I build models that survive evaluation and backends that people actually use. Email: nikabdulaqmal@gmail.com · 📄 Résumé (PDF)
| 0.933 negative recall |
2,000 hand-labelled msgs |
4 models benchmarked |
1,400+ journal lines |
3.79 CGPA · Dean's List |
- Measure before believing. Every model choice gets a held-out benchmark — including the ones that prove me wrong (mBERT beat my pick).
- The LLM proposes, code verifies. Agents never commit an action (a booking, a tool call) until deterministic Python has checked it.
- Privacy is a pipeline stage, not a policy. PII is stripped before any text reaches a model.
- Online and offline paths. Every stage of my voice agent has a local fallback, so it degrades instead of dying.
- Automate the boring parts. This profile card rebuilds itself nightly from the GitHub API via Actions.
plus Hugging Face Transformers · LangChain · ChromaDB · BERTopic · Celery · Ollama · Quarto
Python · Django 6 · PyTorch · PostgreSQL · ChromaDB · LangChain · Celery
Clinical platform for autism therapy centres: imports WhatsApp exports, links messages to clients, and turns them into sentiment, topics, and grounded answers.
- Sentiment — XLM-RoBERTa fine-tuned on 2,000 self-labelled Malay / code-switched therapy messages.
- Topics — BERTopic (UMAP → HDBSCAN → c-TF-IDF) with Sastrawi stemming, resolving to 5 supervisor-confirmed themes across 266 analysed messages.
- RAG assistant — multilingual MiniLM → ChromaDB → cross-encoder rerank (retrieve 20, answer from 5) over clinic treatment documents.
- Booking agent — LangChain + GLM-5.3 with Redis memory; the LLM proposes, Python verifies availability, only then is it confirmed.
- Privacy — a PII wall strips 54 client/therapist name tokens before any text reaches a model.
- Ops — Celery Beat reminders, role-based dashboards (admin / therapist / doctor / client), DSM-5 specifier approval workflow.
PyTorch · Hugging Face · Quarto · Jupyter
My project shipped XLM-R because "it handles Malay natively." That is a hypothesis, so I tested it on the same 432-message held-out split.
| Model | Accuracy | Macro-F1 | Negative recall | Params |
|---|---|---|---|---|
| mBERT | 0.882 | 0.881 | 0.919 | 178M |
| XLM-R | 0.863 | 0.862 | 0.933 | 278M |
| IndoBERT | 0.509 | 0.046 | 0.899 | 124M |
| DistilBERT | 0.345 | 0.171 | 1.000 | 135M |
- The hypothesis lost: mBERT also won on rojak (0.951 vs 0.926) and pure Malay (0.923 vs 0.863).
- The decision held anyway: in a therapy context the expensive error is missing distress, so XLM-R's negative recall won the requirement — documented as a trade-off, not hidden.
- Two baselines collapsed and are reported as-is.
- Topic modelling defended architecturally (no labels): 42% multi-topic rate, NPMI c_v coherence −0.753, honestly interpreted for 3-token documents.
Swift · Ollama · Deepgram · Accessibility API
Jarvis, but local. Voice in → intent → tool execution → voice out, with an online and offline path at every stage: Deepgram Flux / Apple Speech, GLM API / local Qwen3 8B, Aura-2 / AVSpeechSynthesizer. No Xcode project — Swift Package Manager from the terminal.
Python · Textual · Groq · vector search
Personal retrieval-augmented agent with a real TUI: ingest PDFs, semantic search, cited answers with source tracking, persistent local embeddings. Where I learned retrieval quality — not prompt wording — decides whether RAG works.
12 more projects — e-commerce, data analysis, image processing, learning tracks
| Project | What it is | Stack |
|---|---|---|
| django_ecommerce_stripe | E-commerce with payments and order state | Django, Stripe |
| studentperformanceanalysis | Exploratory analysis + modelling notebook | Python, pandas |
| diet-optimizer | Constraint-based meal planning | Laravel, Blade |
| fuzzylogicvisualization | Interactive fuzzy-inference explainer | HTML, JS |
| DIGITALIMAGEPROCESSING | Coursework: filters, transforms, feature work | Python |
| notesprojectcli | CLI notes app | Python |
| learningdjangoblog · fullstackapplication · javascriptlearningjourney · first_rust_program | Learning tracks kept public on purpose | Django · JS · Rust |
| RAGLearning · RAGPersonalAcademy | RAG bootcamp notes and experiments | Python |
| B.Sc. Computer Science (Honours) | National Defence University of Malaysia (UPNM) · CGPA 3.79 |
| STPM | Science stream · 3.92 |
| Awards | Dean's List · First Class Award, UPNM |
| Certifications | MATLAB Fundamentals & Machine Learning · Django & REST APIs · Full-stack (Tailwind, JS, PostgreSQL) |
| Currently learning | FastAPI + Poetry, vector databases, Swift beyond the terminal |
I'd rather understand a pipeline than copy one. The engineering journal for Sentiri is 1,400+ lines of bugs, wrong hypotheses, and the reasoning that fixed them.