# Jeff Kent Buenaflor

> Principal AI engineer who designs and runs AI agent teams that build, test and review production software. Started as a UI designer and still designs the products he ships. Remote, any time zone (US, EU or APAC hours).

I design the agents, the rules they follow and the reviews that check them, then ship real products with them.
Client names are withheld where under NDA. Every number in the case studies comes from project records.

## How to contact or hire

- Email: hello@jeffkent.dev (the best way; I reply within one working day)
- LinkedIn: https://linkedin.com/in/jeffkentbuenaflor
- GitHub: https://github.com/jeffkent
- Open to: Full-time remote AI engineering roles (principal, staff or lead AI engineer, head of AI engineering), fractional AI engineering lead, agent fleet setup for a team, platform rebuild with an agent crew, product design lead for an app or website.
- Location and hours: I'm based in the Philippines (UTC+8) and work remotely in US, EU or APAC hours, 40 hours a week
- Machine-readable resume: https://jeffkent.dev/resume.json
- Full CV with employer names: on request by email

## Experience

15 years of product engineering. CTO and AI engineering lead since 2026.

## Abilities

- AI agents: Multi-agent orchestration, Model routing by risk and cost, LLM evaluation (evals), Prompt and context engineering, Agent memory, MCP, AI code review, Guardrails, Human-in-the-loop approval, AI cost optimization
- AI tools: Claude Code, Anthropic API, OpenAI Codex, OpenAI API, xAI Grok, Open-weights models (DeepSeek, GLM) via Ollama
- Engineering: TypeScript, Node.js (NestJS, Fastify), React, PostgreSQL, Python, Swift / SwiftUI, Kotlin / Jetpack Compose, Playwright
- Cloud and DevOps: AWS (ECS, RDS, Cognito, Lambda), Terraform, Cloudflare (Workers, D1), CI/CD
- Generative media: ElevenLabs, Kling video, Remotion, Blender, ffmpeg, Computer-use agents, Speech model fine-tuning
- Product design: UI and UX design, Product flows, Design systems, Interaction and motion design, iOS and web app design, UI review for wording, spacing and states
- Domains: Regulated fintech and crypto (wallets, KYC/AML, payments), Video streaming (players, DRM, ads), Media automation

## Case studies

- [Rebuilding a regulated fintech platform with a team of AI agents](https://jeffkent.dev/case-studies/fintech.md): As CTO, rebuilt a 9-app regulated crypto platform with an AI agent team: about 4,000 approved changes in under six months, independent AI review on every change, 10+ long-running agents cut from 45+.
- [Orchestrating AI agents to rebuild a streaming platform on web, iOS and Android](https://jeffkent.dev/case-studies/streaming.md): As the sole engineer, directed AI agents to rebuild a streaming service's website, back office, APIs and native iOS and Android apps, with 600+ automated tests across the native apps.
- [An AI agent studio that writes, renders and schedules short-form video](https://jeffkent.dev/case-studies/media.md): Built a three-agent video studio (director, executor, browser uploader) that wrote, rendered and scheduled 28 finished short-form episodes in about two weeks, each passing a 12- to 13-point QC.

## Articles

- [Code readability is no longer the bar](https://jeffkent.dev/writing/readability-is-not-the-bar.md): Why human readability stopped being the first review question for AI-written code, what replaced it (business-rule tests, independent review, security and rendered checks) and what cheaper models do not fix.
- [Where I still stand in an agent team](https://jeffkent.dev/writing/near-the-loop.md): Where a human still stands in a long-running AI agent team: events and clocks start the work, agents own staging, and one fixed gate stays human, the production release.
- [Guardrails for agents that run while I sleep](https://jeffkent.dev/writing/guardrails-for-unattended-agents.md): Eight guardrails for agents that run unattended, each from a measured incident: pre-tool hooks, a usage guard, three-result checks and an orphan sweep.
- [The merge train that landed 3 of 9](https://jeffkent.dev/writing/merge-train.md): Why a batch merge train for an AI agent team landed 3 of 9 trains in four days, and the per-change merge flow with scheduled checks that replaced it.
- [How my Claude, Codex and Grok agents talk to each other through plain tmux](https://jeffkent.dev/writing/tmux-agent-bus.md): How agents from Claude, Codex and Grok send each other messages through plain tmux: a verified send script, signed messages, one address per agent, and the incidents that shaped the rules.
- [The memory palace: how fresh AI agents start with my rules](https://jeffkent.dev/writing/memory-palace.md): How every new or compacted AI agent starts with the rules, the project state and its next step: a size-capped rules file, a note-per-fact palace with a generated index, task checkpoints and a per-prompt hook.
- [How I cut 45 AI agents to 10+ long-running lanes](https://jeffkent.dev/writing/45-agents-to-10.md): Why fewer, stronger agents with clear jobs beat a crowd of helpers, and the rules that made the cut from 45+ agents to 10+ long-running lanes safe.
- [235 billion tokens of agent work: subscriptions vs metered API](https://jeffkent.dev/writing/subscriptions-vs-api.md): How flat-rate Claude and Codex seats compared with metered API prices across 235 billion tokens: about $124,000 of work at list prices for about $4,000 in seats.
- [Benchmarking decision models before trusting them](https://jeffkent.dev/writing/benchmarking-decision-models.md): How I tested a small judgment model against a frontier model (16 of 16 vs 13 of 16), why the trivial baseline mattered, and why I retired it.

## Highlights

- About 31x value versus spend, Feb to Oct 2026: about $124,000 of AI work at API list prices for about $4,000 in flat-rate Claude and Codex seats (personal and employer-paid), 235 billion tokens.
- Decision models tested before trusted: a typed judgment model for routing and risk checks scored 16 of 16 versus 13 of 16 for a frontier model, at 26x fewer tokens. Retired after measuring that about two in three of its holds were unneeded.
- About 20x lower cost per coding task: benchmarked models and settings, then routed each task to the cheapest option that met the review bar, with no rise in review failures. Open-weights workers (DeepSeek V4.1 Flash, GLM 5.3 Flash, via Ollama and direct APIs) ran under frontier reviewers.
- Process: a planner splits each task, low-cost builders implement, an independent reviewer must approve every change, then the result is verified in the real environment.

## Optional

- [Full text of every case study and article in one file](https://jeffkent.dev/llms-full.txt)
- [Resume in JSON Resume format](https://jeffkent.dev/resume.json)
- [Home page](https://jeffkent.dev/)