An endpoint that returned 882 bytes took as long as one that returned 52 KB. That was the clue. The time was not in the response. It was in a cache row that every request read a…
CQRS said the worker owns decisions and FastAPI only reads cache. Users still waited seconds on /dashboard. The compute boundary was fine. The read path was not.
Prep for an Invest walkthrough wanted Map → Micron → 13F holdings. The Micron snapshot had no verifiable 13F citation. The run disclosed the gap and used NVIDIA for holdings ins…
ADR-008 gave the gateway an HTTP control plane. signalwatcher still drafted from Invest AI Top-1 alone while InvestSim evidence stayed "documented for later."
A draft "Daily Stock Market Intelligence" stack looked like a new FastAPI + Alembic + Next app. Invest AI already had Resend digests, report triggers, macro radar, and a reports…
Phase 1 shipped the architecture package: system design, ADRs 0001–0018, C4, contracts. The next temptation is SMS + templates + campaigns before the host can say no to a bad JWT.
A dense Chinese A4 PDF framed around a $200K FI portfolio looks like a new product. New domain. New worker. New Resend project. Maybe a cute investment-report.xingai.app.
You ship a verifier. Dashboard gets Accept / Reject. Engineer patches the verify JSON with review=accepted. Next CLI run overwrites the cache. Human work vanishes. Trust evapora…
On July 11, the Robinhood MCP gateway story was simple: the gateway blocked every write order on purpose. G1, G6, and G7 were real code. G2 through G5 were named, explicit fail-…
Public posters love a clean ladder: Plan → Build → Validate → Operate, twelve boxes, each with a Tools row. LangChain here, LangSmith there, Docker at the end. Fine as a checkli…
XingAI Engineering Communication Coach is not a Grammarly wrapper. The product question is: what should a non-native engineer practice today so tomorrow’s design review, inciden…
We finally wired Resend secrets into the Opportunity Radar GitHub Actions job and forced a send. The mail arrived. It also looked like a shell: date line, empty tables, a canned…
A question that comes up every time someone builds their first MCP server: it's just an HTTP endpoint calling a few functions — why not hand it an API key like any other interna…
[Part 1](2026-07-13-mcp-oauth-vs-api-key.md) and [Part 2](2026-07-13-mcp-auth-api-key-vs-oauth-pkce.md) of this series covered why claims-mcp-oauth-poc uses OAuth and what the p…
[Last post](2026-07-13-mcp-oauth-vs-api-key.md) covered why claims-mcp-oauth-poc uses OAuth 2.1 + PKCE instead of a plain API key — the short version being that MCP's real shape…
Yesterday's post covered [claims-mcp-oauth-poc](2026-07-12-claims-mcp-oauth-poc-same-auth-different-industry.md): real OAuth 2.1 + PKCE in front of four deliberately narrow tool…
Two [education guides](https://github.com/xingaiapp/xingai-enterprise-ai-design/blob/main/guides/2026-07-12-mcp-oauth-pkce-lab.md) shipped yesterday, walking through a full OAut…
[Invest AI ADR-028](https://github.com/xingaiapp/xingai-invest-ai/blob/main/docs/adr/028-robinhood-mcp-execution-gates.md) defined seven gates (G1–G7) a trade order must clear b…
The July 5 [Agent Firewall post](2026-07-05-agent-firewall-helpfulness-attack-surface.md) described a 0din-class block with risk 60. That number was honest — and incomplete.
ADR-003 promised four approval buttons: Approve once / Approve for session / Deny / Deny + add rule. The first three shipped. The fourth sat in the UI copy with no backend.
[ADR-003](https://github.com/xingaiapp/xingai-agent-firewall/blob/main/docs/adr/003-approval-workflow-ledger.md) specified four one-click approval actions back in the PRD. [ADR-…
Mozilla's 0din team [showed](https://www.tomshardware.com/tech-industry/cyber-security/ai-coding-agents-can-be-tricked-into-installing-malware-via-clean-github-repositories-mozi…
Every team argues about shipping fast vs. shipping well, as if the tradeoff were a law of nature. AI-native engineering teams are proving it isn't. The mechanism is first-pass y…
[Engineering System ADR-001](https://github.com/xingaiapp/xingai-engineering-system/blob/main/docs/adr/001-personal-memory-engine.md) defined the architecture. [Invest AI ADR-03…
The Opportunity Radar generates a detailed markdown report every week. Until now, it lived in a private repo. This week we added a newsletter worker that reads that report and s…
Meal AI v4 is a pure Next.js app — no Python backend, no database, all decision logic in TypeScript on the client. The [Decision Ledger schema](https://github.com/xingaiapp/xing…
The AI engineering discipline has moved fast. We went from prompt engineering to context engineering, then to harness engineering, and now to loop engineering — where systems pu…
Meal AI is the health domain of XingAI — nutrition planning, meal recommendations, and caregiver support for family members with specific health conditions.
