Jump indexes: Expand an index to find the relevant worked checks and follow-up explanations.
Data source, region and timestamps
Clean records and calculate
Test the model fairly
Reporting, ethics and next steps
How could a student at Singapore Institute of Technology’s Punggol Campus build a responsible AI project using Singapore haze data? First define a specific problem, preserve the official source, build checks for timestamps, regions and units, establish a simple baseline, evaluate any model on genuinely unseen later data, and make the limitations visible. A high-performing model is not automatically a reliable air-quality warning.
The SIT Applied Artificial Intelligence programme is based at SIT Punggol Campus and includes mathematical, computing and AI study. Its Computing Science programme covers computing foundations and analytics. These are credible contexts for a *hypothetical educational project*, not evidence that SIT assigns a particular haze capstone or has approved this project.
The guide will develop 28 original quality and modelling checks and a complete project workflow, without reproducing official examinations or generating unsafe claims about real Punggol air. Use NEA’s live service for actual health guidance: Punggol is part of the North regional reporting area.
A university project starts with a decision, not a neural network
A weak brief says “predict haze with AI”. A strong one asks a measurable question, such as whether a particular validated model improves a short-range forecast relative to a persistence baseline over a held-out historical period.
The team needs a target variable, unit, forecast horizon, geographic reporting area, available input features and evaluation metric. If any are undefined, a model score may be impossible to interpret.
Another reasonable project is data validation rather than forecasting: identifying missing observations, changes in source schema, duplicate timestamps and confusing interface labels. This can be an excellent engineering challenge because it provides objective software tests.
For public-facing outputs, the system should never impersonate an official NEA health advisory. A research chart can explain assumptions but should link the user to the authoritative live source.
Why SIT Punggol is a relevant educational setting
SIT’s published programmes locate its Applied Artificial Intelligence and Computing Science undergraduate degrees at the Punggol campus. They include foundations appropriate to software, mathematics, databases and AI applications.
An interested learner living or studying near Punggol can recognise why air-quality information matters: it may affect family activities, school routines or outdoor work. But the publicly reported regional measurements cover North, not a unique official sensor for each student residence or campus building.
A real coursework project requires the lecturer’s instructions and relevant approvals. The reader should not infer that a blog post grants permission to deploy a sensor, collect personal health information or publish data gathered on university property.
The advantage of a locally relevant project is motivation; its quality still comes from rigorous mathematics and computing practice.
What NEA actually measures and reports
NEA publishes the one-hour PM2.5 concentration and the 24-hour Pollutant Standards Index under distinct definitions. A PM2.5 value is a concentration usually reported in µg/m³, while PSI is an index derived from monitored pollutants.
The data.gov.sg real-time PM2.5 API provides access to regional hourly data, including timestamps and source metadata. A university project should use the current documentation, not assume a previously saved endpoint has an unchanged response structure.
The definition of a Punggol view must remain honest: its geographic label should say North-region ambient data when that is what the source contains.
A blurred screenshot may be useful for a media-literacy exercise but is not an acceptable substitute for source records when training or evaluating an environmental model.
Start with a data dictionary
A minimal row can include a local or offset-aware timestamp, region, pollutant name, averaging period, measurement unit, reported value, data quality status and source reference.
Each field needs a type and validation rule. Is a timestamp allowed to be absent? Are duplicate timestamps legal if they represent different regions? Which values count as missing, and which represent a genuine zero?
Do not create one universal column called “haze number” for PM2.5, PSI and a foreign AQI. These metrics do not have identical units or scales.
A well-defined schema makes the processing pipeline testable. It also means a later analyst can discover why a record was excluded or revised instead of wondering why a dashboard changed without explanation.
