Josiah Lau | Applied AI Engineering

Portfolio Map

Five implemented systems across one AEC decision journey

The portfolio follows information as it moves from source documents and project conversations into design options, quantities, and simulated site action. Each project remains independently runnable; one tested integration proves a bounded specification-to-cost handoff.

Selected project evidence Complete project index Evidence-linked skills Projects by role Two-page portfolio brief

Selected Work

Three projects, three kinds of engineering evidence

The strongest work combines implementation depth, measured behavior, visual explanation, and explicit failure boundaries. Metrics come from fixed public-source snapshots, bundled synthetic fixtures, or simulator holdouts as labeled beside each result.

Local AEC retrieval interface with a grounded answer and source citations

01 / Flagship

AEC Code Compliance RAG

Public-source ingestion, four retrieval modes, citations, abstention, evals, and focused tests.

Document retrieval is stronger than exact evidence retrieval Document Hit@1 0.952 / MRR 0.976; exact-target Hit@1 0.810 / MRR 0.881; page Hit@1 0.778 on manually labeled fixed-snapshot targets. Public Hit@1 95% interval [0.773, 0.992], n=21. No-answer is 2/2; this is not compliance certification or broad accuracy evidence. Read visual case study

02 / Embodied AI

Construction Embodied Agent Simulator

Shared expert demonstrations, rendered-pixel shift tests, raw and filtered rollouts, and planar MuJoCo command replay.

0.760 filtered success / 51 raw contacts / 0 filtered contacts 96 unseen grids for policy evaluation; 12-scenario physics subset with named rigid-contact telemetry. Rendered-RGB success falls from 0.719 to 0.427 under an unseen palette. Pixels come from state; no camera or hardware evidence. Read visual case study

Constraint-aware massing plan, isometric view, and transparent proxy measurements

03 / Computational Design

Constraint-Aware Massing Explorer

Seeded geometric options, hard constraints, Pareto ranking, editable objectives, and transparent environmental and access proxies.

0.977 feasible rate / 0.19% mean best GFA error 864 candidates per method across synthetic sites; unconstrained baseline feasible rate 0.052 and mean best GFA error 34.47%. Rectangular envelopes and transparent proxies only; no code inference, calibrated environmental simulation, or approvable design. Read visual case study

Measured target-reach and rigid-contact rates for raw, filtered, and reference embodied-agent command traces — Headless planar MuJoCo replay: raw egocentric commands contact named rigid geometry on 51/150 movements; filtered egocentric and A* traces record 0 contacts across 148 and 98 movements. Not a mobile-robot or safety validation.

One-Command Technical Check

python scripts/reviewer_check.py

Checks public claims, documentation commands, links, app titles, and focused tests for the three primary projects without rewriting tracked evidence.

View the artifact ledger

Executed AEC Integration

Approved requirements to a bounded schematic cost input

A deterministic local runner connects the specification, massing, and QS interfaces. Every conversation, site value, generated option, and rate in this trace is synthetic.

Executed AEC system journey with two input streams, source and approval gates, typed project handoffs, traceability, and professional review — 5 approved / 3 mapped / 2 retained; 16/16 sourced fields; 96 candidates / 92 feasible / 42 storey-matched; 7/7 priced takeoff lines / tender not_run.

Budget and accessibility remain outside automation. The QS handoff covers one schematic ground-floor envelope, not a developed plan or whole-building estimate.

python integrations/aec-design-to-cost/run_workflow.py Inspect contract, tests, and trace

Role-tagged project messages, requirement ledger, and source-linked draft clauses

Project Communication

Communication and Specification Assistant

Shared role-tagged chat, versioned requirements, conflict detection, scoped approvals, audit events, and source-linked draft clauses.

33-case language stress: 0.958 F1 / 0.939 exact 5 conversations / 35 messages / 1.000 fixture F1; 2 number-word misses are retained. Manually labeled synthetic cases, not open-domain chat or a professional specification. Read visual case study

Constraint-aware massing plan, isometric option, and proxy metrics

Computational Design

Constraint-Aware Massing Explorer

Hard geometry checks, baseline comparison, Pareto ranking, and inspectable proxy scores.

No code inference, internal egress, calibrated simulation, structure, or professional design output. Read visual case study

Vector quantity takeoff, cost build-up, and synthetic tender comparison

QS Automation

Takeoff and Tender Analysis Workbench

Shared-wall measurement, opening deductions, rate provenance, uncertainty bands, and line-level tender exceptions.

