub Resume