PROJECTS

13 BUILDS

AI / Analytics2026

SPATIALREHAB

A visionOS companion for early-stage dementia, built on one clinical rule: never let the person feel they got it wrong. Eight short games take a silent cognitive baseline, then the daily exercises rehearse what dementia takes first — making kopi one step at a time, mahjong by the real Singapore rules, and the route home studied on a holographic tabletop before being walked at life size through a neighbourhood meshed from OpenStreetMap. The caregiver dashboard holds the only numbers in the app: one trend line per game, never a grade.

visionOSSwiftUIRealityKitSwift ChartsOpenStreetMap

THE TABLETOP — THE ROUTE HOME

OPEN ↗

P-02

Backend / Security2026

AIRLOCK

A safety gate that stands in front of every package an AI agent installs. Each candidate is detonated in a throwaway Daytona sandbox seeded with honeytoken credentials, read line-by-line by a code model on a rented GPU, and reputation-checked against the live web — then a judge weighs all of it and blocks or passes, in plain language the agent understands.

PythonDaytonaQwen 3.5Nosana GPUDoublewordOxylabs

Agent terminal session where Airlock intercepts a pip install of python-pillow and returns a BLOCKED verdict.

BLOCKED — python-pillow

OPEN ↗

P-03

Mobile / Backend2026

BIOHACKK JOURNEY APP

A full-stack mobile health platform: Singpass and Google OAuth, realtime bidirectional sync across four-plus biometric sources — Apple HealthKit, Omron, smart scales and Visbody body scans — and a phase-aware daily protocol with food, sleep, mood and vitals journaling on top. The Supabase and GCP pipelines behind it ingest 10,000+ readings a day; reworking the schemas and query paths took API responses from 3s to 500ms.

React NativeExpoTypeScriptSupabaseGCPNestJSHealthKit

The nutrition detail screen: a photographed meal identified as strawberry kakigori with creme brulee topping, a health score, and a per-ingredient sodium, salt and energy breakdown.

The quests screen listing daily and weekly missions — 10K steps, mindful moment, workout streak, hydration hero — each with XP and coin rewards and a progress bar.

LIVE — DASHBOARD, PHASES, BODY SCAN

OPEN ↗

P-04

Quant Research2026

PORTFOLIO RISK ANALYTICS

A quant notebook that runs AMZN's 29-year return series through the full risk battery against SPY — Sharpe and Sortino, drawdown curves, beta and alpha, a monthly-return heatmap and a 1,000-path Monte Carlo on the trailing year. A 30% CAGR reads very differently once you plot the drawdown underneath it.

Monte Carlo simulation of 1,000 paths over 252 trading days, with the realised return traced in red through a 95% confidence band.

MONTE CARLO — 1,000 PATHS

PythonquantstatspandasNumPyJupyter

HDB RESALE PRICE MODEL

A linear regression over 287,196 Singapore HDB resale transactions, built as an honest end-to-end scikit-learn pipeline: correlation and multicollinearity screening, leakage checks, imputation kept inside the pipeline, and a held-out test set measured against a mean-predictor baseline rather than against itself.

Scatter plot of resale price against floor area for Singapore HDB flats, with the fitted linear model drawn through it in red.

FIT — RESALE PRICE VS FLOOR AREA

Pythonscikit-learnpandasseabornJupyter

MNIST DIGIT CNN

A convolutional net built from the layer definitions up in PyTorch — two conv blocks with dropout into a 50-unit head, 21,840 parameters total — trained for ten epochs on a CPU across all 60,000 MNIST digits and scored against the 10,000-image held-out set after every one. 92% after the first epoch, 97.46% by the tenth, with a loss curve that floors at 1.46 for a reason worth explaining.

Two MNIST test digits rendered from the notebook — a 7 and a 2 — both classified correctly by the trained network.

PREDICTED — 7 AND 2

PythonPyTorchtorchvisionNumPyJupyter

ETL PIPELINE

In build. A scheduled extract–transform–load pipeline: pull from source APIs, land the raw payloads untouched, then parse, validate and reshape them into warehouse tables a query can trust. Reruns are the design constraint — the same window replayed twice should land the same rows once. Numbers, source and captures go up when it ships.

Schematic of the pipeline: three stages — extract, transform and load — connected left to right, with the sub-steps of each listed underneath.

SCHEMATIC — EXTRACT / TRANSFORM / LOAD

SUB-MS VECTOR SEARCH ENGINE

A from-scratch ANN engine with an HNSW index, SIMD distance kernels, and a lock-free query path. Serves nearest-neighbour lookups over 50M embeddings without breaking a millisecond at p99.

RustSIMDHNSWTokioArrow

DISTRIBUTED RAG ORCHESTRATOR

A DAG-based orchestration layer for retrieval-augmented agents: typed tool calls, speculative retrieval, automatic eval gating, and full request tracing across every hop.

PythonasynciopgvectorTemporalOTel

CONNECTION-POOL AUTOPILOT

An adaptive pooler that models pool saturation as a queueing system and resizes itself from live telemetry — killing tail latency spikes during traffic surges.

GoPostgreSQLPgBouncerPrometheus

STREAMING FEATURE STORE

Exactly-once feature pipelines feeding online inference, with point-in-time correctness and a columnar hot cache that keeps reads in single-digit microseconds.

KafkaFlinkDuckDBRedis

LLM EVAL HARNESS

A reproducible eval framework with statistical significance gating — every model/prompt change ships with a confidence interval, not a vibe.

QUERY PLAN VISUALISER

An interactive tool that turns Postgres EXPLAIN ANALYZE output into a navigable cost graph, surfacing the one missing index hiding behind a sequential scan.

TypeScriptWebGLPostgreSQL