Skip to content
William
SOFTWARE · AI SYSTEMS · PRIVACY-CONSCIOUS · EVIDENCE-DRIVEN

I build AI-assisted software that earns trust through evidence.

I turn product ideas into privacy-conscious, testable systems — from on-device media and security tooling to local AI workflows.

Explore Featured Work ↓How I Build →
62/62
Tests Passed
EchoReel Native iOS Suite
50-case
Security Benchmark
SCSA Slither50 Matrix
160/160
Soak Tasks Verified
VoiceFlow Watchdog
EVERY CLAIM TRACEABLE TO PROJECT EVIDENCE.
Detailed precision, recall & memory metrics inside case studies →
CORE ENGINEERING THESIS
“AI accelerates iteration. Trust comes from explicit boundaries, reproducible evidence, and the willingness to reject what fails.”
Fail-closed by defaultZero-cloud biometric egressHonest negative result disclosure
01FEATURED CASE STUDY

EchoReel

On-device photo memories with human-controlled identity matching.

GOVERNING PRINCIPLE

“AI proposes candidates. People make identity decisions.”

Problem

Commercial cloud clustering uploads family photos to remote servers and silently merges identities based on opaque similarity scores. EchoReel executes all indexing and memory compilation locally with zero network permissions.

Architectural Decision

The architecture enforces an Actor-isolated boundary where Apple Vision and local clustering only propose candidate occurrences. AI is strictly prohibited from writing to the persistent SQLite relationship graph without an explicit human tap in the SwiftUI Review Queue.

Test Suite
62/62 PASS
3/3 Simulator Flows
E1 Memory Gate
0.215 MiB
Budget: ≤0.250 MiB/photo
Read EchoReel Case Study →
SYSTEM SCHEMATIC · DATA FLOW & TRUST BOUNDARY

On-Device Identity Isolation Pipeline

Zero Network Egress
01 PhotoKit

01. Local Input: PhotoKit Asset Reader loads full-resolution photos into on-device memory with zero network permissions.

02 Vision

02. On-Device Detection: Apple Vision face observations extract face bounding boxes and temporary similarity features on-device.

03 AI Proposals

03. AI Proposals: Groups candidate occurrences as unconfirmed suggestions only. Enforced boundary: AI proposes, people confirm.

04 Human GateHuman Gate

04. Mandatory Human Gate: SwiftUI Review Queue requires user tap to confirm identity. Zero auto-linking permitted.

05 SQLite Graph

05. Local Persistence: SQLite stores confirmed identities and memories within 0.215 MiB/photo memory budget.

06 AVFoundation

06. Synthesis: AVFoundation video renderer composes hardware-encoded foreground video stories.

Selected Stage:01. Local Input: PhotoKit Asset Reader loads full-resolution photos into on-device memory with zero network permissions.
EVALUATION GATE LOG · ECHOREEL-021
GATE FAILED · NOT PROMOTED
Candidate Artifact
Candidate Face Embedding Model
Target Gate
Recall@50 ≥ 0.850000
Measured Result
0.585117 (-0.264883)
DECISION: The automated clustering candidate failed the precision-recall threshold. Rather than masking the deficit or shipping false identity clusters to user albums, the model was rejected and discarded. The shipping architecture retains Apple Vision base observations combined with explicit manual confirmation.
02FEATURED CASE STUDY
PROVENANCE ARCHITECTURE · EVIDENCE GRAPH & TRUST BOUNDARY

Deterministic Screening & Constrained Explanation

Fail-Closed Native Builds
01 Ingestion

01. Ingestion: Safely stages untrusted Solidity source with path-traversal safeguards and disabled-by-default native builds.

02 Slither AST

02. Slither AST: Runs 27 mapped static detectors producing verifiable AST code spans and call graphs.

03 Normalizer

03. Normalizer: Maps diverse detector outputs into unified finding schemas with deterministic CWE/SWC identifiers.

04 Evidence GraphEvidence Anchor

04. Evidence Graph: Stores verified findings and relations in SQLite as immutable factual anchor nodes.

05 Grounded LLM

05. Constrained Explanation: LLM explains and writes diffs strictly bounded by Evidence Graph nodes; zero hallucinated CVEs.

