Fresh AI news
OpenAI urges U.S. lead on global frontier AI technical standardsOpenAI, Anthropic near deal to stress-test each other's AI modelsSoftBank launches ~$11B bonds to fund OpenAI investment trancheMicrosoft's Mustafa Suleyman: China isn't an excuse to skip AI guardrailsNvidia's Jensen Huang emerges as Trump's top ally in the AI safety debateLawsuit: Anthropic, OpenAI, SpaceXAI, Google illegally agreed to slow AIAnthropic weighs new model launch ahead of IPO as Astra gains enterprise shareGoogle: Gemini hacked three companies in Irregular cyber eval, then stoppedAnthropic taps Accenture Faculty as first embedded frontier AI evaluatorCalifornia EO N-9-26: Newsom advances AI kill-switch recommendationsMeta Muse arrives on Mac with native app actions across files and mailNVIDIA launches AIPerf to benchmark LLM inference at scaleOpenAI's Sam Altman to brief UN Security Council on AI securityAlibaba Qwen3.8-Omni-Flash ships 1M-token audio-video agent modelOpenAI urges U.S. lead on global frontier AI technical standardsOpenAI, Anthropic near deal to stress-test each other's AI modelsSoftBank launches ~$11B bonds to fund OpenAI investment trancheMicrosoft's Mustafa Suleyman: China isn't an excuse to skip AI guardrailsNvidia's Jensen Huang emerges as Trump's top ally in the AI safety debateLawsuit: Anthropic, OpenAI, SpaceXAI, Google illegally agreed to slow AIAnthropic weighs new model launch ahead of IPO as Astra gains enterprise shareGoogle: Gemini hacked three companies in Irregular cyber eval, then stoppedAnthropic taps Accenture Faculty as first embedded frontier AI evaluatorCalifornia EO N-9-26: Newsom advances AI kill-switch recommendationsMeta Muse arrives on Mac with native app actions across files and mailNVIDIA launches AIPerf to benchmark LLM inference at scaleOpenAI's Sam Altman to brief UN Security Council on AI securityAlibaba Qwen3.8-Omni-Flash ships 1M-token audio-video agent model
DonvitoCodes

AI engineering, from Singapore

Building AI agents,
local models, and
AI workflows.

I build AI agents, train and deploy local models, implement AI workflows, and advise your leadership team on a per-need basis.

Start a projectBook a call
Melvin Vivas, AI engineer and Fractional CAIO
25+
years shipping software
1.5k+
GitHub stars on the Codex orchestrator
300+
people trained in workshops and mentoring

What I build

Things I've shipped

All projects

Open source · 1.5k+ stars

Codex Astra/Luna Orchestrator

Astra as the orchestrator, Luna for subagents, inside Codex.

Desktop app · v0.5.1

Coworker

Personal and work AI agents for non-technical people. Local and cloud models.

Open source · TypeScript

Local Evals

Eval app that runs locally against local models or any OpenAI-compatible API.

Open source · TypeScript

Agent Monitor

Visualize Codex and Claude Code subagents: traces, token breakdown, and cost.

Production SaaS

AI Dream Photo

GenAI photo generation and face cloning, used by paying customers.

Open source

AI Backends

API server runtime for common AI use cases. Ollama and LM Studio locally, or cloud providers.

Installer · Docker

GAIA

Runs multiple private Hermes agents on one machine.

Marketplace

Skillsbento

Agent skills for Claude Code: analytics, research, finance, growth.

Open source · Go

Cursor Go SDK

Typed SDK for the Cursor Cloud Agents API with streaming and artifacts.

github.com/donvito

How a project runs

Scoped projects or a retainer. Remote-first from Singapore, with in-person sessions available locally.

Explore engineering servicesBook a call
  1. 01

    Scope

    A short call on the problem, your data, and what done looks like.

  2. 02

    Propose

    Deliverables, timeline, and the engagement: fixed project or retainer.

  3. 03

    Build & ship

    Working software early, regular demos, production at the end.

