neo (@PineapplesDev) header
neo (@PineapplesDev) profile picture

ABC: Always Be Cooking 🧑‍🍳 👻 https://teapaw.org/t.co/CXLbGCp5Km 🍍 https://teapaw.org/t.co/KFgCrif9f7

⚡️vibe it with us 👉
Joined October 2012
Photo by neo (@PineapplesDev)Photo by neo (@PineapplesDev)Photo by neo (@PineapplesDev)Photo by neo (@PineapplesDev)Photo by neo (@PineapplesDev)Photo by neo (@PineapplesDev)Photo by neo (@PineapplesDev)Photo by neo (@PineapplesDev)Photo by neo (@PineapplesDev)Photo by neo (@PineapplesDev)
UCLA at Michigan, 3:30 on CBS. Mute the TV, put on the radio call, and you hear the snap before you see it. The Call fixes that: tap when you hear the snap, tap when you see it, and the radio waits. One page, built with Grok's app builder. wave-comet-clover-frost.grok…
22
Enterprises afford it the boring way: a cap in code and an owner. One client's background model spend has a daily cap. To clear a backlog of 212 graph extractions and 494 figure reads, the owner approved 3x for one day. The revert was in the PR before the raise.
How in the world can enterprise afford Anthropic tokens? They are so rediculously expensive. The cost isn't justified. It's just because they have the best product. Shows how much product matters.
1
12
Our meeting AI turned one call into three sets of to-dos. Then nobody did any of them. One meeting often reached it as three copies: the PDF, the AI notes, the recorder's report. It proposed tasks from each one. In a deal team's shared review queue, 47 of 64 proposals had no owner and expired after 7 days. Nothing got approved. What we changed: 1. One meeting, one run, one review card. Copies match on the calendar event first. Otherwise they need the same day in the owner's time zone plus the same title, people or start time. Same day alone never counts. 2. The card lands in the owner's own to-do list, not a shared queue. 3. Approving re-checks each task, so an existing one gets linked, not copied. 4. The recap stays out of the knowledge base, so the AI never counts a decision twice. This week everyone is launching AI that listens to your meetings and acts: swipe-to-approve decision cards, preconfigured AI chiefs of staff. The listening is solved. Deciding what counts as one meeting, and whose card it is, is the part nobody demos. When your AI proposes tasks from a meeting, who should own the approval: the person who ran the call or the deal lead?
1
16
Platform ROAS also counts things your customer never did. One program this week sent the ad platform 17 Schedule events for 12 people, staff rebooks and tests included. Only 8 were prospect bookings. Platform: about 10 bookings. CRM: 5. Fix: one event per person per program.
Platform ROAS is the most dangerous number you can run a business on. It only measures what that one platform can see, which is a fraction of your actual business. It skews in favor of whichever dashboard you’re looking at. As your ad stack grows across Meta, Google, TikTok, and YouTube, each platform’s attribution model is built to claim maximum credit for any touchpoint it played a part in. Customer journeys rarely close on one click. A shopper sees a TikTok video, clicks a Google search ad 3 days later, then converts off a Meta retargeting ad, and all 3 platforms report that same sale as their own conversion. Add up every platform’s reported revenue and compare it to what your Shopify account shows for that period. Add up what Meta and Google each claim they drove and compare it to what your Shopify account actually shows for that period. Stack in TikTok and Snap on top of that, or widen the attribution windows, and the platform total can climb past what Shopify shows for the same period. So now, alongside platform ROAS, we ask: - Is the number you’re looking at the platform’s reported revenue, or your actual Shopify revenue for that period? - Could more than one platform be claiming credit for the same sale, even when the totals don’t look inflated? - What happens to total Shopify revenue specifically when you pause or cut spend on one platform? - Is the lift you’re crediting to ads something you’ve tested with a holdout, or something the dashboards assumed? Bookmark this post to refer back to next time a platform report looks too good to be true.
1
13
Testers deleted files from a deal room. Our AI kept answering from them. On one client firm, 1,096 facts in its knowledge graph (773 of them financial) outlived the documents they came from. Root cause: when a file was replaced or removed, the database cascade deleted the evidence rows. It never touched the graph edges built from that evidence. Nothing marked the old version as superseded, and the readers never checked whether a fact still had a live source. What we changed: 1. A fact answers only while at least one piece of its evidence sits on a document still live in that project's vault. Every graph read enforces it: statements, contradictions, neighbors, paths. 2. Removing a document retires its graph in the same commit. 3. Deal terms and ledgers hide evidence from a trashed file or mark it "no longer in the vault", with no quote. 4. Cleanup jobs run dry first and wait for the owner's OK. Trash stays restorable for 30 days. 5. One end-to-end test on real Postgres checks 24 surfaces. Two of them fail on the old code. This week a judge approved Google buying Spirit Airlines' old emails and chats for AI training, with de-identification rules attached. Everyone is promising what happens to the source data. The harder part is what the AI already learned from it. If a seller pulls a document from your data room tomorrow, how would you prove your AI stopped quoting it?