21/21 quantities / 5/5 tender exceptions No PDF/CAD/BIM parsing, live pricing, professional QS output, or award recommendation. Read visual case study

Capabilities

Domain reasoning backed by inspectable engineering

Architecture contributes the problem framing: constraints, alternatives, coordination, review gates, and accountable handoff. Software and ML evidence still come from code, fixed datasets, tests, and reproducible artifacts.

01

Source-grounded AI

Page-aware ingestion, metadata-rich chunks, retrieval baselines, citations, abstention, and evaluation.

02

Embodied AI

Observation encodings, behavior cloning, appearance shift, action filtering, and planar physics replay.

03

Computational design

Parametric geometry, hard-constraint validation, Pareto ranking, and explicit proxy objectives.

04

AEC workflow systems

Requirements, approvals, source lineage, quantity provenance, cost build-up, and human review boundaries.

05

Engineering practice

Typed APIs, SQLite state, automated tests, generated evidence, idempotence checks, and local-first delivery.

Open the evidence-linked skills matrix

Primary Project

AEC Code Compliance RAG

This is the most developed project in the portfolio. It applies built-environment domain knowledge to Markdown and page-aware PDF ingestion, Singapore public-source manifests, source filters, TF-IDF, BM25, dense LSA, hybrid retrieval, citation checks, unsupported-scope handling, and a tested local service with bounded durable telemetry.

Evaluation: 51 synthetic cases plus 24 authored public-source cases over a fingerprinted 15-document snapshot
Synthetic check: 51-case synthetic set: Recall@4 1.000 / MRR 0.906 / Hit@3 1.000
Evidence location: Document Hit@1 0.952 versus exact manually labeled target Hit@1 0.810; source-page Hit@1 0.778
Uncertainty: Public Hit@1 0.952 [0.773, 0.992], n=21; MRR 0.976 [0.929, 1.000]. Hybrid-vs-BM25 MRR delta 0.012 [0.000, 0.036], inconclusive
Service: 12/12 local contract checks passed; 9 requests observed before metrics response. The fixed workload returned 48/48 HTTP 200 responses, 0 server errors, and P95 at or below 500 ms
Persistence: 48 payload-free query telemetry rows remained after app reconstruction
Proof: Architecture, eval design, paired uncertainty, ablation, failure analysis, demo answers, service reports, and tests
Limit: No authority validation, professional sign-off, OCR, network load, uptime, external deployment, traffic, or live amendment guarantee

Visual case study Project Evaluation Tests Service tests Service contract Reliability evidence Uncertainty evidence Target-label audit

Local AEC RAG Streamlit interface showing an answer and source citations — Actual local Streamlit run using labeled demo data.

Document-level, exact-chunk, and source-page retrieval metrics for the fixed AEC public-source snapshot — Document Hit@1 0.952 / MRR 0.976; exact-target Hit@1 0.810 / MRR 0.881; source-page Hit@1 0.778 / MRR 0.861. Finding the expected publication is easier than retrieving the labeled supporting passage. Exact targets were authored from the fixed local snapshot and have not been independently reviewed.

Point estimates and 95 percent intervals for the authored AEC public-source retrieval snapshot — Public Hit@1 0.952 has a 95% Wilson interval of [0.773, 0.992] over 21 answerable cases; MRR 0.976 has a 95% bootstrap interval of [0.929, 1.000]. No-answer 1.000 is 2/2 with a Wilson interval of [0.342, 1.000]. The intervals resample this fixed authored set; they are not external-validity claims. Exact-target uncertainty: Hit@1 0.810 [0.600, 0.923] and MRR 0.881 [0.762, 0.976] across 21 manually labeled targets.

Synthetic construction grid, rendered RGB crops, and measured holdout results across four observation families — Generated from evaluator output. RGB pixels are rendered from privileged simulator state, not a physical camera.

Concept illustration of a construction robot following a planned safe route — Generated concept image. It is not a simulator screenshot or hardware evidence.

Embodied AI

Language, state, action, and visible failure

The simulator maps task instructions and structured construction grids to closed-loop actions. It compares engineered-state, semantic-raster, agent-centered local-state, and rendered-RGB imitation on the same demonstrations and holdout layouts. An unseen palette exposes visual brittleness.