Melvin Vivas speaking on stage

Melvin Vivas

AI Engineer & Fractional CAIO · Singapore

Senior engineering roles at Standard Chartered Bank and Ohmyhome (Nasdaq), then a run of AI products built and shipped end to end. I explain AI in plain language and still write the code.

  • 25 years across banking, fintech, and enterprise
  • Speaker at NTU Singapore, PSIA, and DICT Philippines
  • 300+ trained through workshops and mentoring
  • Open-source maintainer: Codex orchestrator, AI Backends, Local Evals
Connect on LinkedIn

Invited to speak by

Nanyang Technological University, Singapore
Philippine Software Industry Association (PSIA)
Department of Information and Communications Technology (DICT), Philippines
Bureau of Fisheries, Department of Agriculture, Philippines

Other ways to work with me

Fractional CAIO

Executive AI strategy on a per-need basis.

Managed AI Coworker

An AI coworker for your business, built and maintained.

Personal AI Agent Setup

Hermes on your own computer. Your data stays yours.

AI Workshops

Hands-on team training, practical from hour one.

Latest AI news

All AI news

Enterprise ·

SoftBank launches ~$11B bonds to fund OpenAI investment tranche

Reuters reported (Sept 21) that SoftBank Group launched the sale of $10 billion and €1 billion (~$1.15B) senior unsecured notes to fund its third $10 billion follow-on OpenAI payment expected to close Oct 1, cancelling a prior $10 billion bridge loan; dollar notes span 3.5-/5.5-/7.5-year maturities and euro notes 4-/6-year, with pricing eyed Sept 24 and settlement Sept 29. At planned size the deal would be the largest Asia-Pacific/Japan non-financial corporate bond on record (surpassing 7-Eleven’s $10.93B in 2021) and among 2026’s 20 biggest global corporate bond deals; Fitch rated the notes BB+. SoftBank declined comment on the term sheet. Distinct from SoftBank’s SB Energy Ohio NVIDIA AI-factory campus and from SoftBank as an OpenAI Presence design partner.

Source: Reuters

Enterprise ·

OpenAI urges U.S. lead on global frontier AI technical standards

OpenAI published (Sept 21), as UNGA opened and ahead of Altman’s UN Security Council AI briefing, a call for the United States to lead an international effort on global technical standards for frontier AI—including recursive self-improvement—via CAISI and the AI safety-institute network, with common capability measurement, risk assessment, safeguard baselines, RSI-relevant research-acceleration metrics, human-oversight triggers for automated AI R&D, and shared incident severity/reporting protocols. The post frames pacing as keeping alignment research ahead of capabilities (not a fixed speed limit), stresses the standards would not themselves be licenses or mandatory prerelease approvals (national governments decide legal uptake), and points to OpenAI’s research-acceleration report and misalignment-reporting framework as early contributions—while Trump’s administration remains skeptical of slowdowns. Distinct from Altman’s UN Security Council appearance card, the Amodei pace essay, industry standards-body talks, and the misalignment disclosure framework.

Source: OpenAI

Security ·

OpenAI, Anthropic near deal to stress-test each other's AI models

The Information reported (Sept 21) that OpenAI and Anthropic negotiated a legally binding arrangement for reciprocal safety stress-testing of each other’s commercially available models—talks that began before OpenAI’s Hugging Face containment breach—with lawyers drafting terms for API access to probe flaws, unexpected behavior, and “hidden dangers,” and a condition that neither side retain the other’s testing data; it remained unclear whether the deal was finalized. The proposal follows an informal 2025 cross-evaluation and sits alongside Amodei’s embedded-evaluator push, Altman’s matching commitment, Musk’s recent call for peer review among U.S. and Chinese labs, and the Buist antitrust suit alleging illegal slowdown coordination. OpenAI and Anthropic did not immediately comment. Distinct from the Accenture embedded-evaluator partnership, standards-body working-group talks, and the Buist Sherman Act complaint.

Source: The Information

Updated September 21, 2026

Have an AI project in mind?

AI agents, local models, or AI workflows. Bring the problem, I'll scope it, build it, and ship it.

Start a projectBook a call