An orientation table for SIT students
| Starting strength | First challenge | Next rigorous step |
|---|---|---|
| Mathematics | Units, percentages, sampling | Error metrics and interval interpretation |
| Programming | Parse structured records | Automated schema and timestamp tests |
| Database design | Unique keys and provenance | Versioned ingestion and deduplication |
| Machine learning | Baselines and train-test split | Leakage and drift evaluation |
| Capstone preparation | Research question and scope | Supervisor approval, ethics and reproducibility |
These are learner orientations, not a formal SIT programme semester or compulsory assessment map. The next step should follow actual module prerequisites and project supervision.
Timezones can corrupt a perfectly sensible graph
Singapore Standard Time is UTC+8. A hypothetical observation at 9am in Singapore corresponds to 1am UTC on the same date. An incorrectly converted timestamp can shift readings into the wrong reporting interval.
For joins with rainfall or wind datasets, keep timezone information explicit. An ISO timestamp with an offset is preferable to a local clock string whose timezone is unstated.
One dataset may report the beginning of an averaging window while another reports its end. Before merging, determine exactly which time each source label represents.
An eight-hour shift in a merged table may create an apparent relationship that was never present in the original data. Software tests should catch this rather than allowing it to feed into a model.
Missing values are not the number zero
A missing reading might reflect unavailable service, an invalid record, delayed publication or some other cause. Substituting zero declares that no particulate mass was detected even though the source may have supplied no measurement.
Keep null values and reasons distinguishable from genuine measurements. Report completeness before calculating an average or claiming coverage of a particular day.
If missingness is related to weather, a naïve complete-case analysis can be misleading. The method of handling gaps must be justified by the research goal and source characteristics.
An educational exercise can show how a mean changes after an imputation, but should label which entries were estimated and avoid passing synthetic values off as actual NEA observations.
The first model should be explainable
A persistence baseline predicts that the next reading will be similar to the most recent available reading, under a deliberately simple assumption. It is not necessarily a good final predictor; it establishes a performance reference.
A more complex model should be compared with the baseline using the same held-out time periods, target variable and error metrics. Comparing its training score against another system’s test score is not fair.
If the model does not improve on a simple baseline or becomes unstable during changing conditions, that is a meaningful research result. The project should report it rather than conceal the weakness behind a complex network diagram.
No model should issue activity recommendations that supersede the current official personal health guide.
Why time-series splits must respect causality
For ordinary static tabular tasks, a random training-test split may be reasonable under suitable assumptions. Forecasting has a different risk: records from later times can leak into what the model appears to know at an earlier prediction time.
A chronological holdout keeps the evaluation period later than the training data. Feature engineering should only use input values that would actually be available when the forecast is made.
For example, a 9am prediction must not use a noon PM2.5 observation from the same day as if it were available at 9am. The result may look accurate, but it cannot be reproduced during real-world deployment.
The validation design must match the actual intended decision time.
A model score is not public-health authority
Even if a hypothetical model achieved high classification accuracy, the relevance of that score depends on the label definitions, prevalence, false-positive and false-negative rates, quality of inputs and cost of mistakes.
An experimental alert category is not necessarily an official NEA haze band. The interface must not silently equate synthetic class labels with public-health instructions.
A user should see the data source, observation time, observed-versus-predicted status, uncertainty and a clear route to the responsible agency. When a model is only a classroom prototype, label it plainly.
A responsible system is honest enough to report “not enough data” or “model unavailable” rather than inventing a confident answer.
What the four indexes cover
The indexes at the top of this article divide the material into data integrity, statistical calculations, machine-learning evaluation and project governance. They are intended to be opened as needed rather than read as a single compulsory checklist.
A novice can use the early exercises to practise arithmetic and timestamps. A more experienced programmer may start with schema tests or idempotent ingestion. A machine-learning learner can focus on leakage, baselines and error metrics.
After solving one task, change its inputs and reproduce the answer. Changed-task success is stronger evidence of understanding than copying the original solution.
The final deliverable should be reproducible
A university-level demonstration should preserve its source versions, data dictionary, processing rules, software dependencies, evaluation split, model parameters and test outcomes.
