Kenneth Hurley

Podonos just raised $2.4 million in pre-seed funding, and the signal is loud and clear: the voice AI industry finally has someone measuring the sound, not just making the noise. Based in Los Gatos, Podonos is building the infrastructure layer that decides whether voice AI feels real, human, and ready for primetime. The industry is racing toward a projected $47.5 billion by 2034, but growth without standards is chaos. Podonos is making sure the hype actually holds up when you press play. At the center is ✦Soohyun Bae, PhD in Computer Engineering from Georgia Institute of Technology and Y Combinator W22 alumnus. His track record stretches from engineering roles at Google Maps to leading AR mapping at Niantic, plus co-founding Bobidi and TickTock AI. He has seen what happens when technology scales without proper #guardrails. With Podonos, he's betting the future of AI isn't just about how smart the models get, but how believable and trustworthy they sound. Podonos evaluates models across naturalness, similarity, emotion, recognition accuracy, pronunciation, tone, and resilience. Add in #personaconsistency, and suddenly you're not just testing machines, you're testing their ability to pass as human. The twist? They deliver results in under twelve hours, when the legacy approach drags on for two months. In AI time, that's the difference between relevance and irrelevance. Investors caught the beat. Serac Ventures, led by Kevin Moore, took the lead, backed by NAVER D2SF, the venture arm of South Korea's NAVER, and KAIST Venture Investment Holdings. Together, they've now backed Podonos twice, with this $2.4 million round stacking on top of a $750,000 raise in March 2025. Total funding sits at $3.15 million, and every dollar is pointed toward scaling #engineering, refining #analysistools, and launching into the U.S., Europe, and Southeast Asia. This isn't theory. Podonos already serves Resemble AI, Play AI, and Sanas AI, hitting six-digit ARR by November 2024 with just four employees. Their secret weapon is reach: 150,000 evaluators spread across nine languages and thirteen locales, forming a human-in-the-loop platform that blends speed, rigor, and global coverage. That reach lets Podonos deliver evaluations at a scale and pace that no competitor has matched. The strategy is ambitious: expand into high-demand sectors like #healthcare, #finance, #gaming, and #advertising, then move beyond voice into #multimodal model evaluation, spanning video AI and large language models. The mission isn't just to check if AI works, it's to decide if AI works for people. #Startups #StartupFunding #EarlyStage #VentureCapital #PreSeed #AI #AIInfrastructure #Voice #VoiceAI #VoiceTech #Infrastructure #Technology #Innovation #TechEcosystem #StartupEcosystem #Hiring #TechHiring If software engineering peace of mind is what you crave, Vention is your zen.

6

1 Comment

This is how Large Language Models (like ChatGPT) work: GPT stands for “Generative Pre-Trained Transformer”. Let’s break down what this means. Generative: this means the model can write (or “generate”) new text. As opposed to simply analyzing or processing inputs, it can output totally new, unique text. Pre-Trained: when you hear about AI being trained on massive data sets, this is what they are talking about. In the pre-training phase, they give the model tons of data, which in the case of LLMs, it uses to learn the patterns of language as well as general knowledge. Transformer: this is the key that has made all modern LLMs and the recent explosion in AI possible. In 2017, a team at Google introduced transformers in a landmark paper called “Attention is all you Need.” Transformers allow the model to analyze every word in a sentence simultaneously in order to understand which ones are most important. (Previously language models tried to analyze words in sequence and wound up having memory issues). — Ok, so how does this all work? At the beginning they start with a shit-ton of data. They can’t just dump it all on the Large Language Model, it needs to be cleaned up and formatted properly. This takes a lot of time. Next, all the words need to be turned into numbers. Computers don’t actually understand words, so the words are broken up into “tokens” and assigned a unique number identifier, which the computer can use to reference that word. They then use these numbers to create an insane multi-dimensional map of the entire language. And when I say multi-dimensional, I’m not talking about 3D space like our puny human brains can comprehend, I’m talking about THOUSANDS of dimensions. They place these word-numbers (“vectors”) into a multi-thousand dimensional space map. This is essentially a meaning map that allows the computer to comprehend language. All existing words are placed somewhere on the map based on their meanings and their relations to each other. This is what it uses to predict words. The next phase of training is basically playing a prediction game where the model is given examples of sentences and asked to guess what the next word will be. If it guesses wrong, its “weights” are tweaked and then it tries again. This happens over and over again billions of times, until it gets really good at predicting words. Once it reaches a good place, its weights are fixed. There’s an additional layer of “fine-tuning” for personality, telling it what its role is, what style it should communicate in, and setting some parameters around how it should “behave”, but this is the basis of the mechanism it uses to generate text in realtime when you send it a prompt. — I’m curious if this makes sense or is still too technical? My goal is to make technical stuff like this more accessible, so would love any feedback if you’re feeling like: “what the hell are you talking about?”

