Brian Bulkowski
Menlo Park, California, United States
2K followers
500+ connections
About
I create things.
Sometimes - even often - I create teams. Sometimes I create…
Articles by Brian
-
Hungry AI: How Machine Learning Technology Is Transforming Payments
Hungry AI: How Machine Learning Technology Is Transforming Payments
2018 was a busy year for the payments providers. For example, last year PayPal—one of the premier global payment…
2 Comments -
Order Of Magnitude: 10x Can Change Everything
Order Of Magnitude: 10x Can Change Everything
An “order of magnitude” means a 10x difference. One inch to 10 inches is an order of magnitude, 10 miles per hour to…
1 Comment
Activity
2K followers
-
Brian Bulkowski posted thisi wrote an ai abstraction layer, Djinnite. when i did my first for reals integration, my code agent reached right around the api contracts and used an internal function. i went back to an agent coding Djinnite and said "this unit will likely only be used by agents. what will improve agents use and understanding". it said use typed python, and an explicit contract document. but type all functions, because ai agents will skip the documentation this is exactly how humans work typing your Python entry points has been a great idea forever. the same concept was at the root of Hack, a now abandoned language specifically built for both speed and robustness. we also discussed if the functionally the user dev reached for should be included in the external and decided against, but had a good reason. what i like about coding with ai is it's ability to introspect and root cause.
-
Brian Bulkowski shared thisWelcome to Djinnite https://lnkd.in/gzvxfA2M Djinnite solves a particular problem I had two weeks ago. A small personal project, but I need to make Gen AI requests programmatically. I tried a few "large requests" and all the services (pro accounts) choked hard with the number of downstream web requests (not context window). Therefore I need, well, a glorified loop, and a little cache. I mean, what could be hard? A little credential store, a little little layer. I found three main problems. Two I expected one I didn't. I expected API lock in. Therefore, a quick abstraction layer. But the fundamental concept is the same, right? Send a string, get a string? Can't be too hard? I then ran into price, to be expected. Some models are 250x (!) more expensive than others. I built a little system to fetch current models, and create a linear numeric comparison (which is inaccurate). For the pure joy of ouroboros, I used the framework itself to call AIs to get estimates. Doing that, I ran into "grounding" aka the knowledge cutoff, which is handled by the web service facets, but not APIs. I then had to build *that* into all the cutoffs, so I could get price estimates. Now I'm back to coding my (now two) core projects. Maybe it helps you too. Thus, I propose to you Djinnite.
-
Brian Bulkowski posted thisBeware the training cutoff! I had a serious tussle in building an AI centric project last week. I ran head-on into training cutoff. The fact I had missed: AI "chatbots" (web services from OpenAI, Gemini, Anthropic) naturally will do web searches on your behalf, if you ask nicely, or if they decide you need up to date information. It's a great feature, I rely on it now constantly (given that Google's index is now fully polluted with AI slop). It turns out the APIs (paid versions 100%) DO NOT do web searches by default. There is no consistency in the APIs for enabling "web search" between the different providers, and they're "weird". Gemini's interface doesn't support JSON return and web_search, but Anthropic does. And OpenAI doesn't have them at all, you have to manually ask what web searches would be needed to satisfy a request, do the searches yourself, run another prompt with the data. It turns out my agentic coding tools - CLINE and VSCode - use APIs - which I knew - but the implication is interesting. My code agents *WILL NOT* use APIs or features or products past their training window, full stop. This is known as the "brain in a box" problem. Your agent might not have heard about a new product, or a new API, that would make your dev faster. It was a more serious tussle than I expected, since my code agent kept telling me the APIs didn't exist or didn't work. It said the models I was asking for cost estimates for didn't exist, even though it had just live-fetched the working model set. Once I fully grokked the issue, I kept the system on track. I still don't know the training cutoff of the different models. I ended up using an AI web interface to get sample code (because those automatically do RAG / web search). I got that code working the old fashioned cut-and-paste way, I put the sample code in my project, then I had CLINE replicate the pattern into my code. Now I have an AI query abstraction in python that works like the chat interfaces, and supports web_search on the three major platforms (OpenAI, Gemini, and Anthropic). There's tooling to get the recent model set, there's tooling to get estimates of the cost of each model (some are 240x the baseline, so costing matters). Good news - this was a lot of complexity, and I (we?) bashed through it in 3.5 hours, wall clock. On to the next phase! THE KICKER - until web_search or RAG is integrated into the APIs and thus available to agentic coders, progress in technology will slow down. Innovation in new APIs and features will be unknown to coders, and they'll fight and tell you they can't write particular code unless you spoon feed them. Interesting, no?
