Verifiable career facts and engineering judgment
Xinwei Xiong
AI Product Systems Engineer
Agent Engineering / Evaluation / Native iOS
I build agent products around user scenarios, with durable context, failure recovery and tool policy across web and native iOS.
Help people and agents recover context and continue work across tools and sessions.
Engineering Signals
Selected Systems
Problem, ownership, decisions, validation, artifacts and boundaries
Ownership / Solo Product Engineer
DayPage
Native iOS capture and AI compilation with a Markdown/YAML vault, draft recovery and concurrent-write protection.
Problem / Constraint
Personal notes were fragmented across capture surfaces, making context hard to continue, inspect or reuse with agents.
Key Decisions
- Chose a Markdown and YAML file-system vault instead of an opaque database.
- Used BGTaskScheduler for nightly AI compilation with backfill and local failure notifications.
- Made the compilation endpoint configurable through an OpenAI-compatible base URL and model so storage and workflows are not bound to one provider.
Validation / Outcomes
- Public automation delivered TestFlight v0.4.74, with the matching prerelease providing a signed distribution IPA under a fixed SHA-256; nine scheduled regressions succeeded from 22 to 30 Jul 2026, with the latest completing 187 XCTest, 277 Swift Testing tests with 4 known issues recorded by the test framework, and 4/4 Maestro UI flows.
- Implemented up to seven-day backfill; Local-to-iCloud vault migration checks file counts and sampled SHA-256 hashes before switching, retains the local backup, and is covered by scheduled CI.
- The owner's fact card records a local vault spanning about 250 days of Markdown; this is context-corpus scale, not consecutive usage or external-user count.
- A pinned September commit adds per-record serialized writes, durable draft recovery and cleanup after persistence confirmation, preserving local conflicts before remote replacement.
Boundary / Not Yet Proven
TestFlight v0.4.74 and the scheduled-regression figures are July 2026 verification snapshots; September changes are pinned source-code evidence. The full Skills → Agent Execution → UI Artifacts flow remains a proposal; external adoption, real-device background reliability and the current distributed build were not verified in this review.
Ownership / Product design and AI-assisted engineering
StripSearch
Person research with traceable evidence, persistent progress and JSON / Markdown / HTML / PDF exports.
Problem / Constraint
Person research spans multiple sources; interrupted runs and withdrawn sources require preserved progress and revalidated conclusions.
Key Decisions
- Persist checkpoints, action reservations and receipts in SQLite; reuse successful actions and stop uncertain paid calls while retaining collected material.
- Drive four export formats from one structured report; invalidate dependent conclusions after source withdrawal and recheck the report around export.
Validation / Outcomes
- Implemented research-controller recovery, cancellation and action records, and delivered an authenticated, account-isolated workspace with JSON, Markdown, HTML and PDF exports.
- Public CI succeeded for the pinned 8 Oct 2026 commit; its PR records 787 passing web tests, one skip and 40 frozen offline replays. These checks cover execution contracts rather than person-research quality.
Boundary / Not Yet Proven
The 8 Oct 2026 deployment has no provider credentials, so live research returns provider_unavailable; the new Fetch pipeline still uses synthetic adapters. Workspace and export acceptance used synthetic reports and does not validate current live research, cross-platform collection, research quality or external adoption.
Ownership / Runtime Design and Implementation
Agentry
A resumable, auditable agent runtime that unifies loops, tools, sessions, policy and multiple transports.
Problem / Constraint
Agent applications often couple model calls, tool permissions, session recovery and protocol access, making them hard to test and audit.
Key Decisions
- Defined Provider, Tool, Transport and Session as four stable seams.
- Modeled sessions as logically appended event streams, persisted as JSONL snapshots and replayed to rebuild context.
- Implemented allow, deny and ask policy primitives plus SSRF screening that rechecks DNS results, redirects and private address ranges.
Validation / Outcomes
- Implemented a bounded agent loop, replayable JSONL sessions, cancellation-aware retries, HTTP and stdio runtime transports, stdio MCP client/server and agent-as-tool primitives.
- The pinned public commit contains 191 top-level tests across 17 packages; build, vet, test and the race detector were reproduced at that commit with 63.2% statement coverage.
Boundary / Not Yet Proven
This remains a prototype inside the DayPage monorepo; repository CI does not run Agentry tests, the CDP browser package is untested, ask approval and sub-agent profiles are not wired into the CLI, and there is no independent release or external adoption evidence.