3
1
62
In deal knowledge, versions bite before search does. A deal-team assistant we build kept answering from the draft LOI. The first fix recovered 3 of 15 questions: the draft's chunks filled the 24-chunk cut. Now the signed copy wins each field and the draft goes to history.
A mid market Private Equity firm in the US came to us with a clear ask: the intelligence platform mega funds pay $1M+ a year for, without their deal data living in someone else’s software. The problem was familiar. Deal knowledge was scattered across CIMs in folders, unsearchable calls and Zooms, banker emails in partners’ inboxes, and a lot of it only in people’s heads. Every Sunday night, an associate rebuilt it all by hand for Monday’s partner meeting. So we’re building one knowledge graph of the entire deal operation. Every fund, mandate, CIM, target, banker, call and email becomes a record, linked by what actually happened. Three jobs run on top of it: Sourcing: Every CIM is scored against mandate, sector, EBITDA range, check size and geography the day it lands, with a drafted brief waiting. Coverage: Every banker becomes a record: deals shown, hit rate, and whether their quoted multiple tracks the clearing price. Synthesis: Every call, Zoom and email on a target merges into one timeline, surfacing what no single call shows. By 7 a.m. every Monday, partners get one brief with drafted actions ready for approval. It runs single tenant inside the firm’s own Azure tenancy, with transcription on their own servers. Nothing leaves their walls. The result: the firm’s knowledge stays when people leave, and Monday runs on synthesis instead of Sunday night assembly. Three jobs. One knowledge graph. One accountable engineering partner. If you lead a mid market fund whose edge lives in calls, documents and relationships, reach out.
1
14
Someone asked our AI deal-team assistant for an 11am PST meeting "today". It booked Nov 7, ten days in the past. Then they asked "is it done?" and it made a second invite. Root cause: the chat never told the model what day it was, so it guessed. Reading production also showed the saved calendar time zone was empty for all 7 users. What we changed: 1. Every chat turn gets the person's date, their zone and the next 8 dates. 2. The calendar tool takes "PST" or "friday" in the person's own words. Code turns that into an exact time and zone. A start in the past is refused. 3. Every confirmation reads back the full date, weekday and zone. 4. Email and tasks got the same rules. "Is it done?" returns the existing task instead of making a new one. 5. Evals now grade the write itself (start, zone, attendees, due date) across 1,500 random clocks and zones. The old evals only checked the words, so they missed this. KPMG just ran a full audit with 100+ AI agents, and humans still stepped in for the key checks. Same idea at our size: the model never decides what day it is. If your AI books meetings or sets due dates, where does "today" come from: the prompt or the tool contract?
12
The first pass is the easy part to hand off. The line we still read ourselves is where the money comes from. Yesterday we changed a client checkout to price on the server from the catalog and ignore the amount in the request. If the paid total doesn't match, fulfillment stops.
The tech leads at a PE-backed US healthcare data company were reviewing every pull request by hand. They asked us to take some of that work off their plate. We built an AI reviewer to handle the first pass. By November: - It had reviewed 110 pull requests on their main product. - The median wait for its first comment was 22 minutes. - The team had fixed or closed more than half of its comments. The leads now have a first review waiting when they open a pull request, so they can focus on design and risk. Getting there took more than switching a reviewer on. The customer had three conditions: don't hold up delivery, keep source code and secrets inside their perimeter, and keep AI spend predictable. So the reviewer can suggest, but it can't block a merge. If it fails, delivery carries on. And on a new repository, it posts nothing until the team gives the go-ahead. It reads each change alongside the linked ticket and checks whether the code does what was asked. Every finding has to point to actual code in the change, or it gets thrown out. Then we asked the developers what was useful and what was annoying. Most of the changes from that feedback made the reviewer talk less fluff. Dismissed comments stay dismissed. Findings don't repeat. Each pull request gets one summary thread. That matters because a reviewer nobody reads isn't taking work off anybody's plate. We also launched their own AI platform inside their cloud account. It gives each product a key and a spending limit, and records model usage and cost without storing prompts. The reviewer is its first use case. A second repository is getting the same reviewer now, with one configuration file and no new infrastructure. Their own DevOps engineer deployed both systems with us and runs them day to day. The code is in their repositories, the changes were reviewed, and both systems have runbooks. They can keep running it without us.
22
Our AI financial planner saved "$7,000 a month" as $7,000 a year. The chat asked for monthly expenses. The save tool had no period field, and the service stores everything as annual. Whether to multiply by 12 was left to the model. QA found two more in the same feature: 1. A chat update added a second expense line instead of replacing the first. In the test case the record summed both, $174,000 instead of $90,000, while the agent said it had replaced the figure. 2. One page called an education goal "On Track ($0/mo)". Two pages later, the same plan showed a $225,000 shortfall. Two functions, two answers. The fixes: 1. Every line carries monthly or annual. Code annualizes, not the model. 2. The save returns a receipt, "$84,000 a year ($7,000 a month)", and the agent reads it back. 3. One shared calculation for every page that shows the goal. Anthropic's Claude Code team says verification loops are what let agents write most of its code. Same lesson here: the model never decides units. If an AI writes numbers into your system of record, where are units enforced: the prompt or the tool contract?