Polymarket AI scans prediction markets, estimates edge, and sizes positions with fractional Kelly. The tempting failure mode: wire a cron job to the CLOB and let edge > 5% auto-…
After [ADR-026](https://github.com/xingaiapp/xingai-invest-ai/blob/main/docs/adr/026-decision-engine-integration.md), Invest AI can show two different regime signals on the same…
Robinhood [Agentic Trading](https://robinhood.com/us/en/support/articles/agentic-trading-overview/) lets connected agents read portfolios and place trades — including auto-execu…
Invest AI ships a production dashboard with macro consensus. Decision Engine owns multi-timeframe technical scoring and backtests. Two repos on purpose — [ADR-002 on the Decisio…
Robinhood's [Agentic Trading MCP](https://robinhood.com/us/en/support/articles/agentic-trading-overview/) lets third-party agents read portfolio data and place trades in a dedic…
Insurance adjusters don't guess. They read the claim, pull the policy, check history, apply rules, and decide — with paperwork that shows why. Our new POC automates that workflo…
We now maintain 10 Cursor Skills in [xingai-engineering-system](https://github.com/xingaiapp/xingai-engineering-system) and plan MCP servers for Invest AI data and (eventually) …
The marketing site puts locale in the path (/zh/…). Product apps like Research AI switch language in the header and persist to localStorage. SEO still has to ship.
After you type a topic, Research AI returns a verdict and a 0–100 Learning ROI score. Both come from the same worker run — not from the browser guessing.
The first version of our daily-picks email worked technically. It read the worker cache, selected top signals, and sent an email. But the product shape was wrong: it looked like…
Model Context Protocol (MCP) gives agents a standard way to discover and call tools. That part is easy to demo. Production is harder: scoped permissions, chained servers, multi-…
[T Today](https://t.xingai.app) reads brokerage screenshots and returns a 做T plan — buy zones, sell targets, cash notes. A confident wrong ticker is worse than “I can’t read thi…
[T Today](https://t.xingai.app) ships rule-engine logic, Prisma schema, and OpenAI vision routes in one Next.js repo. TypeScript can look fine in the editor while tsc fails. A r…
We ship more than one app. Meal Coach, Cook AI, Travel AI, Invest AI, the marketing site at xingai.app — each repo had its own habits for navigation, languages, themes, and lega…
[T Today](https://t.xingai.app) is the screenshot-first coach: upload brokerage holdings, get a 做T plan with buy zones and rule checks. It’s paper-only — no orders.
T Today’s holdings coach uses OpenAI with responseformat: { type: "jsonobject" }. The model returns a big JSON blob with zh and en blocks. Users switched language — still empty …
In 2023, a lot of AI engineering looked like prompt writing. You opened a chat box, tuned a system prompt, added examples, and tried to make the model answer better.
It is easy to assume that “Connect Git Repository” on Vercel only works for public GitHub repos — especially when your dashboard still shows Connect Git because you have been sh…
[Invest AI](https://xingai.app/apps/invest-ai) answers “what does today look like?” — signals, dashboard context, and (eventually) richer decision surfaces. A different question…
[ADR-004](2026-05-12-v1-hosting-fly-io.md) explains why we picked Fly.io for the Python side and Vercel for the Next.js side. It does not replace copy-pasteable commands, projec…
This post documents a decision we accepted, not something already running in production. It belongs in the same story as our CQRS market cache: derived data should have one writer.
Invest AI surfaces BUY / HOLD / SELL style signals and confidence scores. In many jurisdictions, anything that looks like personalized investment advice can drag you into regula…
Our V1 system uses a fallback chain: try Ollama (local, free), then Gemini, then OpenAI. Whichever model responds first wins. It's resilient, but every model does the exact same…
An endpoint that returned 882 bytes took as long as one that returned 52 KB. That was the clue. The time was not in the response. It was in a cache row that every request read a…
CQRS said the worker owns decisions and FastAPI only reads cache. Users still waited seconds on /dashboard. The compute boundary was fine. The read path was not.
