[Weekly] What are you building this week?
What are building at work or on the side? Drop a link if it's public!
85 comments
this is a clever way to solve the cold-start problem. one subtle issue though: the model's initial ranking also determines which comments get enough exposure to collect votes, so the eventual human signal isn't entirely independent of the judge.
have you considered reserving a bit of randomized exposure to measure whether the model's quality judgments actually correlate with reader preferences?
Definitely want to test that out once there is enough activity to get some statistical significance!
Also I think thumbs up/down are not working properly (or nobody liked anything ?) I currently only see my own votes
So the design of this forum is a bit different than traditional upvote / downvote forums like Reddit.
I’m using a decision model to grade every thread / comment and that accounts for 100% of the ranking in the beginning. Then over time as more people view the comment more and more of the score weighting shifts to upvotes / downvotes.
This gives more incentive to comment on older threads (since a high quality comment will still get sent to the top immediately), and improves ranking overall (LLM judge is a more sane default than waiting for a few people to scroll to the bottom of a long comment thread or look at the “New” feed).
Will tweak this a bit over time, I'm also working on an About page atm that will explain this a bit more clearly!
Okay, that's extremely interesting ! Would love to read and understand which decision model you selected and how you configured it.
That's the first live example of a decision model I am witnessing :)
I'm teaching AI agents to run sophisticated simulations of nuclear fusion devices called tokamaks. Running simulations is not difficult, so an agent can figure it out by reading documentation and source code, but tokamaks are complex, and very often to achieve precise simulation a physicist or AI physicist needs to generate and test many hypotheses. AI can do this very well, but in many cases, it is not enough to have only one agent. Instead, I build chains of agents that generate hypotheses, review and prioritize them, independently estimate their impact, and so on.
woah! this reminds me of when I was at CERN, it seems a lot like the AI models the particle physicists are working with there, not just to simulate collisions but also filter petabytes per second down to MB/GB in multiple stages.
one big challenge they always face is the mix of "filter the right data out of it to train on", "get all the inputs and outputs right", and "cross-check the most valuable outputs (human)"... is it the same battle? do your experiments test real-time control / temporal constraint hypotheses too?
The timescale of tokamak plasma processes is 10^-5 s, so we have enormous data processing problems too, but so far I have mostly focused on validating our simulator (old scientific code, Fortran + Python) against the experimental data.
ahh, makes sense! when you say "validating", are the hypotheses generated like proper hard-math / logic, or behavioral, both, or neither i.e. just oracle cross-check to make it go "yep good enough"? given how chaotic the mystical models of the underlying EM/plasma equations seem, I'd imagine some blend of tree exploration in that...
(and if you need someone to pick brains or assist w/ what I'm guessing would be high-vol stream processing btw, lmk! that stuff is super fun to crack at the bottom end, know some folks too 🍻)
This is super cool! Are you finding AI helpful in reducing the cost of running simulations too or primary for generating hypotheses?
I'm working on TLDR Reader (kingfisher.tldr.fm), a place for fresh tech news.
Feel free to browse or search (you can use queries of arbitrary length). You can make one-off searches or use them to tune your feed. The stories are expandable, so that you can get the most important information about a story without clicking through. We want it to feel like a TLDR newsletter, but more customizable and even fresher.
All comments welcome! As with this forum, everything from bug reports to interface details would be great to hear about. I'll be watching this thread closely.
I'm building software for terabit-level free-space laser links to manage AI infrastructure! It's a hugely strange challenge that I never thought I'd be working on, but it's a cross functional system that involves distributed systems, network engineering and lasers :)
been building KRYN, an open-source coding workspace running Qwen3.5-9B locally on Apple Silicon via oMLX + OpenCode.
the interesting challenge isn't getting a 9B model to write code. it's making it reliable across long-running tasks: preserving file/check evidence through compaction and restarts, managing memory pressure, and independently verifying outcomes instead of trusting the agent's "done" message.
one surprising result: stabilizing the system prompt cut new input tokens by 60%, but strict task acceptance dropped from 3/4 to 2/4 in a small dev eval. reverted it.
curious how others building agent harnesses are measuring actual reliability beyond benchmark scores and successful tool calls.
Building an AI powered digital tutor for high school and college students with ADHD to save them from the mediocrity and complexity of LMS systems like Canvas.
Been working on this for about 9 months now. Solo project and company made entirely possible by Claude code. Most fun I’ve had in years. Psyched to have one school already piloting it and dozens of students using it daily.
I’m working on Walmart’s internal AI coding agent, we have several surfaces, CLI, JetBrains plugin, VSCode plugin, Web, Desktop app… and layers to all of those! We are building a common harness to sit under those surfaces and it has been a big fun project to tackle!
the common harness across that many surfaces is a really interesting problem. i've been building a local coding workspace on OpenCode, and one thing i've learned is that session continuity and verified task state are two very different things. a session can resume perfectly while its previous test results are already stale.
are you centralizing execution state and verification in the harness, or letting each surface manage some of that independently?
this is sick, glad to see the big retail giants I worked w/ in past roles moving faster!