1
31
The renewal test for SaaS is shifting to: can the customer's own agent use us? Our MCP server listed 433 tools. ChatGPT caps at 128, so it couldn't load it. We now list 38 core tools and the rest sit behind search, with the same approvals.
We have 2 agents we pay five figures each for. They are great products and we recommend them highly. Great products. However 10K has slowly been taking over their functionality, day-by-day, bit-by-bit. He said not to renew in the coming months. He’s got it.
4
3
48
145 documents in our deal platform's vault had failed with no reason written down. Search could never return them, and nothing said so. A file marked processed with zero chunks read as indexed. A pending row hours old read as "still working." The agent had no idea those files existed. Trending today: SaaS turns into an invisible backend and AI agents run the conversation. Fine. Then the backend has to be honest about what it can't find. What we shipped: 1. A health report that lists every file search can never return, with 13 failure modes written before the code. 2. A repair job that is dry run by default. 3. A nightly alarm when any failure kind stays above zero. After the repairs: 145 down to 36. We chased the last 36 read-only. All 36 were files people had already put in the trash. No retrieval tuning until the index is honest. Buying AI over your documents? Which do you check first: answer quality, or what share of your files search can actually return?
2
31
Lions in Munich tomorrow, 9:30 AM. Mute Fox, put on 97.1 The Ticket, and the radio is ahead of the picture. The Call fixes that: tap when you hear the snap, tap when you see it, and the radio waits. Today: Michigan at Oregon, 3:30, FS1. wave-comet-clover-frost.grok…
44
Domain specific, and what you leave out matters as much. Our deal chat answered its own "read the source first" note as if the user wrote it: "going forward I will call..." Now notes are labeled app checks, and the next turn's history keeps only the question and final answer.
Overheard this morning: “We need to retain control of the investigation.” Ah yes I’m definitely in DC. “Optimizing the compaction summary turns out to be very domain specific.” I’m in DC?!
13
Our AI chat replied to its own safety check instead of the person who asked. The question was simple: how many files are in this folder? The agent read the folder with two tools our check didn't count as reads. So the app sent its "read the source first" note. The model answered the note: "That last question only asked how many files... going forward I will call..." PluraiAI just launched a simulator to stress-test agents against simulated users. Useful. Our break came from inside: the guardrail text leaked into the answer. The fix: 1. Every app note says it is an automatic check, repeats the person's question, and says not to answer or mention it. 2. After a note, the reply is held until it is final. 3. A reply that talks to the note goes back once. A second one is shown, so it can't loop. 4. The next turn keeps only the question and the final answer. Never the note. Guardrails are product surface. We wrote 6 failure modes and run them through the real router and tool loop. Rolling out AI in a portco? Ask your team: can a user ever see one of your hidden checks?
12
8
414
The small-company version of this shows up in tech diligence: what only works because one person is there? We ask who can ship a release, who holds the cloud and vendor logins, and who can rotate the production keys. One name on all three is the succession plan.
I really don't see this as the same as Jobs leaving Apple. Three main reasons: 1/ Jobs knew he was sick and had been planning a transition. Nvidia has no succession plan. 2/ Jobs was more a marketer and visionary than hardcore technologist. Jensen was all three, but the last part mattered a ton. 3/ Apple wasn't a "central bank" to tech in the same way Nvidia has become, so its success/failure wouldn't have quite as many ripples.
22
A payment webhook we run answered "not found" 41 times in one week. Nothing crashed. A flag said billing was off. a16z partners say AI agents now drive most web traffic. When software does the buying, the webhook is your record of the sale. And webhooks fail quietly. What happened on a client platform: 1. The webhook route checked an internal "billing on" flag before doing anything. 2. The flag shipped off, so every delivery got a 404. 3. Those payments still mattered. Purchases from a separate checkout tool set the weights for lead routing. 4. The processor saw a week of refusals and scheduled the endpoint to be disabled in three days. The fix: - Verify the signature. - Store the event once. A duplicate gets OK and is not queued again. - Queue it, answer OK, process in a worker with retries and a dead-letter queue. - Bad signature gets a clear rejection. A database failure gets an error, so the processor retries. No feature flag at the door. The route receives and records. Whether billing is "on" is the worker's call. If you run a portco, ask your team one question: what does every inbound webhook do when a feature is turned off?
1
2
38
What doesn't roll over between apps: Live Activities. iOS gives apps no way to ask for them. On a client app we start the activity when the user enters a capture flow, and show a row linking to Settings when they're turned off.
🌠 Supernova's second app hit a $1M revenue run rate after about three months of development. Turns out it helps when login, payments, and push all rolled over from their first app. Now they have three apps built with Expo and a great blog post about the experience. ✍️ @TheAnirudh, @shubhporwal24, and Sai Vasudatta. 📚 expo.dev/blog/supernova-ios-…
20