Prep for an Invest walkthrough wanted Map → Micron → 13F holdings. The Micron snapshot had no verifiable 13F citation. The run disclosed the gap and used NVIDIA for holdings ins…
ADR-008 gave the gateway an HTTP control plane. signalwatcher still drafted from Invest AI Top-1 alone while InvestSim evidence stayed "documented for later."
A draft "Daily Stock Market Intelligence" stack looked like a new FastAPI + Alembic + Next app. Invest AI already had Resend digests, report triggers, macro radar, and a reports…
Phase 1 shipped the architecture package: system design, ADRs 0001–0018, C4, contracts. The next temptation is SMS + templates + campaigns before the host can say no to a bad JWT.
A dense Chinese A4 PDF framed around a $200K FI portfolio looks like a new product. New domain. New worker. New Resend project. Maybe a cute investment-report.xingai.app.
You ship a verifier. Dashboard gets Accept / Reject. Engineer patches the verify JSON with review=accepted. Next CLI run overwrites the cache. Human work vanishes. Trust evapora…
On July 11, the Robinhood MCP gateway story was simple: the gateway blocked every write order on purpose. G1, G6, and G7 were real code. G2 through G5 were named, explicit fail-…
Public posters love a clean ladder: Plan → Build → Validate → Operate, twelve boxes, each with a Tools row. LangChain here, LangSmith there, Docker at the end. Fine as a checkli…
XingAI Engineering Communication Coach is not a Grammarly wrapper. The product question is: what should a non-native engineer practice today so tomorrow’s design review, inciden…
We finally wired Resend secrets into the Opportunity Radar GitHub Actions job and forced a send. The mail arrived. It also looked like a shell: date line, empty tables, a canned…
A question that comes up every time someone builds their first MCP server: it's just an HTTP endpoint calling a few functions — why not hand it an API key like any other interna…
[Part 1](2026-07-13-mcp-oauth-vs-api-key.md) and [Part 2](2026-07-13-mcp-auth-api-key-vs-oauth-pkce.md) of this series covered why claims-mcp-oauth-poc uses OAuth and what the p…
[Last post](2026-07-13-mcp-oauth-vs-api-key.md) covered why claims-mcp-oauth-poc uses OAuth 2.1 + PKCE instead of a plain API key — the short version being that MCP's real shape…
Yesterday's post covered [claims-mcp-oauth-poc](2026-07-12-claims-mcp-oauth-poc-same-auth-different-industry.md): real OAuth 2.1 + PKCE in front of four deliberately narrow tool…
Two [education guides](https://github.com/xingaiapp/xingai-enterprise-ai-design/blob/main/guides/2026-07-12-mcp-oauth-pkce-lab.md) shipped yesterday, walking through a full OAut…
[Invest AI ADR-028](https://github.com/xingaiapp/xingai-invest-ai/blob/main/docs/adr/028-robinhood-mcp-execution-gates.md) defined seven gates (G1–G7) a trade order must clear b…
The July 5 [Agent Firewall post](2026-07-05-agent-firewall-helpfulness-attack-surface.md) described a 0din-class block with risk 60. That number was honest — and incomplete.
ADR-003 promised four approval buttons: Approve once / Approve for session / Deny / Deny + add rule. The first three shipped. The fourth sat in the UI copy with no backend.
[ADR-003](https://github.com/xingaiapp/xingai-agent-firewall/blob/main/docs/adr/003-approval-workflow-ledger.md) specified four one-click approval actions back in the PRD. [ADR-…
Mozilla's 0din team [showed](https://www.tomshardware.com/tech-industry/cyber-security/ai-coding-agents-can-be-tricked-into-installing-malware-via-clean-github-repositories-mozi…
Every team argues about shipping fast vs. shipping well, as if the tradeoff were a law of nature. AI-native engineering teams are proving it isn't. The mechanism is first-pass y…
[Engineering System ADR-001](https://github.com/xingaiapp/xingai-engineering-system/blob/main/docs/adr/001-personal-memory-engine.md) defined the architecture. [Invest AI ADR-03…
The Opportunity Radar generates a detailed markdown report every week. Until now, it lived in a private repo. This week we added a newsletter worker that reads that report and s…
Meal AI v4 is a pure Next.js app — no Python backend, no database, all decision logic in TypeScript on the client. The [Decision Ledger schema](https://github.com/xingaiapp/xing…
The AI engineering discipline has moved fast. We went from prompt engineering to context engineering, then to harness engineering, and now to loop engineering — where systems pu…
Meal AI is the health domain of XingAI — nutrition planning, meal recommendations, and caregiver support for family members with specific health conditions.