3

Microsoft's AI chief Mustafa Suleyman said in an interview with the Financial Times that he predicts that most, if not all, white-collar tasks will be automated by AI — WITHIN THE NEXT YEAR OR YEAR AND A HALF. Immediate reaction: He's writing checks that Copilot can’t cash. Second reaction: He knows he's a white-collar worker, right? This myth that AI is ready to take your job is repeated often, mostly by people interested in selling it. They want to drive so much hype — make it seem so inevitable and indispensable — that they can take our salaries and divide it up with our bosses. It's a kind of theft, replacing you with an inferior product driven by an overwhelming marketing blitz. Actually, I find AI helpful for some tasks. I like to get preliminary research or run ideas by it. Sometimes, I create imagery for stories I'm working on. But the output is never great, especially once the novelty is gone. If you think an AI video is that great, I admit that I suspect your taste. And what I'm hearing from my other white collar creative friends is similar: AI is not public facing. It's a helper chat bot. That'd be OK if trillions hadn't been sunk into it. Maybe I'll be proven wrong as the tech improves, but it sure doesn't seem like it's going to improve that much in a year. I think AI pushers need to start getting realistic. AI adoption will increase when they stop being so creepy. Here’s something I truly believe: if you're a wealthy person who believes that AI will displace millions of workers, then you better be sure that you have a solution besides "lol, get a new job." Training programs. Unemployment funds. Basic income. If nothing else, do it out of self-preservation! Do you really want a whole class of displaced workers angry at you?

6

1 Comment

A reliable agentic workflow builder shouldn't separate those who know "how to build" and those who know "what to build". Instead, it should hand them paper and pen and let them create independently or collaboratively on a shared canvas. In Dify, there's always a space for both of them. #AgenticWorkflow #AIApplication #Dify

2

🚀 BoxLang 1.8.0 Is Officially Released! https://lnkd.in/eNWq9Zdg This is one of our most impactful releases ever, delivering next-generation HTTP capabilities, full SOAP/WSDL integration, dynamic class loading, and over 100 stability and compatibility fixes across the platform. 🌐 Revolutionary HTTP Client The new HTTP engine is built for modern distributed systems: * Fluent, chainable API: http("url").post().body(data).send() * HTTP/2 support with streaming and multiplexing * Server-Sent Events (SSE) for real-time applications * Connection pooling + persistent connections * Middleware-like response transformations This unlocks high-performance service-to-service communication inside your BoxLang apps. 🧼 SOAP & WSDL Integration (Enterprise-Ready) SOAP is still everywhere in enterprise environments — and BoxLang now speaks it natively. * Automatic WSDL parsing * Full method introspection & discovery * Fluent invocation: ws.getUserInfo(userID: 123) * Support for SOAP 1.1 and 1.2 Perfect for integrating with legacy systems, ERPs, banking APIs, and government services. 🔌 Dynamic Class Loading Extend your runtime on the fly: * Load JARs at runtime with no restart: getRequestClassLoader().addPaths(["/path/to/jar"]) * Enables plugin systems, modular apps, and rapid experimentation 📦 MiniServer JSON Config A simpler, more powerful configuration model: * Declarative server setups * Environment-specific overrides for dev/stage/prod * Cleaner DX for building BoxLang-based services 🐛 100+ Critical Fixes This release includes major improvements across the entire stack: * 30+ date/time fixes and enhancements * 15+ query engine improvements * 15+ CFML compatibility updates Stability and performance upgrades everywhere 💎 New BoxLang+ Modules Our ecosystem continues to grow with new optional modules: bx-redis • bx-couchbase • bx-csv • bx-ldap • bx-spreadsheet • bx-pdf • bx-rss • bx-charts ⬇️ Download BoxLang 1.8.0 Today: https://lnkd.in/dswq5wkZ 📘 Full Changelog: https://lnkd.in/eNWq9Zdg #BoxLang #JVM #HTTP #SOAP #WSDL #APIs #DeveloperTools #Java #DynamicLanguages #OrtusSolutions #EnterpriseJava #ModernDevelopment