Ownership / Workflow design and AI-assisted implementation
agent-kit
A shared skill tree for six coding agents, with dry-run installation, idempotent symlinks and conflict preflight, plus third-party source/license validation and public-content scanning.
Boundary / Not Yet Proven
Installation and validation workflows are project work; third-party skills retain their authorship. Public CI verifies tool behavior without establishing external adoption, cost savings or model-quality gains. Rule-based scanning cannot guarantee removal of every secret.
Ownership / Solo Product Engineer
Solo Compass
An AI-curated living map for solo travelers, validating prompts and schema with a bot before expanding to web and native iOS.
Validation / Outcomes
- Public automation completed signing, Fastlane and TestFlight upload, version tagging, GitHub prerelease and IPA archival through v0.1.289.
- Public CI validates TypeScript types, formatting, schema dry-runs, TypeScript-to-Swift data parity and gitleaks secret scanning.
Boundary / Not Yet Proven
The TestFlight upload is verifiable, but the separate iOS CI currently masks an xcodebuild destination error and cannot prove build or test success; the README also warns against production use, and external adoption remains unverified.
Ownership / Design and Implementation
CCT
A public working prototype that captures sessions through Claude Code Hooks and provides local search, browsing and statistics over SQLite and JSONL.
Ownership / Solo Product Engineer
Telepace Next
A voice-native, agent-first user-research prototype exposed through MCP, web and REST/webhooks.
Boundary / Not Yet Proven
Public evidence currently shows repository activity in July 2026, not adoption or payment outcomes.
Ownership / Open-source Intern / Community Contributor
OpenIM
Owned deployment, CI and contributor governance in a large Go IM project; reached #1 on GitHub's contributor graph.
Problem / Constraint
As the project grew, deployment, testing, protocol explanations and community knowledge needed standardization for distributed collaboration.
Key Decisions
- Treated deployment guides, protocol documentation and contribution workflows as engineering systems.
- Coordinated across roles through issues, pull requests, biweekly meetings and asynchronous documentation.
Validation / Outcomes
- Ranked first on the OpenIM Server default-branch contributor graph with 358 contributions; a public retrospective documents 100+ PRs and 100+ issues.
- Representative merged work includes Helm Charts and Helmfile orchestration for 11 services (#1291), API and E2E testing in GitHub Actions (#1359), and the contributor baseline for RFCs, DCO, PR review and releases (#366).
- Published eight long-form OpenIM videos from Oct 2023 to Jan 2024; a public-platform snapshot on 20 Jul 2026 records the top three at 4,736, 3,308 and 959 views.
Boundary / Not Yet Proven
The contribution count does not imply every commit was core IM server code; the resume separates engineering, documentation and community scope.
Ownership / Founder / Product Engineer
Telepace AI
Led an AI feedback analysis SaaS spanning asynchronous data pipelines, real-time web interaction and Kubernetes deployment.
Problem / Constraint
Customer feedback arrived through fragmented forms and APIs and needed asynchronous aggregation, analysis and live presentation.
Key Decisions
- Used a Django and Celery pipeline with a Next.js frontend and WebSocket delivery.
- Built dual-region Kubernetes and multi-registry delivery too early, later identifying this as a sequencing error.
Validation / Outcomes
- Led the core backend, system integration and deployment while collaborating on the web product; without reaching a first paying user, the initial venture stopped after 28 Oct 2025.
- Turned the failure into explicit validation-first sequencing: a later public strategy sets gates of at least 100 signups or 20 paid preorders per concept within three weeks, and a pivot if a direction misses $2K MRR by month four. These are forward-looking kill criteria, not achieved results.
Boundary / Not Yet Proven
Paused refers only to the 2024 to 2025 initial venture. The 2026 prototype is listed separately; interview and trial counts still require owner confirmation.
Career timeline
From deployment and open-source collaboration at OpenIM to an AI SaaS venture and independent products; recent work connects user stories to evaluation cases and focuses on task recovery and result verification.
Telepace AI
Founder / Product Engineer
Founded an AI feedback analysis SaaS, owning the core backend, asynchronous data pipeline, system integration and Kubernetes delivery while collaborating on the web product. The stack worked end to end, but no first paying user emerged, so the initial venture stopped in 2025 and the lesson became explicit validation gates and kill criteria.
The Plant
Engineering Team Member
Briefly joined the international team of a Tokyo-based commerce software company during the final university year, then moved into independent entrepreneurship. Exact dates, formal title and a representative outcome remain pending.