Polymarket AI scans prediction markets, estimates edge, and sizes positions with fractional Kelly. The tempting failure mode: wire a cron job to the CLOB and let edge > 5% auto-…
After [ADR-026](https://github.com/xingaiapp/xingai-invest-ai/blob/main/docs/adr/026-decision-engine-integration.md), Invest AI can show two different regime signals on the same…
Robinhood [Agentic Trading](https://robinhood.com/us/en/support/articles/agentic-trading-overview/) lets connected agents read portfolios and place trades — including auto-execu…
Invest AI ships a production dashboard with macro consensus. Decision Engine owns multi-timeframe technical scoring and backtests. Two repos on purpose — [ADR-002 on the Decisio…
Robinhood's [Agentic Trading MCP](https://robinhood.com/us/en/support/articles/agentic-trading-overview/) lets third-party agents read portfolio data and place trades in a dedic…
Insurance adjusters don't guess. They read the claim, pull the policy, check history, apply rules, and decide — with paperwork that shows why. Our new POC automates that workflo…
We now maintain 10 Cursor Skills in [xingai-engineering-system](https://github.com/xingaiapp/xingai-engineering-system) and plan MCP servers for Invest AI data and (eventually) …
The marketing site puts locale in the path (/zh/…). Product apps like Research AI switch language in the header and persist to localStorage. SEO still has to ship.
After you type a topic, Research AI returns a verdict and a 0–100 Learning ROI score. Both come from the same worker run — not from the browser guessing.
The first version of our daily-picks email worked technically. It read the worker cache, selected top signals, and sent an email. But the product shape was wrong: it looked like…
Model Context Protocol (MCP) gives agents a standard way to discover and call tools. That part is easy to demo. Production is harder: scoped permissions, chained servers, multi-…
[T Today](https://t.xingai.app) reads brokerage screenshots and returns a 做T plan — buy zones, sell targets, cash notes. A confident wrong ticker is worse than “I can’t read thi…
[T Today](https://t.xingai.app) ships rule-engine logic, Prisma schema, and OpenAI vision routes in one Next.js repo. TypeScript can look fine in the editor while tsc fails. A r…
We ship more than one app. Meal Coach, Cook AI, Travel AI, Invest AI, the marketing site at xingai.app — each repo had its own habits for navigation, languages, themes, and lega…
[T Today](https://t.xingai.app) is the screenshot-first coach: upload brokerage holdings, get a 做T plan with buy zones and rule checks. It’s paper-only — no orders.
T Today’s holdings coach uses OpenAI with responseformat: { type: "jsonobject" }. The model returns a big JSON blob with zh and en blocks. Users switched language — still empty …
In 2023, a lot of AI engineering looked like prompt writing. You opened a chat box, tuned a system prompt, added examples, and tried to make the model answer better.
It is easy to assume that “Connect Git Repository” on Vercel only works for public GitHub repos — especially when your dashboard still shows Connect Git because you have been sh…
[Invest AI](https://xingai.app/apps/invest-ai) answers “what does today look like?” — signals, dashboard context, and (eventually) richer decision surfaces. A different question…
[ADR-004](2026-05-12-v1-hosting-fly-io.md) explains why we picked Fly.io for the Python side and Vercel for the Next.js side. It does not replace copy-pasteable commands, projec…
This post documents a decision we accepted, not something already running in production. It belongs in the same story as our CQRS market cache: derived data should have one writer.
Invest AI surfaces BUY / HOLD / SELL style signals and confidence scores. In many jurisdictions, anything that looks like personalized investment advice can drag you into regula…
Our V1 system uses a fallback chain: try Ollama (local, free), then Gemini, then OpenAI. Whichever model responds first wins. It's resilient, but every model does the exact same…