George Trammell

The problem

AI assistants are brilliant at planning and hopeless at following through, because the place your plans actually live is a closed app they can’t touch. Meanwhile every task app still treats your list as a flat pile of text, when real tasks belong to projects, wait on people and depend on each other.

HITList (short for Human-in-the-List) fixes both. It’s a task app built from day one to be worked on by AI as well as by you.

What it does

• A built-in MCP server. HITList speaks the Model Context Protocol, so Claude and other assistants can read your list, break a goal into steps, rebalance an overloaded week or turn a meeting transcript into action items, using the same tools you do.
• A personal knowledge graph. Every task links to its projects, people, places and notes. Ask “what’s blocking the launch?” or “what did I promise Sam?” and get an answer built from real relationships, not keyword search.
• Human in the list. Every change an agent proposes arrives as a card you accept, edit or flick away. Nothing moves without you, and every change records where it came from.
• Flip, don’t scroll. Your day is a deck: one task front and centre, so the next thing is always obvious and everything else gets out of the way.

How it’s built

• SwiftUI across iPhone, iPad and Mac from a single codebase, designed for each.
• SwiftData on-device, so it’s instant and works offline.
• CloudKit sync, private to your iCloud account, with nothing to sign up for.

Try it

HITList is in a private TestFlight beta for friends and family. Request an invite and you’ll be among the first to use it.

The idea

Stories and short-form apps want clips; your camera roll is full of long videos. Trimming them one at a time in Photos is tedious, and most “video splitter” apps want an account, a subscription and a copy of your footage on their servers.

SnapSplit does one thing. Pick a video, scrub a thumbnail timeline to drop your split points, and save every clip you want straight back to your library in one go.

How it works

• Thumbnail timeline. See the whole video at a glance and place cuts exactly where you want them.
• Pick-and-save. Export only the clips you want, all at once.
• Projects that remember. Leave mid-edit and pick up exactly where you were.

Privacy

SnapSplit works entirely on-device and only ever sees the one video you choose through Apple’s Photos picker, never the rest of your library. No account, no uploads, no tracking.

Try it

SnapSplit is in a TestFlight beta. Request an invite and start cutting.

What it is

Two little creatures that hang out at the bottom of your Mac menu bar and share one braincell. They only appear when you and a friend are both using Claude Code at the same time. They act out what each Claude is doing, and they nap when one of you stops.

What the twins do

Mostly they hang out and pass the braincell back and forth. Sometimes they act out the task at hand: gazing up to think, tapping their hands to type, squinting at a command. Whoever stops first goes to sleep, and five minutes later they both disappear.

A few things happen rarely: a hug after an hour apart, a high five for each hour spent working together, a slump when one of you runs out of usage. There are a couple of easter eggs too.

Click or hover to read what each of your Claudes is up to.

How it works

• Small enough to read. About 2,000 lines, nearly all Swift, compiled on your own Mac by the install script. No account, no analytics, no admin rights.
• Claude Code hooks. They report whether Claude is busy and what kind of work it is doing. Prompts, code and commands never leave your Mac.
• One encrypted line to your friend. Busy or not, the kind of work, a mood, finished to-dos and, if you allow it, a few words from Claude’s to-do list. ntfy.sh relays it and can’t read it.
• Light. Zero tokens, under 15 MB of memory, and no CPU at all while the twins are hidden.

Try it

The one-line install is in the README on GitHub. It needs macOS 13 or later, the Xcode Command Line Tools and Claude Code. Works for pairs only, for now.

MIT licensed. Not affiliated with Anthropic.

The problem

Getting an app onto the cloud still means knowing which of hundreds of services to wire together, and then writing the infrastructure-as-code to do it. That’s a wall for students, researchers and small teams who just want to ship.

What we built

CloudWeaver is an assistant that takes a plain-English description (“I’m a student who wants to launch a website to sell my handmade crafts”) and generates infrastructure-as-code you can actually deploy.

The interesting part is hybrid context retrieval. Pure vector search is good at finding similar text but bad at knowing that a load balancer needs a target group, which needs a VPC. So I led development of a retrieval layer that combines:

• a Neo4j knowledge graph of cloud services, their dependencies and valid configurations, and
• vector search over documentation and reference architectures,

orchestrated with LangChain. The graph keeps generated infrastructure structurally valid; the vectors keep it relevant to what the user asked for.

Outcome

• Won BU Spark! Demo Day (Spring 2024) with the Judges’ Choice award for AI development.
• Presented at DevConf.US 2024, Red Hat’s annual open-source community conference.
• The graph-plus-retrieval approach became the foundation for the knowledge-graph work I do now.

The problem

Herbaria hold millions of pressed plant specimens, many collected centuries ago. Each one carries a label with the species, where it was found and who found it, usually handwritten, often faded and in any of a dozen languages. Until that label is transcribed, the specimen is invisible to researchers. Doing it by hand takes years.

What I built

Working with the Harvard University Herbaria through BU Spark!, I built a multilingual pipeline that:

1. Finds the label on a high-resolution specimen scan, ignoring the plant, stamps and rulers around it.
2. Reads it, using OCR models suited to old handwriting and to scripts beyond Latin, including Chinese.
3. Structures it, turning the raw text into fields (taxon, collector, date, location) that drop straight into a collections database.

Outcome

The pipeline is open source, and six international labs now use it instead of manual transcription.

What it is

One server at home, run like a tiny cloud provider. It started as a place to keep files and grew into infrastructure a group of friends and family rely on every day, which means real uptime expectations, real support requests and real 11pm incidents.

Architecture

• Hypervisor: Proxmox VE on bare metal, with each service isolated in its own Ubuntu VM or container so one experiment can’t take down everything else.
• GPU passthrough: a dedicated GPU passed straight through to a VM for hardware video encoding, so streams transcode in real time without pinning the CPU.
• Services: media streaming and requests, private file storage and sync, a retro-game library and on-demand game servers for friends.
• Access: everything runs in the browser behind sign-up and authentication. Nothing is open to the public internet without a reason.

The dashboard

The part I’m proudest of is a custom dashboard that ties it together:

• Live sports scores for the teams people care about.
• A daily-games microservice with a catalogue of 500+ games, picking something new for everyone to play each day.

Why I run it

Because running real services for real people teaches things a tutorial can’t: capacity planning, backups you’ve actually tested, graceful upgrades and what to do when something breaks at 11pm on a Saturday.

Context

In early 2023, ChatGPT arrived in every classroom at once, and nobody had rules for it. Instead of a ban, BU’s Faculty of Computing & Data Sciences asked students to help write a policy that treated generative AI as a tool to be used well, not hidden.

My part

As a student co-author, I helped draft the Generative AI Assistance (GAIA) policy: what students should disclose, how instructors should adapt assignments and how grading should account for AI help. The policy was adopted by the faculty in 2023 and drew attention well beyond BU as one of the first of its kind.

It’s also where I started thinking seriously about provenance: if an AI helped, you should be able to say exactly how.