OpenIM
Open-source Intern, then Community Contributor
The owner confirms sustained OpenIM community work from junior year through the final spring; a public retrospective records a six-month internship within that period, while exact employment dates remain pending. Currently ranked first on GitHub's contributor graph, with work across deployment, test automation, protocol and documentation clarification, and the Chinese community.
Work in Progress
agent-diff-guard
A local agent-governance tool that gates high-risk drift before push, deduplicates token usage by message ID, and ranks projects and sessions by estimated cost to expose runaway work.
voiceflow
Designed provider interfaces for a voice AI pipeline so STT, TTS, LLM and storage implementations could be swapped through shared abstractions and configuration.
Writing & Open Source
I publish agent source-code research, product practice and monthly reflections. Recent writing connects user stories to success criteria, failure analysis and regression cases, and studies task and authorization boundaries in personal assistants.
Core Stack
- Agent Systems
- Agent loop / MCP / Tool policy / Session lifecycle / Computer Use / Provider abstraction
- Product Systems
- Native iOS / Web / Backend / TestFlight / Local-first
- Validation
- Task-specific eval design / Failure analysis / Regression testing / Artifact verification / CI/CD / XCTest / Swift Testing / Maestro / Go race detector
- Languages
- Swift / TypeScript / Go / Python
Claim Ledger
Public evidence and owner attestations are explicitly separated · Expand ↓Collapse ↑
claim.daypage.testflight
A public pipeline uploaded DayPage TestFlight v0.4.74 on 22 Jul 2026; the matching prerelease contains an 18,339,826-byte DayPage.ipa under a fixed SHA-256. A reproducible audit of the public IPA reads com.daypage.app, 0.4.74 (1879), minimum iOS 16.1, an App Store distribution profile and get-task-allow=false. This proves a distributable artifact, not external adoption or real-device reliability.
claim.daypage.ci
Nine consecutive public scheduled regressions succeeded from 22 to 30 Jul 2026; the latest completed 187 XCTest, 277 Swift Testing tests with 4 known issues recorded by the test framework, and 4/4 Maestro UI flows.
claim.daypage.recovery
DayPage implements up to seven-day background-compilation backfill; Local-to-iCloud vault migration checks file counts and sampled SHA-256 hashes before switching, retains the local backup, and is tested in public scheduled CI.
claim.daypage.vault
DayPage stores raw entries, daily pages and short-term memory in a Markdown and YAML file-system vault.
claim.daypage.dogfood
DayPage is used for daily dogfooding; duration and external tester count remain pending.
claim.daypage.vault_scale
The owner's fact card records a DayPage local vault spanning about 250 days of Markdown; this measures context-corpus span, not consecutive usage or external-user count.
claim.agentry.runtime
Agentry implements a bounded agent loop, replayable JSONL sessions, cancellation-aware HTTP retries, HTTP and stdio runtime transports, stdio MCP client/server, HTTP-tool SSRF defenses and agent-as-tool primitives.
claim.agentry.tests
The pinned public commit contains 191 top-level tests across 17 packages; go build, vet, test and the race detector were reproduced at that commit with 63.2% statement coverage, but repository CI does not yet run this suite.
claim.agent_diff_guard.usage_ranking
agent-diff-guard parses local Claude Code session JSONL, deduplicates token usage by message ID, aggregates input, output, cache tokens and estimated cost by project and session, and sorts results by cost and output tokens.
claim.openim.rank
cubxxw currently ranks first on the OpenIM Server GitHub contributor graph with 358 contributions.
claim.openim.internship
A public retrospective records a six-month internship within the OpenIM experience; exact employment dates remain pending.
claim.openim.academic_stage
The owner confirms sustained OpenIM community participation from junior year through the final spring; the verifiable default-branch contribution window is Apr 2023 to Apr 2024.
claim.openim.scope
OpenIM contributions covered engineering practices, deployment, CI and testing, contributor workflows, protocol and documentation work, and the Chinese community.
claim.openim.representative_prs
Representative merged work includes Helm Charts and Helmfile orchestration for 11 services (#1291), API and E2E GitHub Actions (#1359), and the contributor baseline for RFCs, DCO, PR review and releases (#366).
claim.openim.community_reach
Eight long-form OpenIM videos were published from Oct 2023 to Jan 2024; a public-platform snapshot on 20 Jul 2026 records the top three at 4,736, 3,308 and 959 views.