-
Brian Bulkowski posted thisIs our next frontier of pronouns "thing" vs "one"? "I" vs "we"? "someone" sent me a mail? "something" sent me a mail? "I" sent you a mail? "we" (me and claude) sent you a mail? My brain and primary actions now are about 50% automated, I feel less comfortable saying that *I* did something, but "we" is weird too? Gender pronouns are so aughty....
-
Brian Bulkowski shared thisMotion I've been using your product for maybe 6 months now. It's not working for me. At least, not what I hoped, and not what I need. I need a "coach", I need someone to look through my projects and provide insights, I don't need another backlog manager. here's my writeup: https://lnkd.in/g5W5T8iT Happy to chat ~ even IRL, I'll get out to NYC maybe in a few weeks ~ If this isn't a solution for Motion, maybe it can be a solution someone else builds? I feel the need.
-
Brian Bulkowski posted thisModel Performance Matters AI "wasn't a thing" for decades because, although many core techniques were known, building high performance (eg, working) ML models was restricted to small specific cases. Then a series of great innovations (eg, transformers, now Mom, integrated search, digesting, etc) pushed AI into usefulness. Now, I am finding different AI services are *vastly* different in accuracy. Right now, OpenAI 4.5 preview is doing VERY POORLY. This morning it got 4 technical questions *in a row* wrong. Not by outright hallucinating - although it did hallucinate an option to a command - but by simply saying core things that were misleading (like recommending SDKMAN! for java version management on windows in powershell - where it doesn't work - SDKMAN does work on windows, but only in WSL). Some details are right - but then I would lose time researching and installing SDKman. Gemini 2.5 Flash (preview) has gotten every answer right on my tasks this morning. It correctly identified 'jabba' as the right package manager for my solution. Thus - as a plea - ***please continue to use multiple vendors, multiple models***. The battle for the best is still being fought - don't think "ai works" or "ai doesn't work" based on a single vendor. There's a *WIDE* difference in capabilities. [ And OpenAI went back to #1 for me about 4 weeks ago - not a gemini fanboi - the models before 2.5 Flash aren't good - I will play the field a little more with OpenAI's model set ] ( Which reminds me, since OpenAI is not doing well at the moment, to fire back up my Anthropic account and get it in the shortcuts )
-
Brian Bulkowski posted thisAI coding ("non vibe") My friends are saying, basically, that the "trap" is thinking you don't have to understand the code. You do. AI isn't at a point where it can make code congruent with your long term goals. Yesterday I had a pretty good session. It's embedded code, ESP32 code for audio. I was using Google products: Google Code Assistant plugin to VS Code, plus Gemini 2.5 flash (experimental). The two-part AI was perfect. Code Assistant was... half bright. It did lay down some pretty reasonable large sections of code. It made trivial blunders along the way. It was *really* good at template auto-correct. It failed at most basic refactor steps ("move structure from header X to header Y and include header Y where needed"). When I hit a spot where I needed a container structure for arguments, it was hopeless, I spent an hour just "cut and stitch" coding to get it right. There's no way to know the model under "Code Assistant", they don't expose that. This code AI, at least, is not capable of writing maintainable code. When you ask for modifications, it simply starts writing wrong code. Next stop, MCP (cursor) where I can pick and choose the models that work best for the code I'm writing. 2.5 Flash, though, was pretty good. I would discuss the theory of what I was doing, and any particular issue. I was dithering about how to use ESP-IDF ring buffers, or just code my own. I have now learned to read both the FM, and consult with AI, and triangulate between the two. But AI reminded me that I don't have to hand-code atomics anymore, C11 has all that. Then I hit a fairly hard bug. It was a pure embedded thing - ESP32's have several memory pools with different properties, and I really wanted my ringbuf's memory to be DMA capable for zero copy shovel to the I2S system (and the smart AI said it's required, Code Assist said it was a nice to have, smart AI is correct). But it turned out the allocator that allocates with memory specifiers often returns memory that is 32-bit addressable not 8-bit, and atomics (used within the ESP-IDF ringbufs, I didn't roll my own) don't work on that kind of memory. This was my immediate hunch, which I then proved, which I then solved, with the AI literally refusing to help "I already reviewed the code where you allocate that structure, it's correct. Please post your code in this other area". Bzzz. This is where you need to know your code, and what you're doing. I follow my own hunches, they're pretty good. We were also near some context limit, it started repeating answers and getting sideways. Overall, it was a good and productive day. With 2.5 flash, it felt like I was pair coding, or in a nice office with a good coworker willing to take a break and discuss fine points. At the end of the day, the code I got to was a serious step above, in quality, where I would have been without AI. And I had more fun. It'll be very interesting where we get to in another 6 to 12 months, because the "smart AI" was pretty smart.