Where personal or restricted information is involved, do not expose it merely to publish a reproducible notebook. A synthetic test fixture and documentation can support authorised verification without disclosing identities.
A reader should be able to determine why a row disappeared, which period a chart covers and what assumptions underlie a forecast. A visually polished output without those answers is not a strong engineering artefact.
The project’s success is therefore the clarity and reliability of its evidence, not simply the number of AI tools used.
Worked Design Check 1: What is the model target?
Question. A project says it predicts ‘haze’. Which elements are still undefined?
Worked answer. Specify a numerical or categorical target, region, time horizon, averaging period, units and intended use.
Why the check matters. Without those labels, performance cannot be checked. A synthetic class called ‘unsafe’ must not be substituted for an official health category.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 2: What region describes Punggol?
Question. An API reports North, South, East, West and Central. The app labels one row ‘Punggol HDB reading’. Is that justified?
Worked answer. No. Display North region if that is what the original source actually reports.
Why the check matters. A more precise geographic label creates false confidence. Local interest does not create block-level data.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 3: Convert SGT and UTC
Question. A synthetic timestamp is 20 October 2026 at 09:00 Singapore time (UTC+8). Convert to UTC.
Worked answer. Subtract eight hours: 20 October 2026 at 01:00 UTC.
Why the check matters. The value is a clock-conversion example. Real records must retain their authoritative timestamp conventions.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 4: Check daily completeness
Question. Twenty-four values are expected but only twenty-one exist. What is completeness?
Worked answer. 21/24 ×100 = 87.5%; missingness is 12.5%.
Why the check matters. Disclose missing hours before presenting any daily summary. Do not silently fill with zero.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 5: Find exact duplicates
Question. A static export has 100 rows, including five repeat entries with identical source keys and timestamps. How many unique rows remain?
Worked answer. Under the stated exact-duplicate rule, 95 unique rows remain.
Why the check matters. Do not delete repeated concentration values that legitimately occur at different times.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 6: Validate the pollutant unit
Question. A synthetic record has field `psi_24h=150` and unit `ug/m3`. Should the system accept it?
Worked answer. Reject or quarantine pending correction; PSI is an index, not a fine-particle concentration.
Why the check matters. Code should validate semantics, not only that 150 is a numeric value.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 7: Flag a negative concentration
Question. A parser returns PM2.5=−8 µg/m³. Does the number represent a valid absolute concentration?
Worked answer. No. An absolute mass concentration cannot be negative.
Why the check matters. A signed change from a reference could be negative, but it must use a different field and definition.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 8: Keep null separate from zero
Question. One sensor reading is unavailable. Should it be stored as concentration zero?
Worked answer. No. Keep a missing-value marker and metadata about availability.
Why the check matters. Substituting zero creates a fabricated clean-air observation and biases later calculations.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 9: Find an arithmetic mean
Question. Three invented hourly values are 40, 60 and 80 µg/m³.
Worked answer. Mean=(40+60+80)/3 = 60 µg/m³.
Why the check matters. A mean of three one-hour values is not the officially calculated rolling 24-hour PSI.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 10: Weight by interval length
Question. A hypothetical concentration is 30 for one hour and 50 for two hours.
Worked answer. Time-weighted mean=(30×1+50×2)/3 = 43.33 µg/m³.
Why the check matters. The weighting assumes the stated piecewise-constant periods; a real sensor may vary within each interval.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 11: Measure missing data
Question. Three of thirty expected records are absent.
Worked answer. Missing proportion=3/30 = 10%.
Why the check matters. A quality dashboard should show the numerator, denominator, expected frequency and missing period.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 12: Compute a rolling summary
Question. For invented values 30, 50 and 70 calculate their three-observation mean.
Worked answer. (30+50+70)/3 = 50 µg/m³.
Why the check matters. The statistic can smooth a graph but may hide a short spike. It does not create a new certified air-quality measure.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 13: Investigate an outlier
Question. Synthetic values are 30,31,29,32,145. Should 145 be discarded because it is inconvenient?