claim.telepace.initial
The initial Telepace venture was built in 2024 to 2025, did not reach its first paying user, and stopped activity after 28 Oct 2025.
claim.telepace.validation_gates
In a later public product strategy, the owner sets validation gates of at least 100 signups or 20 paid preorders per concept within three weeks, with a pivot if a direction misses $2K MRR by month four; these are future gates, not achieved results.
claim.telepace.responsibility
An authenticated owner audit of the now-private Telepace repositories ranks the owner first in feedback-collection-api and server (577 / 279 contributions) and second in the multi-contributor web repository (162). This supports the responsibility boundary of leading the core backend and system integration while collaborating on the web product; because the repositories are not public, the figures remain owner-attested for public readers and do not demonstrate commercial adoption.
claim.telepace.nexus_alpha
The owner's fact card records real non-paying use of the Nexus alpha within the Telepace product line. Because no verifiable user count, activation cohort or retention cohort was retained, this resume treats it only as a qualitative contact signal rather than an adoption outcome.
claim.telepace.next
Telepace Next is a publicly active July 2026 voice-native, agent-first user-research prototype.
claim.solo.testflight
Solo Compass has a public TestFlight v0.1.289 upload plus passing TypeScript, schema-parity and secret-scan checks; its README still labels the product Pre-alpha, and a separate iOS workflow has a false positive that masks xcodebuild failure.
claim.voiceflow.lifecycle
The Git history currently visible in the public repository spans May 2023 to Apr 2025; GitHub does not mark the repository as archived.
claim.cct.prototype
CCT is a public working prototype using Claude Code Hooks, FastAPI, SQLite/JSONL and React.
claim.theplant.employment
The owner reports working with The Plant K.K.'s engineering team for about two months in 2024; exact dates, formal title and outcomes remain pending.
claim.theplant.company_context
The Plant K.K. is a Tokyo-based software company building commerce systems for the Japanese market.
claim.education.graduation
The owner reports completing undergraduate study at Wuhan College of Arts & Sciences in 2024 and states in a public annual review that Computer Science was the chosen major; the exact degree title remains pending.
claim.theplant.transition
The owner's public annual review describes leaving a foreign-company role around graduation to pursue independent work and entrepreneurship; it does not name the employer, so it supports the career transition rather than formal The Plant employment.
claim.basics.public_profile
On 31 Jul 2026, the owner confirmed Shenzhen as the current location; the public GitHub profile still lists @Singapore with hireable=true, so that public location remains to be updated.
claim.daypage.draft_recovery
The pinned 20 Sep 2026 commit implements durable draft recovery, cleanup after persistence confirmation, per-record serialized writes and local conflict preservation before remote replacement. This review checks code rather than certifying real-account sync or device reliability.
claim.stripsearch.recovery
Persist checkpoints, action reservations and receipts in SQLite; reuse successful actions and stop uncertain paid calls while retaining collected material.
claim.stripsearch.exports
Drive four export formats from one structured report; invalidate dependent conclusions after source withdrawal and recheck the report around export.
claim.stripsearch.validation
Public CI succeeded for the pinned 8 Oct 2026 commit; its PR records 787 passing web tests, one skip and 40 frozen offline replays. These checks cover execution contracts rather than person-research quality.
claim.stripsearch.live_boundary
The 8 Oct 2026 deployment has no provider credentials, so live research returns provider_unavailable; the new Fetch pipeline still uses synthetic adapters. Workspace and export acceptance used synthetic reports and does not validate current live research, cross-platform collection, research quality or external adoption.
claim.agent_kit.installation
agent-kit provides dry-run and idempotent installation from a shared skill source for six coding agents. All-target preflight blocks writes on conflicts, resolved paths are rechecked before writes, and source commits, tree hashes and licenses are validated. Public CI checks tool behavior; third-party skills retain their authorship.
claim.writing.eval_method
September public notes discuss code assertions, human calibration and failure regression; October notes connect user stories to evaluation cases and commits. These sources support evaluation-method research, not a complete production evaluation platform or improved live quality.
claim.writing.personal_agent_design
Public research uses pinned OpenClaw source to examine run recovery and tool policy, and proposes a reference model for mandates, tasks, runs and result verification. The article explicitly labels this a design example rather than an author-deployed, tested system.
claim.writing.figwright
The public article records use of free Figwright; the personal design-system workflow and DayPage card exercise remain proposals rather than a delivered cross-platform design system.