I helped a big telecom solve an adjacent problem ("anyone with genai is a developer, so how do we ensure those apps are governed and integrate with auth & data sources correctly"), so I'm curious:
how does your harness hook into company context? is it modular core + integrations, w/ each one wrapping any messy ol data source?
that's what we did and it worked very nicely, but know there's a dozen ways to do it so I want to hear what works
Honestly so jealous of people that get to build coding harnesses, feels like such an interesting design space. Are you guys using something like Pi as a base or fully rolling your own?
I'm working on an activity logger app! It sound simple but I specced it out... probably way too much, to be able to have a massive amount of features that I personally want 😅 I started this project in a more "pre-AI" world (aka 2024) and so it's been wild to deal with scope creep and figuring out which things I can hand off to agents, which things I need to rework for today, and so on. It's fun!
Also, my own newsletter (cassidoo.co) is a passion project of mine. I've been working on it for almost 10 years and it's just a weekly, personal newsletter that has dev articles of the week, cool random links, a practice coding question, a joke, and some other goodies!
I'm curious about the activity logger! It reminds me a bit of a project of mine that (i) started pre-AI, (ii) is a sort of practical everyday use kind of application, (iii) came to be because I wanted one of its type with a certain big custom feature set, and (iv) is very amenable to piecewise enhancement with / replacement by AI, though it's not at all trivial to figure out which pieces, and how. It's a very interesting category of application to me, largely for technical reasons but also for emotional ones (they're often passion projects, and grafting AI onto a passion project one started without any idea that AI would be what it is can feel strange).
I'm looking forward to reading more about this project as you choose to write about it.
Launched this integration with ChatGPT and Coinbase! https://x.com/coinbase/status/2122691614426751448?a=-Ls31L_P3v7hQ5cBIgA&s=46
Slowly then all at once, ChatGPT is becoming the new app store!
I think this will actually be huge for trading in particular because order books look super scary and chat is probably already much more sophisticated than the average retail trader at getting decent execution.
Yup agreed, very bullish on agentic trading and the idea of taking high-level strategies and automating them without the user having to get deep into the weeds with order books and charts and so on and so forth
iceberg-mcp-server (github.com)
iceberg-mcp-server is an MCP Server for Apache Iceberg, enabling users to read, query, and manipulate data within Iceberg catalogs. It supports reading and data manipulation using catalog types supported by PyIceberg, and supports SQL queries using catalog types compatible with DuckDB.
Also just put out my first blog post at https://www.dragonejt.dev/, but I need to find more time to write stuff lol.
This week I'm mostly working on advocacy around the AI version of the SDLC and the OpenCode harness we are building for the company. Turns out it's almost as important as the quality of what you're building in AI.
In this 2026 time we always put in the spotlight AI enthusiast people, but in most large company there are still a lot of people that are that much AI friendly. Working on their conversion. Wololoooh (old ref)
Many reports from every conversation I have at big tech and big media and even startups around #NoAI boils down to therapeutic counseling.
Demonstrate AI first successes to build confidence; be a transparent mentor and educator to build trust; pursue corporate goals through learning, using, sharing, and aligning to build momentum.
multiple things! the most interesting two so far:
Orbweaver, an autonomous satellite constellation platform for commercial & defense that self-keeps any shape of clustered satellites (I see you cowboy!), handles collision avoidance (I see you orca!), and maximizes ground link via space networks + GNSS integrations
-> virtual lab complete, moving towards physical validation, engaged w/ more design partners
vergo, experimental Golang engine/lang modification inspired by https://arxiv.org/abs/2608.22630 that bakes static, temporal, and distributed-system verification directly into the language (think a mix of Hoare, TLA+, and SMT with reinforced type checking/assignment)... because I can only take my agent stack so far building e.g. orbweaver or client software w/ minimal+focused behavior/property-driven outcome testing guidance in multiple review rounds before the token use & time frustrates me 😂 -> lots of nuanced problems, will be open-source
Leading copilot and agentic AI R&D as part of our advanced AI development acceleration initiative, e.g., some recent work includes:
two main projects, with a third that's updated less frequently now but underpins all my other work..
first, my own agent orchestration system in rust, crusteo github (github.com) - overview vid on youtube (youtube.com)
second, the fastest way to work with parquet, period - zpq github (github.com) - project site (zpq.chicagobuss.dev)
third, my agentic doc & task store, tracker (github.com)
it's been cool watching our team at work adopt tracker. It's pretty light on features but that's kind of the point. Agents figure it out really well, and start coordinating with it beautifully.