Worked answer. No. Investigate source, time, units and instrument quality first.
Why the check matters. An unusual point can be a real event, a sensor issue or a transfer error. Explain any exclusion.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 14: Find signed sensor difference
Question. A hypothetical reference is 50 and a test sensor reads 58.
Worked answer. Signed difference = +8 µg/m³; relative difference = 16%.
Why the check matters. One comparison does not demonstrate overall accuracy or satisfy a regulatory-grade monitoring requirement.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 15: State a baseline
Question. A team starts with a sophisticated forecasting neural network. What simpler baseline could be tested first?
Worked answer. Use a persistence baseline predicting the next relevant reading from the latest available observation.
Why the check matters. Benchmarking must use the same evaluation period and target as the candidate model.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 16: Use chronological training
Question. A synthetic data record covers 120 consecutive days. Reserve 90 for training and the next 30 for testing.
Worked answer. 90 earlier days for training; 30 later days for evaluation.
Why the check matters. This helps test future-style forecasting and reduces look-ahead contamination.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 17: Find information leakage
Question. A supposed 9am forecast uses the day’s noon observation as an input.
Worked answer. The model suffers future-information leakage.
Why the check matters. Its test accuracy may be artificially strong because the input would not exist at the decision time.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 18: Calculate a residual
Question. A synthetic model predicts 65, while the held-out observed value is 59.
Worked answer. Residual observed−predicted = −6 µg/m³.
Why the check matters. Always establish sign convention before comparing residual series.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 19: Calculate MAE
Question. Residuals in a fictional evaluation are +2,−3,+1,−4.
Worked answer. MAE=(2+3+1+4)/4 = 2.5 in the output variable’s unit.
Why the check matters. Absolute values prevent positive and negative errors from cancelling.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 20: Calculate RMSE
Question. Use fictional residuals +2,−3,+1,−4.
Worked answer. RMSE=√[(4+9+1+16)/4] = √7.5 ≈ 2.739.
Why the check matters. RMSE penalises large errors more strongly than MAE. Compare models under the same conditions.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 21: Calculate percentage error
Question. A model predicts 55 when the fictional observation is 50.
Worked answer. Absolute percentage error=|55−50|/50×100 = 10%.
Why the check matters. Percentage error needs a nonzero meaningful denominator and can mislead near zero.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 22: Calculate accuracy
Question. A synthetic confusion matrix has TP18, FP6, FN2 and TN74.
Worked answer. Accuracy=(18+74)/100 = 92%.
Why the check matters. The invented class labels are not official haze risk thresholds. A high accuracy can conceal poor performance on a rare class.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 23: Calculate precision
Question. Use TP=18 and FP=6 from the previous matrix.
Worked answer. Precision=18/(18+6)= 75%.
Why the check matters. This measures the share of predicted positives that are true positives under the synthetic label definition.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 24: Calculate recall
Question. Use TP=18 and FN=2.
Worked answer. Recall=18/(18+2)= 90%.
Why the check matters. Recall focuses on the portion of actual positives detected and is not identical to precision.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 25: Calculate F1
Question. With precision .75 and recall .90, calculate the F1 score.
Worked answer. F1=2×.75×.90/(.75+.90)≈ 0.818.
Why the check matters. Whether that is useful depends on the task and the consequences of different errors, not a generic league table.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 26: Find publication latency
Question. A record’s observation timestamp is 09:00 and an app receives it at 09:12.
Worked answer. Latency= 12 minutes.
Why the check matters. Show the observation time and update time separately; ‘live’ need not mean instantaneous.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 27: Quantify drift
Question. A synthetic mean changes from 40 to 55 across periods.
Worked answer. Relative increase=(55−40)/40×100 = 37.5%.