2

Automation tools with agentic LLM support — same canvas, very different trade-offs. Zooming out, these tools still fall into two tribes: automation-first vs LLM-app-first — with one interesting hybrid. n8n Code available (Not open source). Automation platform that grew AI muscles. Best when the real work is moving data between lots of systems—triggers, retries, approvals, logs, on-prem—then sprinkling LLM/agent steps inside. Massive integration library + mature ops. 👉 If your workflow is 70% plumbing, this wins. Windmill Open source (Mix of Apache-2.0 and AGPL-3.0). Developer-first automation with a code-native core (TypeScript/Python/SQL) and optional UI/workflows on top. LLMs fit naturally as functions or agents, versioned like software. Strong CI/CD, secrets, RBAC, self-hosted story. 👉 Ideal when you want real code + real ops, not a pure node canvas. Langflow Open source (MIT License). Visual LLM graph builder close to Python/LangChain. Great for fast iteration on RAG, agents, custom components. Engineers like it because nodes feel like code, not magic. ⚠️ Weak on ops as graphs grow; strong on experimentation speed. Dify Open source (Apache-2.0). An LLM app platform—not just a canvas. Prompt + workflow + RAG + tools + deployment + team features in one box. 👉 Best for shipping internal or customer-facing AI apps with governance & observability baked in. ⚠️ Trade-off: you buy into their way of doing things. FlowiseAI Open source (Mix of Apache-2.0 and Commercial license). Fastest way to prototype LangChain-style agents on a node canvas. Fantastic for demos, PoCs, internal tools. ⚠️ Gets messy at scale unless you harden or migrate to code. Now under Workday, so expect more enterprise gravity. Sim.ai (YC X25) Open source (Apache-2.0). AI-first workflow builder with a “ship workflows like products” angle. Visual agents, many connectors, heavy enterprise/compliance messaging. 👉 Makes sense if packaging, sharing, and compliance matter early. ⚠️ Maturity and lock-in risks. How I’d choose (cold pragmatism): ·      Mostly integrations + automation (low-code) → n8n ·      Code-first automation with LLMs + serious ops → Windmill ·      Mostly LLM logic + experimentation → Langflow ·      Production AI apps for teams → Dify ·      Fastest demo-to-agent → FlowiseAI ·      Enterprise-postured AI workflows from day one → Sim.ai (YC X25) Main product pain points: ·      n8n: deep AI nuance ·      Langflow/FlowiseAI: ops when graphs sprawl ·      Dify: flexibility vs structure ·      Sim.ai (YC X25): flexibility vs platform assumptions ·      Windmill: requires engineering discipline (but repays it at scale)

16

1 Comment

Apple opens Foundation Models framework to developers, letting them access on-device AI with just 3 lines of code while maintaining privacy. Everyone's racing to the cloud for AI. Apple just bet $3 trillion on the opposite direction. The new Foundation Models framework lets any app access Apple's on-device AI with just 3 lines of code. No cloud costs. No data sharing. No latency. Completely offline capable. Here's the contrarian take: While competitors burn billions on massive data centers, Apple is democratizing intelligence at the edge. They're not trying to build the world's smartest AI. They're making "good enough" AI available to every app, everywhere, for free. This could flip the entire business model. Why pay OpenAI $20/month when your iPhone can already handle 80% of AI tasks locally? Why send sensitive data to the cloud when your device can process it privately? The real disruption isn't better AI - it's cheaper, private, always-available AI that works without WiFi. Apple just turned every iPhone into a personal AI server. While others compete on model capabilities, Apple is competing on accessibility and trust. And in enterprise environments where data privacy matters, that might be the winning move. #EdgeAI #PrivacyFirst #MobileAI #DeveloperTools #AIStrategy Will on-device AI eventually replace cloud-based models for most tasks?

1