Hi, I'm Ng Yi Jie.
I engineer rigorous verification layers.

Buying a laptop or PC has become an overwhelming experience for everyday users due to technical specification overload and the rigid nature of traditional e-commerce filters.

• Context-blind Filtering: Conventional multi-select filters fail to interpret nuanced, multi-factor constraints (e.g., "a lightweight laptop for a Computer Science student on an RM3,500 budget that can handle light 3D rendering and has good battery life").
• Information Asymmetry: Non-technical consumers struggle to translate subjective software needs into concrete hardware requirements (RAM, TGP, thermal headroom, display color accuracy).
• The Illusion of Choice: Naive keyword searches often return biased or sponsored results rather than calibrated, value-driven alternatives, leading to decision paralysis or purchase regret.

PickWise shifts the paradigm from passive browsing to an interactive, agentic AI consultant. Instead of relying on brittle keyword queries or naive RAG pipelines, it uses a conversational ReAct agent to conduct a guided needs discovery interview, translating ambiguous constraints into deterministic hardware scoring parameters.

• Multi-Step Agentic Reasoning: Built a ReAct-style tool-calling agent (LangChain create_agent, Gemma 4 via the Gemini API) with 4 tools: search_laptops (the retrieval pipeline below), get_review_evidence (YouTube review excerpts), calculate_custom_apple_price (Apple upgrade pricing) and search_malaysian_market_price (catalog price history plus live retail listings). Responses stream over SSE, and the agent decides per turn whether to search again or answer from the conversation and current shortlist.
• 4-Stage RAG Pipeline: pgvector top-50 cosine recall → rule-based reranking (similarity × budget/weight penalty + purpose/brand bonus) → constraint relaxation (weight, then budget, only when no candidate is viable) → relevance gating, using 768-dimension Gemini embeddings on PostgreSQL. The 0.53 gate threshold was set by spot-checking relevant and irrelevant queries; below it, the agent asks a clarifying question instead of recommending weak matches. Every search logs its gate decision, top score and any relaxation.
• Bilingual Eval Harness: Built an offline evaluation harness over Chinese, English and mixed-language queries, combining an LLM judge with deterministic rule checks, including one that flags RM prices the agent states without tool output behind them across multi-turn conversations. Three repeat runs showed about one item in six varies between runs, so single-run changes are treated as noise until they reproduce.
• Automated Data Ingestion: Built Playwright scrapers (with uploaded-HTML parsing for a site that blocks automation) and an LLM extraction step (Gemma 4 via the Gemini API) that turns raw product pages into structured records, prompted to emit one SKU per listed RAM/SSD upgrade option. Catalog of 370 SKUs (Apple, ASUS, Acer), running in Docker Compose.
• Explainable Scoring Engine: Designed an 8-factor deterministic scoring engine with a per-factor score breakdown for each laptop. Replaced min-max normalization, which scaled factors against outlier catalog extremes, with percentile rank within the catalog.

• Cross-Platform MVP: Built and validated a working cross-platform MVP with a Node.js/Express backend, MongoDB, and LangChain RAG pipeline delivered through a Flutter mobile app.
• Architectural Redesign: Identified scalability bottlenecks and document schema limitations in initial vector search, establishing the foundation to scope and engineer the production-grade PostgreSQL/pgvector architecture in PickWise v2.

• Traceable AI Analysis (Planned): Designed to separate data extraction from AI reasoning: financial figures are collected and stored first, then fed to a self-hosted Large Language Model (LLM) (Qwen3.5) for analysis, so every number the AI uses can be traced back to its source.
• Clean Architecture Backend: C# / .NET backend following Clean Architecture (Domain, Application, Infrastructure, Api layers) with xUnit tests, exposing a REST API through ASP.NET Core Minimal APIs.
• Provenance-Tagged Data: PostgreSQL schema where each quarterly result records the source of its figures (Bursa filing, KLSE Screener's parse, or computed by FinLens).
• Annual Report Extraction (Planned): Python extraction service to pull narrative sections, segment breakdowns and management discussion from annual report PDFs.
• News Scrapers: Scheduled scrapers for financial news (The Edge Malaysia, Nanyang Siang Pau), with relevance flagging for Bursa-related articles.
• Portfolio Module: Pulls positions, trades and fees from the Moomoo OpenAPI.
• Planned: Next.js frontend, a quality-vs-valuation Four-Quadrant Matrix, and a daily policy-driven news briefing.

• Forecasting Pipeline: Built an end-to-end pipeline forecasting the next 30 days of MRT Kajang Line ridership from daily open data (data.gov.my, 2019–2025), using the Kajang Line plus four neighbouring rail lines.
• Leak-Free Features: Engineered 24 past-only features, including lags (1, 7, 14, 30, 60 days), 7/30/90-day rolling means, sine/cosine day-of-week and month encodings, weekend flags and Kuala Lumpur public holidays for every year.
• Model Comparison: Compared a stacked LSTM with a self-attention layer (TensorFlow, 730-day input window, 30-day output), a multi-output Random Forest and a weighted hybrid of the two, tuned on a separate validation set.
• Fair Evaluation: Chronological 80/10/10 split with gaps so target windows never overlap; every model is scored on all 30 forecast days of the same held-out test period (Sep 2024 – Feb 2025).
• Results: The hybrid and the Random Forest beat the seasonal-naive baseline by about 9% and 8% Mean Absolute Error (MAE) (hybrid MAPE 6.83%). The LSTM on its own did not beat the baseline.
• Collaboration: Original university project co-authored with Goh Kian Wee.

• Removed Data Leaks: An audit of the original code found future information leaking into training: a centred seasonal-decomposition trend, scalers and outlier thresholds fitted on the full dataset, and a 2025-only holiday list. Each was replaced with a past-only or training-only equivalent.
• Comparable Scores: The original scored the LSTM and the Random Forest on different targets and tuned and reported the hybrid on the same data. The rebuild scores every model on the same days, on a test set kept apart from tuning.
• Ratio Target: With the leaks removed, both models lost to the baseline because 2024–25 ridership was above anything in training. Predicting the ratio to the same weekday last week fixed this. The corrected results reversed the original conclusion: the Random Forest now outperforms the LSTM.