# Mark Sulleza

> Technical lead with eight years in software. Builds AI tooling that is cheap to run and hard to misuse: a DeepSeek-native coding agent, a hybrid model gateway for Claude Code, and MCP servers with hard limits.

The site at mark-web.com is a vintage desktop interface; every page below also has a plain Markdown version. Facts on these pages come from each project’s own README and tests.

## About

- [About Mark](https://mark-web.com/about.md): who he is and how he works
- [Old BMWs](https://mark-web.com/bmw.md): the cars he wants to own and work on, and why he hasn’t bought one yet
- [Work history](https://mark-web.com/work.md): SensorFlow, Etch, Refocus, Talino Venture Labs, Infor, Finastra; BSc Computer Science, UP Los Baños

## Projects

- [px](https://mark-web.com/projects/px.md): A coding agent built around DeepSeek's prefix cache, where the repo's own tests decide when the work is done, not the model.
- [Hybrid gateway for Claude Code](https://mark-web.com/projects/hybrid-gateway.md): A local LiteLLM proxy that runs Claude Code with a strong planner and cheap executors, and tracks what every session costs.
- [Waypoint](https://mark-web.com/projects/waypoint.md): A debt-payoff and multi-currency cashflow planner that walks every account forward day by day, with an MCP server so an assistant can record money through the same checks as the app.
- [personal-gmail-mcp](https://mark-web.com/projects/personal-gmail-mcp.md): A read-only Gmail connector for Claude Code. One OAuth scope, a refresh token in the macOS Keychain, and email the agent treats as data.
- [ESP32 air-con controller](https://mark-web.com/projects/esp32.md): An ESP32 that learns an air-con remote's IR codes, replays them, reads the room, and runs its schedules even when the internet is down.
- [Tapos na ba?](https://mark-web.com/projects/taposnaba.md): "Is it done yet?" A phone-first tracker for the household's cat-care tasks, with photo evidence and a review queue.
- [Solar inverter dashboard](https://mark-web.com/projects/solar-inverter.md): Live telemetry from a hybrid solar inverter whose Wi-Fi logger allows no local access, read off its own cloud uplink on the LAN.
- [kiln, a launcher for the TrimUI Brick](https://mark-web.com/projects/kiln.md): A replacement home screen for a retro handheld, as one static C binary that draws straight to the framebuffer. It runs on the stock firmware, and the launcher lives entirely on the SD card.
- Debt tracker: Waypoint’s predecessor: a personal cashflow and debt-payoff tracker across two currencies, with a month-by-month projection, scenarios, sensitivity and Monte Carlo runs. (private, no public page)
- Private lending projector: Tracks private loans, projects earnings, and models what-if scenarios for deploying capital. (private, no public page)
- Sakto: Budgeting for bi-weekly earners in the Philippines: upcoming bills against paydays, instead of a log of every past transaction. (private, no public page)

## Writing

- [A read-only Gmail connector for Claude Code](https://mark-web.com/writing/read-only-gmail-mcp.md): One OAuth scope, a refresh token in the Keychain, and email the agent treats as data. Also why this site has a privacy policy.
- [Seven weeks with a self-improving agent](https://mark-web.com/writing/living-with-hermes-agent.md): Nous Research’s Hermes on my Mac: 295 sessions, 105 skills of which it wrote 41, a bot per role, and why it now asks before it runs code.
- [Letting an assistant record my money without letting it break the books](https://mark-web.com/writing/waypoint-mcp-checked-writes.md): Waypoint’s MCP server writes as the signed-in user, through the same database checks as the app, and answers every write with what it changed.
- [Two agents, one memory, and why I re-read every write](https://mark-web.com/writing/two-agents-one-memory.md): Hermes and Claude Code share one pool of notes. Mining 148 chat threads into it produced about 400 writes; 361 of them needed fixing when checked against their sources.
- [px: a coding agent built around DeepSeek’s prefix cache](https://mark-web.com/writing/px-deepseek-coding-agent.md): A 30-turn session costs $0.0149 off-peak. Most of the saving comes from never changing a byte the cache has already seen.
- [What DeepSeek’s API won’t tell you](https://mark-web.com/writing/what-deepseeks-api-wont-tell-you.md): Notes from building px: cache hits at 2% of a miss, a model that spent its whole output budget thinking, images that vanish, and the $16 day that taught me to price a sweep before running it.
- [Running a 27B coding model on my laptop](https://mark-web.com/writing/local-27b-model-on-a-laptop.md): Qwen3.8-27B on an M4 Max: 20.8 tokens a second, a 20,000-token floor before any work starts, and the defaults that quietly broke it.
- [Running Claude Code on a cheaper stack of models](https://mark-web.com/writing/claude-code-hybrid-model-stack.md): A local LiteLLM proxy where a strong planner hands bounded coding jobs to cheap executors, with spend tracked per session.

## Optional

- [Everything in one file](https://mark-web.com/llms-full.txt)
- [GitHub](https://github.com/mSulleza)
- [RSS feed](https://mark-web.com/rss.xml)
- [Privacy policy](https://mark-web.com/privacy)