Why the check matters. This shift could reflect environment, source changes or data processing. A retraining decision needs investigation.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
Worked Design Check 28: Protect personal data
Question. A student proposes publishing friends’ exact addresses and private medical histories with indoor PM2.5 data.
Worked answer. Do not publish identifying personal data without a lawful, authorised basis and proper governance.
Why the check matters. A demonstrator can use fictional or appropriately authorised data, and the public interest does not override privacy obligations.
Independent retest. Change one valid input, define the expected result before running code, and ask whether the software should accept, reject, quarantine or flag the record. Preserve the source, timestamp and unit rather than forcing the answer into a generic pollution score. Each example is synthetic and is not today’s official Punggol air quality.
A Five-Stage ETL Pipeline a Reviewer Can Audit
Extract: Retrieve authorised data or load an explicitly labelled synthetic fixture, preserving its retrieval date and original fields. Validate: Check schema, time, region, pollutant, unit and duplicate rules. Transform: Create derived variables with an audit trail, including how missing values were treated. Store: Write a versioned table or database without silently duplicating records. Review: Confirm data completeness, unexpected values and provenance before modelling.
A stable ETL process should be idempotent, meaning rerunning an unchanged input does not create a second copy of the same observation. It should also detect legitimate source corrections rather than erase them.
The system needs a defined failure mode. An unfamiliar field, wrong unit or missing timestamp should lead to a traceable warning or quarantine action, not a fabricated default number.
A strong university deliverable includes a data dictionary and a test fixture showing exactly how each problematic record is handled.
How to Compare Models Against an Honest Baseline
Evaluate a model on data that were not available during its training. For a forecasting problem, choose a chronological holdout and avoid constructing features with information from the future.
A persistence baseline is simple but interpretable. A new model should beat that baseline on a suitable error metric and across more than one favourable interval. Compare both on the same target variable, period and units.
Use error measures such as MAE and RMSE carefully. RMSE gives more influence to large errors, while MAE summarises absolute error without squaring. Neither automatically tells a parent whether an outdoor activity is medically suitable.
Report failed experiments as well as successful ones. If a neural network does not improve upon a simple baseline, say so and investigate the reason rather than claim the project has discovered a better haze forecast.
Test the Interface, Not Just the Numerical Model
A family-facing view can mislead even when the underlying computation is correct. Suppose the project displays a predicted concentration without saying it is experimental, labels North as a particular Punggol block or hides the time of the last observation. Each is a communication defect.
Write acceptance tests for the title, pollutant, unit, region, observed-versus-predicted status and official-source link. Test the same interface on a small screen, in low connectivity and with missing data.
Avoid using colour alone to represent categories; add text labels and explanatory context. A blue line and red warning badge can look authoritative even when the underlying information is fictional.
The responsible design should show research demonstration, not official public-health advice and route people to NEA for the current personal guide.
Reproducibility: Make a Second Student Able to Check
A research team should record the project version, dataset or retrieval timestamp, dependencies, processing scripts, validation checks, train-test split and evaluation settings. That information lets another student reproduce the results and identify when a new run differs.
If the underlying data cannot be shared for legitimate privacy or licence reasons, provide a nonidentifying synthetic fixture and document the processing logic. Reproducibility does not justify disclosing protected information.
A release note should describe whether a change affects the schema, a formula, a model parameter or the user-interface wording. The ability to explain why an output changed is more useful than claiming an AI system updates itself magically.
A classroom demonstration is strongest when a reviewer can run the tests, inspect failures and verify the output without needing the original author’s undocumented assumptions.
A Safe SIT Punggol Learning Project Timeline
Stage 1 — Frame: Define the user and question, choose a source, document the geographic coverage and list prohibited claims. Stage 2 — Prototype: Generate synthetic records with known errors and design the validation checks. Stage 3 — Analyse: Produce descriptive summaries and a simple baseline.
Stage 4 — Evaluate: Separate earlier training and later testing periods, review performance across conditions and audit failure cases. Stage 5 — Present: Demonstrate a clear data dictionary, an accessible interface and a limitations section, then obtain any necessary supervisor feedback.