-
Brian Bulkowski posted thisGood operational luck to all my friends in advertising. Seven days out from Black Friday, you should be all locked down, your runbooks at hand. I hope / sense it'll be a good year, still with desire for gifting families after so long apart, and with a greater appetite for online. Hopefully the root economic circumstances are in favor as well. Go get 'em!
Experience
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Other similar profiles
Explore more posts
-
Mai Medhat
If you're building with Pipecat, this one is for you. We just launched the Tuner <> Pipecat SDK 𝘱𝘪𝘱 𝘪𝘯𝘴𝘵𝘢𝘭𝘭 𝘵𝘶𝘯𝘦𝘳-𝘱𝘪𝘱𝘦𝘤𝘢𝘵-𝘴𝘥𝘬 Now you can plug in observability and analytics in minutes and get real visibility into how your voice agent is performing in production, at scale. Because shipping the agent is only step one. Understanding what’s happening in production is where the real work begins. Link in the comments 👇
12 Comments -
Rohit Shekhar
Live-streaming isn’t just about capturing frames, it’s about acting on them in real time. Read my co-founder Punit Agarwalla new blog, “How Stream-First Observability Achieves Live Streaming Reliability at Scale”, He shows how Chiron’s architecture flips the typical observability model: Detects root cause while the stream is still flowing. Cuts MTTR from tens of minutes → under 2 minutes Works alongside Datadog, New Relic & others : no rip & replace If you’re working in OTT, live video, or streaming platforms and still relying on logs-first observability, this is a must-read: [https://lnkd.in/g6-p5JBQ] #Streaming #Observability #SRE #PlatformEngineering #DevOps #LiveVideo
-
Romans T.
I was at GDC in San Francisco last week, and it reinforced something I’ve seen at conferences for years: Most people think the value is in the event itself. It’s not. The real connections—the ones that lead to ideas, opportunities, and partnerships—happen outside the main stage. Over dinner. In smaller conversations. In the moments between sessions. And often, they happen because someone did a little homework before the event even started. The people you’d most like to meet are often in the same place as you for a short window. That’s a rare opportunity. But if you just show up and hope for the best, you’ll likely miss it. A simple shift: Don’t just attend—prepare. Identify who you’d like to meet. Reach out in advance if you can. Be intentional with your time while you’re there. Conferences can be incredibly valuable—but only if you treat them that way. What’s one thing you do to make the most of events like this?