Policies: Random forest plus world-raster, egocentric-state, and rendered-RGB MLPs fitted from shared A* demonstrations
Holdout: 96 disjoint procedural scenarios with raw and filtered closed-loop rollouts
Finding: Egocentric action accuracy 0.834 versus world-raster 0.478; filtered success 0.760 with 943 interventions
Pixel ablation: Rendered-RGB action accuracy falls from 0.813 to 0.474 when pixels are replaced by training means
RGB shift: Action accuracy 0.813 to 0.417; raw success 0.635 to 0.000; shifted filter interventions 3315
Physics replay: 12 unseen scenarios; raw contact rate 0.340 versus filtered 0.000 in a planar MuJoCo command adapter
Limit: Pixels are rendered from privileged state and the filter has full rules; physics uses a planar body and static proxies, not a physical camera, mobile robot, ROS, hardware, or safety validation

Visual case study Project Evaluation Tests

Architecture Background

Design discipline carried into AI engineering

Architecture provides the domain context for the AEC and embodied-AI work: requirements, constrained alternatives, versioned information, human review, and accountable handoff. It does not replace software evaluation or robotics validation.

Read the project statuses and image provenance

Bamboo Factory thesis visualization — **Bamboo Factory**2019 M.Arch thesis; AI-assisted visualization anchored to original drawings.

Tideline Commons adaptive-reuse concept — **Tideline Commons**Self-initiated 2026 concept; generic context, not commissioned or built.

Monsoon Works steel-hall concept — **Monsoon Works**Self-initiated concept; not structural design or a fabrication package.

Professional Delivery

Experience across design, coordination, tender, and handover

The AI work is informed by direct responsibility for drawings, project information, stakeholder coordination, procurement, progress, defects, and delivery. These roles provide domain context; the repository remains the evidence for software and ML claims.

2024 - presentDirector / 3JAI StudioAI workflow systems, review criteria, and client delivery.
2022 - 2024Project Manager / People's AssociationTender, cost, progress, defects, audit, and handover.
2020 - 2021Architect / HDBGIS, estate planning, environmental studies, and agency coordination.
2020Graduate Architect / aKTa-RCHITECTSAuthority submissions and construction packages.
2018 - 2019M.Arch Intern / Tierra DesignGrasshopper CAD automation and drawing QA.

Singapore M.Arch MSc Design + AI Revit + Rhino/Grasshopper + Python

Project Index

A direct route into every implemented system

Five selected projects carry the portfolio narrative. Eight smaller experiments test focused engineering patterns without inheriting flagship claims.

Flagship / Source-grounded AIAEC Code Compliance RAGRetrieval, citations, abstention, service contracts, telemetry, evals, and tests. Primary / Embodied AIConstruction Embodied Agent SimulatorState and rendered-pixel policies, holdouts, filtering, failure analysis, and MuJoCo replay. Primary / Computational DesignConstraint-Aware Massing ExplorerGeometric search, hard constraints, Pareto ranking, proxy metrics, and visual outputs. Supporting / Project RequirementsCommunication and Specification AssistantRole-tagged messages, conflicts, scoped approvals, audit events, and language stress tests. Supporting / QS AutomationTakeoff and Tender Analysis WorkbenchVector measurement, rate provenance, uncertainty bands, and tender exceptions. Tested IntegrationApproved intent to bounded cost inputSource-linked requirements, typed handoffs, rejection tests, and one checked-in synthetic trace.

Eight focused experiments

BIM Schedule Rule Checker Building Energy Regression Construction Grid Route Planner Robot Telemetry Rule Monitor Visual Provider Contract Deterministic Research Workflow Local Model Serving and Monitoring Local Text Classification Lab

These systems are runnable and tested, but remain experiments rather than first-screen portfolio claims.

Project Scope

What this portfolio does and does not show

Implementation: runnable code, tests, fitted classical models, persistence, metrics, and simulation.

Data: project pages identify synthetic conversations, sites, rates, tenders, public subsets, and locally downloaded sources.

Mock providers: deterministic LLM and VLM substitutes test interface contracts, not model quality.

Simulation: robotics results do not imply hardware, perception, or physical safety validation.

Professional work: no customer adoption, deployed reliability, compliance outcome, QS sign-off, tender advice, or approvable design is claimed.

Contact + Evidence

Continue into code, evaluations, and project context

Repository Project Guide Skills Matrix Portfolio Brief LinkedIn Email