Bryan Dale.

Local demo · Streamlit · synthetic chatCode ↗

Customers ask for shipping quotes on WhatsApp in short, scattered messages — in Indonesian, Mandarin or English, often with a photo instead of an item name. This pipeline turns the chat into a structured order, prices it from the rate card, and drafts a reply that asks only for what is still missing. It runs free on a laptop CPU with an open-source model (Qwen3.5 9B on Ollama) and is evaluated on real, anonymized customer chats from Pelita Cargo.

• Model reads, code countsThe LLM only copies numbers and units as written (斤, mm, m). Unit conversion, CBM, minimum charges and the price are tested Python, which also flags impossible answers.
• Field order mattersWith the box count before the size in the JSON schema, “65 × 45 × 50 cm” came back as 65 boxes. Putting the size first fixed it; more prompt examples did not.
• Photos helpReading the customer’s photo first raised item-type accuracy from 65% to 76–82% on 17 chats — a small test, not a held-out one.

Overall

208 of 231 details customers stated are read correctly (90%) · the weak spot is the admin’s pricing calls — tariff category and branded rate

Evaluation

Real chats anonymized by rule and checked by hand, never committed · 8 prompt versions compared field by field · 81 automated tests that run without the model

Tried & dropped

Adding similar past admin decisions to the prompt (retrieval with local embeddings) did not beat the plain prompt, so it stays off by default.

Stack

PythonOllamaQwen3.5 9BPydanticStreamlitpytest

gastro-tmb-tmb-prediction.vercel.appOpen ↗

A screening tool that predicts whether a gastric cancer tumor is TMB-High or TMB-Low from the mutation status of 10 genes. A Random Forest built from 774 patients across five cBioPortal cohorts powers a bilingual (Indonesian / English) web app that shows the prediction’s confidence, each gene’s contribution, and a therapy recommendation.

1. 5 cBioPortal cohorts
2. 10-gene mutation features
3. Random Forest
4. Flask backend
5. TMB-High / Low + therapy

Results

On a held-out 20% test set: balanced accuracy 0.82 vs. 0.50 for a majority-class baseline · ROC-AUC 0.93 · catches 17 of 23 TMB-High patients

Stack

PythonFlaskscikit-learnpandasHTMLCSSJavaScript

A backend for property listings, split into three Spring Boot services: one stores listings, one stores users, and a public API combines the two for outside clients. Each service owns its own database; the public API has none and gets everything from the other two over REST.

1. Client request
2. Public API · no database
3. Listing service · one page
4. User service · all owners in one call
5. Listings with owners

• BatchedA page of listings always costs two internal calls, however many owners it has.
• FailuresA service that is down or times out gives a clear 503, a server error gives 502, and a rejected request passes its 4xx and message through.
• IsolatedDocker Compose runs all three services but exposes only the public API.

Tests

29 automated tests with JUnit, MockMvc and MockRestServiceServer · the public API tests run without the other services

My part

Built the user service and the public API, with their tests and the Docker setup. The listing service was given as starter code.

Stack

Java 17Spring BootSpring Data JPAFlywayH2JUnitGradleDocker Compose

pelitacargo-track-dashboard.vercel.appOpen ↗

A shipment tracking dashboard for China–Indonesia freight forwarding, modeled on how Pelita Cargo works. Customers search the manifest and follow each shipment through six stages, from pickup to delivery; admins add shipments and move them forward. Data lives in Supabase, and the cost estimator uses Pelita Cargo’s real all-in rates — per CBM for sea freight, per kg for air.

• TrackSearch by waybill or client and see a timestamped stage timeline.
• AdminAdd shipments by marking code, volume and rate, then update their status.
• EstimateCalculate all-in shipping cost from real per-CBM and per-kg rates.

Stack

React 19ViteReact RouterSupabasePostgreSQLCSSVercel

A comparative study of six models from three families — classical machine learning, a BiLSTM and pre-trained Transformers — classifying five emotions in Indonesian tweets (EmoT from IndoNLU, 4,401 tweets). The models are ranked on macro-F1, because the five classes are imbalanced.

• Domain beats sizeIndoBERTweet, pre-trained on Indonesian tweets, beats the general-domain IndoBERT: data close to the task mattered more than model size.
• Small dataOn 4,401 imbalanced tweets, Complement Naive Bayes outscores the BiLSTM — a strong baseline when data is limited.
• Not luckOn the same 440 test tweets, IndoBERTweet’s lead over Naive Bayes holds up: if the two were equally good, a gap this big would show up less than 1% of the time (McNemar’s test, p = 0.0066).

My part

Exploratory data analysis and slang normalization — mapping colloquial Indonesian to standard words with a slang lexicon before any model sees the text.

Stack

Pythonscikit-learnPyTorchHugging Face TransformersTensorFlowFastTextGradioJupyter

bryandalee.pythonanywhere.comOpen ↗

A monitoring system for the distribution of medical equipment to public health centres. The provider’s admin keeps working in the Excel file they already use, uploads it, and checks a preview before anything is saved; from there the app tracks every unit through eight delivery stages and shows each role — the provider, the health office and the health-centre staff — only what belongs to them.

• ImportUpload the working Excel file and review it row by row before it is written.
• RolesSeparate views for the provider admin, the health office, and each health-centre PIC.
• TrackingEach unit moves through eight delivery stages, updated automatically.

Background

Built from a real brief for a medical-equipment provider. The contract fell through, so it lives on as a public demo with sample data only.

Stack

DjangoPythonPostgreSQLopenpyxlHTMLCSS

Named-entity recognition for crypto assets in Indonesian tweets. Distant supervision labels 25,726 tweets against a CoinGecko knowledge base, heuristic denoising filters out everyday words that share a ticker, and an IndoBERTweet-CRF model is then fine-tuned on 1,461 manually annotated tweets.

1. Indonesian tweets
2. Distant supervision labels
3. Heuristic denoising
4. IndoBERTweet-CRF + gold fine-tuning
5. Crypto entities

Results

Entity-level F1 85.4% (precision 83.3%, recall 87.7%) in 5-fold cross-validation · annotator agreement Fleiss’ κ 0.97

My part

Worked with the team on the data pipeline — preparing the tweet data, distant-supervision labeling, heuristic denoising — annotated the Gold Standard as one of three annotators, and co-wrote the paper.

Stack

PythonPyTorchIndoBERTweetCRFJupyter