DESKTOP EXPERIENCE
Open the AI lab on a computer.
These chatbot demos are available on desktop. You can explore all the data projects on this device.
Explore data projects →AI LAB / THREE WORKING DEMOS
AI chatbots
Explore the work. Solve a question. Learn something.
The first load downloads about 786 MB plus the runtime. Messages stay in this browser. A recent desktop browser is recommended.
Choose an assistant, then load the model to start.
How these chatbot projects work
A local keyword retriever selects one relevant note. A pinned, quantized Qwen2.5 model uses those notes to generate an English answer in a background worker. Each AI answer shows its reference note for checking. Drafts with unsupported numbers, excessive repetition, or an incomplete SQL join comparison are withheld; the source note is shown and labeled instead. These checks cannot catch every factual error. Conversations are kept in memory for this page session; each assistant has a separate history.
The shop policies are authored demonstration content. The coach uses authored study notes. The portfolio guide uses the published project methods and findings. These are retrieval-assisted prototypes; the small model can misunderstand a question or add unsupported claims.
Hugging Face serves the model files and jsDelivr serves the inference library. Those services receive normal download requests, including your IP address. This app does not send your chat messages to either service. No analytics, API tokens, or chat database are used.
Download and inference speed depend on your connection, memory, and processor. The browser may cache the model. Clearing a chat clears its conversation; clearing site storage in your browser removes cached model files.
Download chatbot source & reproduction guide ↓ · Transformers.js documentation ↗
Read the sample shop policies
Returns and eligibility
Northstar is a fictional shop used for this portfolio demo. An unused item in its original packaging can be returned within 30 days of delivery. Proof of purchase is required. Personalized items and gift cards are excluded. An exchange follows the same eligibility rules. This is a sample policy, not a real merchant offer.
Delivery and late parcels
Sample Northstar policy: standard delivery takes 3–5 business days after dispatch. The dispatch email contains tracking. If a parcel is more than 2 business days beyond the estimated arrival date, the customer should contact the shop support team with the order reference. This demo cannot access tracking or contact the team.
Refund timing
Sample Northstar policy: after a returned item passes inspection, the refund goes to the original payment method in 5–7 business days. A return request alone does not mean a refund was approved. This chatbot cannot inspect a return, approve a refund, or check a payment.
Damaged or incorrect items
Sample Northstar policy: report a damaged or incorrect item within 7 days of delivery. Keep the packaging and take photographs of the item and shipping label. The shop support team reviews the evidence before arranging a replacement or refund. Do not submit personal order details in this portfolio demo.
Cancelling an order
Sample Northstar policy: the support team may cancel an order before dispatch. After dispatch, use the return process if the item is eligible. This demo cannot cancel an order, change an address, or verify dispatch status.
When a person should take over
This portfolio chatbot is a fictional support demonstration. There is no real Northstar support address or order database attached. An actual support product would hand disputed charges, missing refunds, exceptions, and account changes to an authorized human agent through a verified channel.
Read the data interview study notes
Evaluate on future observations
For a forecasting task, train on earlier dates, select the model on a later validation period, and evaluate once on a still later test period. Random splitting can place future information in the training set. Features must be available when each forecast would be issued. Fit preprocessing on training data. Compare the model with a seasonal naive baseline and report errors in the target units.
LEFT JOIN and INNER JOIN
INNER JOIN returns matching rows from both tables. LEFT JOIN keeps every row from the left table and adds NULL values when no right-side match exists. With customers(customer_id) and orders(order_id, customer_id), use customers LEFT JOIN orders ON customers.customer_id = orders.customer_id to keep customers who have no orders. Multiple matching orders produce multiple rows per customer.
Aggregate at the right grain
GROUP BY defines one result row per group. WHERE filters individual rows before aggregation; HAVING filters the groups after aggregation. For orders(order_id, customer_id, amount), a draft query is SELECT customer_id, SUM(amount) AS total FROM orders GROUP BY customer_id HAVING SUM(amount) > 100; Check the grain before joining, because duplicate join matches can inflate SUM. This chat does not execute SQL.
Evaluate an imbalanced classifier
Accuracy alone can hide a model that predicts only the majority class. Precision is TP/(TP+FP); recall is TP/(TP+FN). Review precision-recall performance, class-specific results, calibration, and a baseline. Choose a decision threshold on validation data using the actual cost of errors or review capacity. Measure it once on the test set. A risk score does not establish that an intervention helps.
Identify target leakage
Target leakage occurs when a model receives information unavailable at the prediction time or derived from the outcome. For booking-time cancellation prediction, final reservation status and later booking changes are not valid booking-time features. Split before fitting preprocessing, fit transforms on training rows only, and use grouped or chronological splits when the use case requires them.
Plan an A/B experiment
State the decision, primary outcome, randomization unit, minimum effect of interest, and observation window before running an experiment. Randomly assign eligible units to treatment or control and analyze them as assigned. Report uncertainty and check guardrails. An observational correlation or predictive model cannot by itself establish the effect of an intervention.
Preserve the meaning of missing data
A missing reading is unknown, not necessarily zero. Examine the missing-data pattern and document exclusion or imputation rules. Fit any imputer on training data only. Report coverage alongside model errors. In the electricity capstone, only hours with 60 valid minute readings are used for scoring; incomplete hours are excluded and counted separately.
What is demonstrated here?
Portfolio guide: search over published case-study facts, retrieved references, and local answer generation.
Shop support: policy-grounded replies, separate conversation history, and explicit boundaries around order actions.
Data interview coach: short explanations grounded in SQL and evaluation notes, with the underlying notes available for review.
All three use the same free pretrained model. They are demonstrations built by Djellon Pllana, not deployed customer-service systems or newly trained foundation models.