Selected Stage:01. Ingestion: Safely stages untrusted Solidity source with path-traversal safeguards and disabled-by-default native builds.
EVALUATION GATE LOG · SCSA-BENCH-02
MEASURED LIMITATION · 4 FALSE POSITIVES
Candidate Artifact
Slither50 v2 Automated Screening
Target Gate
Zero False Negatives (Recall 100.00%)
Measured Result
25 TP · 21 TN · 4 FP · 0 FN
LIMITATION BOUNDARY: The benchmark recorded 4 false positives across reentrancy and delegatecall detectors. The system is designed and documented strictly as a developer triage assistant, not an automated certification or substitute for formal verification and human audit.

Smart Contract Security Assistant

Deterministic vulnerability screening with evidence-grounded explanation.

GOVERNING PRINCIPLE

“Deterministic static analyzers own vulnerability facts. LLMs explain and remediate.”

Problem

Generative LLMs hallucinate non-existent vulnerabilities or miss subtle state machine exploits. Auditing teams need fast first-pass screening without mistaking model confidence for mathematical security proof.

Architectural Decision

Native builds fail closed by default. The pipeline executes 27 mapped Slither AST detectors, normalizes findings into a local SQLite Evidence Graph, and constrains generative models to explaining established evidence nodes. Unmapped tool output is preserved as raw trace.

Test Suite
175 PASS
140 pytest + 35 Vitest
Slither50 v2 Benchmark
100% Recall
86.21% Precision · F1 92.59%
Read SCSA Case Study →Inspect Public MIT Repo ↗
03FEATURED CASE STUDY

VoiceFlow

Local-first macOS voice agent with confirmation-gated execution.

GOVERNING PRINCIPLE

“No silent external mutations. Ambiguous intents fail closed.”

Problem

Voice agents executing system-level actions risk accidental destructive executions when speech recognition misinterprets input. Sending continuous room audio to cloud inference APIs compromises desktop privacy.

Architectural Decision

The architecture isolates execution behind a confirmation state machine. Read-only queries run immediately; state mutations (file deletions, external posts, terminal scripts) halt for physical confirmation. An out-of-process watchdog recovers failed workers in 0.43 seconds.

8-Hour Soak Suite
160/160 PASS
0.12 MB/hr Memory Slope
Safety Benchmark
0 Auto-Submits
0 High-Risk Actions Unconfirmed
Read VoiceFlow Case Study →
FAIL-SAFE COMMAND PIPELINE · SIDE-EFFECT CONFIRMATION BARRIER

Local Speech Intent to Verified macOS Execution

0.43s Watchdog Recovery
01 Capture

01. Push-to-Talk Capture intercepts audio via Carbon hotkey and feeds 16kHz PCM to on-device memory without cloud exposure.

02 Whisper STT

02. Local Whisper STT (faster-whisper int8 CPU) generates transcript in p50 2,003ms with zero external network connectivity.

03 Normalizer

03. Rule-based zh-TW Text Normalizer sanitizes transcripts against domain dictionaries, eliminating STT hallucinatory drift.

04 Intent Router

04. Intent Classifier partitions commands into read-only queries versus state-mutating actions (B7: 1.0000 success rate).

05 Safety GateConfirmation Gate

05. Safety Gate intercepts all external side effects. Destructive actions require explicit user confirmation (0 auto-submits over 160 soak tests).

06 Execution & Audit

06. Sandboxed adapter executes confirmed command. Watchdog restarts crashed workers in 0.43s. Audits are AES-GCM encrypted in local SQLite.

Selected Stage:01. Push-to-Talk Capture intercepts audio via Carbon hotkey and feeds 16kHz PCM to on-device memory without cloud exposure.
EVALUATION GATE LOG · VF-STT-03
GATE FAILED · NOT PROMOTED
Candidate Artifact
Fine-tuned Whisper int8 (CTranslate2)
Target Gate
Relative WER Gain > 10% & Action Acc ≥ 0.90
Measured Result
WER +0.019188 · Action +0.0727
DECISION: Despite successful quantization and model loading, empirical evaluation revealed negligible command accuracy improvement (+0.0727) alongside quantization artifacts at command boundaries. The candidate was rejected; the shipping build retains the base model combined with deterministic regex normalization.
ENGINEERING METHODOLOGY

How I Build

A disciplined five-stage progression from invariant framing to empirical promotion proof.