This is a suggested independent learning sequence, not an SIT-sanctioned capstone schedule. Module assessment, project approval, data access and intellectual-property requirements belong to the institution.
A pupil should be able to explain each stage in plain English before moving to a more sophisticated ML architecture.
Research and Ethics Are Engineering Requirements
An educational project that mixes environmental readings with identifiable people’s activities or health details creates privacy and ethical questions unrelated to model accuracy. It may not be appropriate to collect those data at all.
Do not publish home addresses, personal health records, student identifiers or precise movements simply to make a heatmap appear interesting. Seek relevant authorisation for any genuine project involving people or restricted systems.
A project may also affect public perception. If it presents experimental classifications as official warnings, readers could change their activities based on misleading information. The communication design should explicitly protect against that confusion.
Ethics is not a paragraph appended after the machine-learning demonstration. It influences data selection, user expectations, permissions, accuracy standards and the choice of what not to predict.
Frequently Asked Questions
Does SIT require an official haze-AI project?
This is an independent teaching example. The university determines its actual modules, capstones and approval procedures.
Is SIT’s Applied Artificial Intelligence degree based in Punggol?
SIT’s official programme page lists its Punggol Campus for the named degree. Specific module details and arrangements should be checked directly with SIT.
What is the difference between PSI and one-hour PM2.5?
PM2.5 is a fine-particle concentration in a physical unit, while PSI is an index derived under NEA’s method.
Can the software label a North-region reading as the Punggol HDB value?
No. That would add geographic precision unsupported by the source.
Why use synthetic data for initial tests?
They allow known expected outputs and deliberate error cases without misrepresenting public measurements.
Does a high classification accuracy prove the model is useful?
No. Label quality, prevalence, recall, precision, false positives and consequences of errors still matter.
What is data leakage?
It occurs when information unavailable at genuine prediction time enters training or evaluation in a way that unfairly helps the model.
Can the app replace NEA’s health guidance?
No. A research prototype should explicitly link to official current advisories and not offer unsupported personal health rules.
Why does the model need a baseline?
It shows whether extra complexity delivers a meaningful improvement over a simple reproducible method.
What should a final project report contain?
Purpose, source and permissions, data dictionary, quality tests, modelling, results, validation, limitations, ethics and reproducibility instructions.
Continue Through the eduKate Punggol Education Ecosystem
For the earlier skills, read Polytechnic Applied Statistics Through Punggol Haze and ITE Higher Nitec Applied Numeracy.
For real data access, visit the Punggol Haze Python API Tutorial or the Excel and Google Sheets Haze Tutorial. For atmospheric prediction limits, use the separate AI Forecasting Guide.
The Understanding Punggol SIT and Applied University Degrees article covers the educational pathway. The Punggol Haze Complete Guide connects this branch to health, schools, work and neighbourhood life.
Official References and Publication Limitations
Reviewed 20 October 2026. All model scores, sample rows, calculations and timestamp examples are invented for teaching; none are the actual conditions at SIT campus or a certification of its facilities.
- SIT: BSc (Hons) Applied Artificial Intelligence
- SIT: Computing Science Programme
- data.gov.sg: Real-Time PM2.5 API Developer Guide
- NEA: Air-Pollution FAQs and Punggol’s Reporting Region
- US EPA: Enhanced Air Sensor Guidebook
- NEA: Live Haze Information
The best applied AI project does not have to make the boldest prediction. It should make the clearest, most reproducible and most appropriately limited claim its evidence supports.
Connected Punggol Haze Mathematics and Research Guides
Continue from familiar Mathematics and air-quality concepts to ITE workplace numeracy, polytechnic inference, SIT applied AI quality systems and university causal-research methods. These articles use explicitly fictional worked datasets; real outdoor decisions still require NEA’s current North-region guidance.