I also find it interesting that they're all in different languages - right tool for the job has never been more possible. Lots of interesting debates about software factories at work as I'm sure most of you are having as well.
the minimal task store idea really resonates. something i've run into with coding agents is that they can preserve a perfectly coherent task summary while losing track of what's actually been verified. i've been experimenting with carrying file hashes and observed check results across compactions rather than trusting completion claims.
does tracker distinguish agent-reported progress from independently verified state, or is that intentionally left to the agents?
100% - I hit the same problem all the time before tracker, and even with good use of it, still run into it sometimes.
But with a few simple rules and nudges, it can be a pretty amazing companion for anywhere between one and hundreds of agents..
I use one workspace in each of my personal and company trackers for recording every agent turn in a folio per agent, fed through harness hooks. This is great for deeper analytics and siphoning off what the agents are all doing for deeper analysis.
The next use case is a rudimentary task queue - agents can tell another agent "you have assignment x" and the receiver takes the task and owns it until it's done, which automatically signals the primary agent.
Finally, the standard way to use it is to just say things like "make a doc called foo in the bar folio under workspace baz" and it'll do that. There's no more hierarchy than that; anything else is just labels, and some of those are automatic (like your bearer-token for author).
It's pretty delightful -- I like to think of it like "notion, but stripped down the bare minimum, lightning fast, and self-hosted so as cheap as you need"
Chipping away at my agentic hardware harness Panio (github.com); it's built for agents to be able to autonomously control systems to do automated testing, adding support for new hardware to your favorite operating system, and anything else you need an agent to be able to do without you having to swap cables or push buttons. My favorite party trick is to have it use a KVM to futz around with the BIOS settings on a PC.
The recommended / tested hardware section is super helpful, I wish more libraries came with a "this is exactly what I use" section!
The recommended / tested hardware section is super helpful, I wish more libraries came with a "this is exactly what I use" section!
I figure it's only fair to call out what I end up using myself! I have ended up trying so many things that didn't work well, and I'd feel bad if people ended up tripping over the same problems I did.
I'm building GSD Task Manager (gsdtaskmanager.com), a free, local-first Eisenhower Matrix app for web, iPhone, iPad, and Mac. Tasks stay on your device by default, sync is optional and doesn't need an account, and there are no analytics or ads.
This week I'm bringing web, iOS, and Android to feature parity so the native Android app can leave closed testing on Google Play. I'm running that sprint with a team of Claude Code agents, one per platform repo.
The part this crowd might like is the MCP server. One command (
npx gsd-mcp-server --setup) lets Claude list, search, create, and complete tasks and pull deadline analytics. Writes support a dry run, so you see the change before it happens. The web app and MCP server are both MIT-licensed.Happy to compare notes with anyone who has tried getting an agent to manage their to-do list.
Internally we use Linear for project management but I don't think anyone really logs in anymore, almost all usage is via MCP!
How are you syncing without an account?
Good catch, and my wording was off. The app doesn’t need an account, and sync is off by default, so your tasks stay on your device. Turning on sync means an OAuth sign-in with Google, GitHub, or Apple, so there’s no GSD password to manage, and the server keeps a copy for your account. If you’d rather not use mine, you can self-host the web app and sync backend with the Docker stack in the repo.
Linear going MCP-first is great to hear. Does anyone on your team still open it for planning or roadmap views, or has MCP replaced that too?
I don't think our usage is complex enough to need a full blown version of those views, I can imagine much larger teams still needing a UI for that type of stuff!
okay hear me out... as almost all of our agentic tool usage is via CLI tools (notion, playwright, linear, cloudflare, aws, and so on) because it saves lots of time and tokens + makes validation easier for multi-call sceipt usage from experience
drawing from inspo from The Pragmatic Engineer's Pulse: I think there should be a weekly (or perhaps daily at scale) submitted debate topic
this made me think "MCP or CLI?" would be a sick thread to see everyone speak to
Count me in for that thread. My current split is CLI when I’m scripting or running a loop, since kbruce is right about the token savings and easier validation. I reach for MCP when the agent has no shell, like Claude on the web or mobile, or when a write needs a typed schema and a preview first. That second case is why I added dry-run support to GSD’s MCP writes.
Dan, if you start it, I’ll bring the first hot take.
The iOS app for Muesli (muesli.ing), a reader you talk to. The part I'm on this week: reading a physical book with my phone locked beside me, saying what I think out loud, and having it turn into notes I can act on. No screen needed.
I really like this idea, I do find myself wanting my phone to just be locked and listening surprisingly often (especially while reading or doing something that requires full concentration), wouldn't be surprised if voice becomes the dominant mode of interacting with the phone long term.
For Muesli it's been a great way to bridge physical books and digital notes. Claude's first suggestion was to take a photo of the page you're on, but that pulls you right out of the flow and kind of defeats the point.
