01Configured agent runtime
Realtor OS
A lead-engagement operating desk for property agents: inbound WhatsApp and voice leads, assistant-led qualification, bookings, tasks, inventory, cost visibility, and CRM sync.
Read the case study
Wu Yekai · AI data & systems · Singapore
My work has included AI systems—from data and evaluation to the interfaces people use.
I am a computer science student and AI data intern working across LLM training data, multimodal inspection, agent workflows, and decision-support products—from the first system boundary to the tested interface.
01LLM data work
Model-scored curation, quality checks, file conversion, and evaluation pipelines.
02Multimodal inspection
Gated VLM inference, camera-pose geometry, and visual verification.
03Product systems
Agent workflows, interfaces, APIs, automation, and reproducible local setup.
Selected work
Systems with meaningful engineering depth.
These projects are prioritized for technical complexity, system design, evidence of validation, and relevance to product-facing engineering teams.
02Forecasting engine
Jump Cup
A multi-model probability engine for World Cup markets, designed around calibration rather than overconfident winner picks.
03Autonomous · Phase 3
Job Application Assistant
An autonomous job-search pipeline that discovers roles from public ATS boards, filters for relevance and eligibility, tailors resume bullets per role via an LLM, and fills applications through a Playwright session. Every field routes to known, LLM-inferable, or personal-unknown, so it never guesses a personal fact.
04Full-stack AI application
Chess Coach
Engine-backed analysis translated into contextual chess coaching, commentary, tactics, and chat.
View repository
05Multimodal routing engine
MRT + Bus Router
"Leave now" Singapore transit routing over a custom bounded multi-modal search — live LTA bus arrivals, GTFS rail topology, and OneMap walking data — with zero third-party runtime dependencies. Benchmarked against the Google Maps Routes API on 100 checkpointed cases.
06Personal booking automation
NUS Tennis Booker
A personal-use Playwright automation that keeps an authenticated session ready for a university tennis-court booking portal, watches for a court to open up, and hands the payment step to a human over Telegram. A separate scheduled health check alerts over Telegram if the site changes.
Case studies
How the systems are put together.
Notes from the projects themselves: the architecture choices, the parts that were validated, and the limits worth knowing before making a claim.
01
Realtor OS · agentic lead operations
Make assistant behavior configurable without making active conversations unpredictable.
Context
Property agents need a single operating desk for inbound leads, qualification, bookings, follow-ups, inventory, and CRM records—while retaining control when a lead needs a human.
System design
A workflow definition, prompt template, and runtime variables compile to a system prompt plus tools. Threads pin their workflow and prompt versions, so a configuration edit does not change an in-flight conversation.
Operational choices
The runtime owns conversations, tool calls, costs, and audit data; Twenty holds durable CRM records. It runs with local JSON when no database is configured and uses deterministic adapters when SaaS credentials are absent.
Evidence in the repositoryNext.js + Zod contracts, Twilio and voice webhooks, Cal.com and Twenty boundaries, typed tool paths, Vitest coverage, smoke/readiness/route scripts, and GitHub Actions for typecheck + tests.
Status: The documented local demo/runtime path works without external credentials. A public source link needs confirmation because this checkout’s Git remote is owned by a different account.
02
Jump Cup · calibrated forecasting
Forecast probabilities that can be inspected, calibrated, and safely revised.
Context
Probability contests reward estimates that are correct at their stated confidence—not aggressive guesses. The system converts free-form market questions into structured prediction tasks.
Pipeline
An LLM/heuristic parser path classifies the question, routes it to suitable models, ensembles available signals, then applies subtype/family calibration, shrinkage toward a base rate, and probability clamping.
Evaluation
The dashboard exposes the parse path, data match, each model output, and exact pipeline math. The calibration gate compares engine and crowd Brier scores across 198 settled markets from 20 captured matches.
Quality barThe gate requires finite probabilities and encodes a defined crowd-relative Brier tolerance. The project documentation also names current model and data limitations rather than presenting a profitability claim.
03
MRT + Bus Router · multimodal optimisation
Route “leave now” journeys while treating live data, schedule data, and API quotas differently.
Context
Singapore public-transit routing needs bus arrivals, rail topology and schedules, walking access/egress, transfer penalties, and a clear path to compare against Google transit results.
Routing choices
The router builds bus, rail, walking, and transfer candidates; takes live LTA arrival lookups only for realistic boarding stops; uses bounded OneMap pedestrian calls; and falls back conservatively for exploratory edges and timeouts.
Benchmark discipline
A repeatable Google comparison is paced, retry-aware, and checkpointed. The source documents known GTFS incompleteness and limits of scheduled rail waits, avoiding unsupported production claims.
Recorded evidenceAutomated tests cover MRT-only, bus-only, mixed routing, topology loading, benchmark reporting, and rate-limit handling. One five-destination routing-fix snapshot records 4 wins, 1 tie, and 0 losses; the README explicitly says it is not a production win rate.
04
Chess Coach · engine-to-language interface
Pair deterministic analysis with adaptive, context-aware explanation.
Context
Centipawn evaluations and move trees are useful to engines but opaque to learners. The product makes those signals navigable through evaluation, commentary, coach chat, and tactics views.
System design
FastAPI route modules and dependency-managed singleton analysers coordinate chess state, Stockfish, and LangChain/LiteLLM output. The React client keeps board and analysis state in typed components.
Delivery evidence
Docker setup, environment examples, health and API routes, backend/frontend/integration GitHub Actions, and a scheduled security scan provide a clearer path from prototype to operated application.
Capabilities
Comfortable across the path from data to product.
Tools are grouped by the systems they support, not by a self-reported proficiency scale.
- AI & agent systems
- Vercel AI SDK, OpenRouter, LangChain, LiteLLM, structured output, tool orchestration, prompt/version management, evaluation gates
- Backend & APIs
- Python, TypeScript, FastAPI, Next.js, Node.js, Express, REST, webhooks, Zod, Pydantic
- Data & quantitative
- pandas, NumPy, SciPy, scikit-learn, Brier scoring, calibration, backtesting, market data, GTFS routing
- Product & delivery
- React, Vite, Tailwind CSS, HTMX, Supabase, PostgreSQL, SQLite, Playwright, Docker, GitHub Actions
How I work
Build it so it can be inspected, tested, and changed.
- Testable boundaries. Risk logic, routing, normalisers, integrations, and application states are separated into units that can be tested with fixtures or deterministic fallbacks.
- Inspectable decisions. The forecasting engine exposes its prediction trace; research recommendations preserve facts and risk flags; agent workflows record versioning, tool calls, and audit context.
- Honest constraints. The portfolio calls out data gaps, evaluation scope, and status rather than converting partial benchmarks or planned work into claims.
Let’s talk
Open to software engineering, AI engineering, data, quantitative, and product-focused technical conversations.