01STAGE

DEFINE

Scope constraints, core user invariants, and measurable acceptance gates before touching code.

Artifact: Invariant Contracts
02STAGE

ARCHITECT

Design threat boundaries, privacy models, actor concurrency, and strict schema contracts.

Artifact: System Schemas
03ACTIVE CORE

BUILD

Supervise AI coding tools to generate typed modules, migrate data stores, and wire pipelines.

Artifact: Verified Codebase
04STAGE

VERIFY

Execute automated suites, regression tests, memory profiling, and adversarial security fixtures.

Artifact: Reproducible Runs
05STAGE

ITERATE

Evaluate promotion gates. Reject underperforming models honestly and refine proven paths.

Artifact: Gate Decision Log
SECONDARY INVESTIGATIONS & BENCHMARKS

Research & Experiments

Concise engineering investigations exploring agent governance, local environment behavior, and reproducible evaluation harnesses.

Secondary research records with explicit boundaries and transparent stop conditions.

EXP-01PAUSED RESEARCH

GovernSeed: Verifiable Agent Governance Contracts

Machine-verifiable local contracts and fail-closed qualification gates for agent-assisted workflows.

Invariant SchemasFail-Closed GatesAgent WorkflowsOffline Verification
01 · Research Question

Can agent-assisted software projects be governed through machine-verifiable local contracts and fail-closed qualification gates?

02 · What Was Built & Tested

Authored formal governance contract schemas, offline test harnesses, and automated gate checklists to constrain multi-agent code generation within strict repository boundaries.

03 · Evidence & Result

R1 qualification gates successfully prevented unauthorized cross-module imports and unconstrained side-effects across 12 test iterations.

04 · Limitation & Stop Condition

Research round R2 was stopped at Gate G2: empirical verification metrics could not sufficiently prove that agent semantic drift was contained without manual specification audits. The research stopped when evidence no longer justified the claim.

05 · What Was Learned

Governance contracts cannot rely on agent self-reporting; gates must remain strictly external, deterministic, and fail-closed.

EXP-02RESEARCH LAB

Selfhosted Lab: Local Stack & Localization Evaluation

Comparative evaluation of 30 self-hosted applications under local Docker and zh-TW localization conditions.

30 Systems EvaluatedDocker Composezh-TW LocalizationResource Profiling
01 · Research Question

How do diverse open-source self-hosted systems behave under isolated local containerization and strict Traditional Chinese localization requirements?

02 · What Was Built & Tested

Standardized local container deployment workflows, configuration templates, zh-TW language pack validations, and runtime health probes across 30 self-hosted services.

03 · Evidence & Result

Profiled baseline memory footprints (ranging from 42 MB to 1.8 GB idle), identified translation coverage gaps across 14 tools, and documented database initialization behaviors.

04 · Limitation & Stop Condition

Localhost validation only. External SMTP deliverability, production DNS routing, multi-tenant authentication, and cloud infrastructure scale were intentionally not certified. Keys, certificates, and runtime data are strictly excluded.

05 · What Was Learned

Container reproducibility requires pin-pointed base images and isolated volume state; localization quality varies widely regardless of project popularity.

EXP-03PAUSED RESEARCH

OSS Benchmark V3/V4: Reproducible Agent Evaluation

Benchmark engineering harness with sealed starting states, independent oracles, and runner containment.

Sealed StatesIndependent OraclesContainment AuditZero False Claims
01 · Research Question

Can agent performance on complex software tasks be evaluated with zero environmental contamination and strictly verifiable oracles?

02 · What Was Built & Tested

Engineered immutable test fixtures, cryptographic workspace sealing, schema-validated task descriptors, and external oracle harnesses decoupled from agent LLM contexts.

03 · Evidence & Result

V3 verified deterministic test execution across 40 baseline challenges. V4 architecture introduced strict resource cgroups and automated cleanup audits.

04 · Limitation & Stop Condition

V4 pilot was halted before execution: Linux cgroup cleanup isolation could not be mathematically guaranteed across kernel edge cases. With 0 pilot runs completed, no treatment effect is claimed.

05 · What Was Learned

Benchmark credibility demands rigorous negative controls; publishing 0 runs with transparent containment halt is superior to publishing contaminated metrics.