Voice is pretty universal too. We've got a Laravel web app and a native SwiftUI app, each with its own idioms, but talking is just the same on both.
OpenHands personas at home on a mini-PC, DGX Spark running vLLM/Qwen3.8-flash-next, and workstation w 3090. Over 1B tokens in the last two weeks but that's prompting it when I have time.
"Research" was the first one; I asked (OpenHands) chat to reference the default system prompt and write a new one. Then created a custom persona. It's working pretty good and seems to code less.
"ComfyUI skills" to convert and test comfy-skills to OpenHands and ensure environment support for a local comfy workstation.
"Warcraft II web multiplayer"
"Web In Browser Local voice chat"
"Bible Study"
... pretty much any random concept that makes room in my head for real world issues.
I'm building Ceviche (ceviche.ai), AI software that helps PE and VC finance teams split shared expenses across their funds, with the backup for every allocation, in place of excel spreadsheets. This week, we just rolled out a read-only MCP server, so finance teams can query their bills, allocations, and fund invoices from an AI assistant, and we are demoing it with our customers now
I’m building a judge model, a coding task classifier. Super fun. Feel free to give me ideas to get coding session dataset ;)
The top thing on my wishlist would be a smart way for a model to give me an ETA for every task. It's would be super useful to understand if a task is likely to take 5 min, 30 min or 2 hrs, even if the error bars are quite wide, I lose a huge amount of productivity by checking in hours later on something that took a subagent 5 minutes to finish.
We've had enough Hurdle customers (check us out (hurdle.ai)) ask for Slack integration that I've had to suck it up and build that out. Already cursed myself as someone is now asking for a Teams integration...
I haven't published this yet but we run a bunch of evals on complex sales tasks, I'll cut out a public version of the tasks to show how the frontier models are performing at real sales work. I had the chance to run Astra 6.2 today and found some good results (better substantiated claims in assessing deal progression) but interesting less financial/data details in an ROI case compared to Fable 5.2.
Great rule of thumb from back when I ran a SaaS company: anything you really don't want to do goes behind the Enterprise "Call Us" plan and gets a custom price!
I think Patrick Mackenzie once wrote something along the lines of "if you find yourself resenting a customer, you simply have not charged them enough".
Currently working Lyre (lyrehq.com) it lets you learn a new language by listening to podcasts you already love. Same shows. Same hosts. Just in the language you want to learn. It’s still early. Happy to share access if anyone wants to try it!
I'm working on a real estate development engine to try and solve the country's housing shortage. Lots of homes don't get built on available lots because the process of determining if a deal works financially is long, manual, and inaccurate - developers need to spend a lot of time and money just to find out a deal wont work (architects, civil, zoning/planning boards). We identify and monitor land parcels, optimize building(s) to the site, and then custom tailor outreach for developers. (We're also a robotic homebuilder, so we build them, we just need developer $$ so we keep our capital in our tech instead of tied up in the buildings/land themselves)
I am working on a replacement for zoom teams and meet, both as a product and a developer sdk. www.hiyve.io is live and had hosted over 300,000 calls. Our angle is to embed an mcp and agent that monitors facial expressions, speech patterns, mood and sentiment to provide real time coaching during sales calls. With an sdk for companies to build out their own video conferencing products.
I'm working on a better onboarding flow for my app ivee (ivee.jobs)
It helps non-technical people use AI like experts. ivee lives inside your AI tools, turning messy thoughts into sharper prompts, so you get better output, faster.
The new MacOS update stole the hotkey I was using for the new Siri update (command + shift + space)!!
If anyone has any hotkey ideas (that aren't already taken on Mac or by other popular apps) then let me know. Users can change the hotkey themselves in settings but ideally we have one that works for the majority of people.
Q: User validation and approval of autonomous agent actions but isolated authorization?
Agent runs autonomous all day then needs access to resources. Like AWS SSM Automation maybe? But isolated in a way agent has minimal access.
Mechanisms to give agents credentials that require human in the loop approvals. Maybe like a special action in AWS IAM permissions... "write-with-approval" along with templates that control presentation to user approval.
I would like this when running vscode copilot autopilot mode, OpenHands, etc.
And approval gates for:
GitHub merging to main.
AWS ec2 launch, add domain, etc.
write access to these approval templates themselves, so I can review an agent attempt to write these templates.
approve urls for domains.
mount/file access.
sql queries.
Or....
Is there something I haven't discovered yet that eliminates risk completely... like an AI company that provides full indemnity. Someone mentioned a popular standard called "roll the dice"?
Right now I'm working on this forum!
It's super early, so if anything feels clunky / broken please tell me! UX nitpicks, bug reports, etc. are all super useful.
Reply to this message and I'll prioritize fixes.