2 Comments -
Yossi Rotlevy 🇮🇱🇺🇦
🎯 The 2026 Vision: The First Voice-to-Action Automation Platform (Part 4 of 4) Over the past 3 weeks, I’ve shared the technical pillars of ActiveLens. To wrap up this series, I want to share our 2026 Roadmap and the shift we're seeing in the Voice Intelligence market. ━━━━━━━━━━━━━━━━━━━━━ 🚀 Our 2026 Roadmap: Driving the Work We are moving beyond "AI Note-taking." ActiveLens is a Voice-to-Action Work Driver. Traditional AI note-takers are passive advisors -they give you summaries to review later or let you ask questions about past conversations. More advanced tools attempt automation, but run it as a black box without visibility or control. A Work Driver is fundamentally different: it surfaces actions in real-time with full transparency, pre-fills forms you can review, and launches workflows directly on your desktop. It doesn't just document work or automate blindly - it drives work forward with you in control. ActiveLens 2026 Roadmap: Driving the Future of Work 🚀 💻 Active Desktop Integration: Transitioning from a "side tool" to a core utility that triggers real-time task wizards and notifications directly on your workspace. 📱 Mobile & WhatsApp Power: Bringing "Work Driver" automation to iPhone and Android to empower field reps to turn any conversation into an immediate action. 🌍 Global Public Launch: Moving from our successful closed beta to full public availability, making Voice-to-Action accessible to every professional. ━━━━━━━━━━━━━━━━━━━━━ 🌊 Why the Desktop Shift is the Future of Work The market has realized that to change how white-collar work happens, you must live on the user's desktop. Here is why the "Meeting Bot" era is ending: From "Side Tool" to "Main Driver": Browser-based meeting bots are passive. We are now seeing why a desktop-installed AI tool is a MUST. Recent moves like CLAWDBOT are just the tip of the iceberg. In the coming years, we will see more AI tools that perform work directly on your computer. Beyond the API: ActiveLens was built from Day 1 for these missions. We won’t stop where an API is available- we help automate any task using your voice. "Consent, Not Permit": No more waiting in "waiting rooms" for a bot to be admitted. You shouldn't need a host's permission to use your own productivity tools. Universal Coverage: ActiveLens works where bots can't- Slack, WhatsApp or even a quick FaceTime call. ━━━━━━━━━━━━━━━━━━━━━ 🔓 Transparency as Strategy In 2026, proprietary “secrets” aren’t your moat- execution is. We share our vision openly because we know our execution speed and clarity of focus will keep us ahead. If competitors follow, great. Users win, the market advances, and we’ve already moved to the next milestone. We didn’t stumble into this space. We built ActiveLens specifically for this moment, with this exact playbook in mind. Now watch us run. ⚡ #StartupJourney #VoiceAI #WorkDriver #B2BSaaS #Automation #2026Goals #AIInnovation
8 Comments -
Jack Vanlightly
There's been some talk around storage unification and zero-copy lakehouse integrations recently, and I wanted to better define these terms as well as look at ways we should evaluate different design decisions in this space. So I’ve published a new post: A Conceptual Model for Storage Unification Storage unification isn't about “one storage format to rule them all,” but virtualizing them into a unified logical model. In this post, I lay out a conceptual framework for thinking about storage unification (aka data virtualization). I explore: 🔹 Internal vs. shared tiering (and why shared tiering is so tricky). 🔹 When materialization makes more sense. 🔹 Client vs. server-side stitching. 🔹 Lifecycle and schema management across evolving formats. 🔹 Seven key considerations for evaluating different approaches. This is the groundwork for a set of posts where I’ll dig into how some real-world systems deliver storage unification, primarily about unifying real-time and historical data. https://lnkd.in/dBiS2kxJ
-
Rhonda Coleman Albazie
Misinformation and exaggerated claims The article notes examples of viral social-media claims (like false attributions to Nvidia’s CEO) warning about the impending end of wealth-building opportunities. These ideas reflect anxiety, even if not grounded in direct statements from leaders.
-
Jason Livingood
Final slide from my recent latency-related presentation - this one very much my own personal opinion with plenty of speculation. Fixed ISPs continue to deliver ever-better reliability, more bandwidth, better latency - at least to a demarcation at the edge of the home. But IMO user wants & needs have shifted and are now broader than the industry has fully recognized (though companies like CUJO AI® are exploring it). For example, a user may not be fully satisfied until they know that X application session on Y device and a point in time T is consistently great - across the full day, across all the devices they use, across all the apps they use, across the entire footprint of their home (or travels). Food for thought... #Bandwidth #Latency #NetworkResponsiveness #TemporalInternet #HolisticInternet #Broadband
5 Comments -
Keoni Mahelona
TLDR monopolies are a path to more scale and more profit. This thinking is Imperialism. It's Empire. It says let's own everything for "efficiency" or whatever word we need to tell ourselves to sleep at night versus let's help society thrive collectively via eg federated social networks instead of centralized ones.
2 Comments
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content