Pat B.

Appreciate Rackhouse for the thoughtful spotlight. I started TorqueAGI after a realization that kept repeating across labs and real deployments: robots don’t fail because they lack intelligence in controlled demo settings. They fail because small, unavoidable distribution shifts across perception, reasoning, and action compound over time. Even minor changes in real operating conditions quickly expose how brittle most robotics stacks really are. Over the last few years, foundation models transformed digital AI by unifying representation and generalization. Physical AI represents the analogous inflection point for robotics. In physical systems, intelligence must internalize geometry, dynamics, uncertainty, and temporal context as first-class signals, not edge cases handled downstream. My focus with TorqueAGI has been to build a shared intelligence layer that connects sensing to action through physical reasoning, runs at the edge, and learns efficiently from real interaction rather than massive curated datasets. This is what allows systems to stay online as conditions evolve, instead of relying on constant resets, retraining, and patches. Grateful to Rackhouse for backing this direction early and for recognizing that scaling robotics in the real world requires rethinking the intelligence layer itself, not just optimizing existing pipelines. Kevin Novak Kofi Asante

8

We are excited to launch LoftxAI "Revenue Management AI" agent - part of suite of AI agents we are building to assist with rental back office operations. Learn more below how our customer Preston Walls is using it to optimize revenue and improve productivity https://lnkd.in/gYtxgUvH #realestate #realestateinvesting #rentcontrol #VC #proptech #rentmanager #appfolio #multifamilyinvesting

27

1 Comment

Building and Scaling AI Products in today's world is no easy job. You must serve hot masala dosas today while building the industrial kitchen that will feed millions tomorrow. Your product teams (pods) are the snack cart. agile, fast, and focused on immediate value. However the organization needs industrial Udupi kitchen: Built for quality, consistency, and scale. Both are essential. Leaning too far either way breaks the model. The Risks of only running Snack Carts : If we only operate snack carts, we feed customers immediately but breed chaos. - Multiple lentil suppliers : Fragmented model hosting and duplicated vector databases. - Inconsistent taste and quality : model drift, lack of evaluation discipline, and security gaps. - Secret batter recipes : Fine-tuning done in silos with no shared checkpoints or governance. - Sky-high spice costs : Uncontrolled API usage and unmanaged cloud spend. Speed without structure burns out teams and budgets alike. The Risks of only building Industrial Kitchens: If we only focus on the industrial kitchen, we’ll have the most sophisticated MLOps facility in town, serving no one. A perfect central kitchen gives you control but kills creativity. You trade speed for bureaucracy and risk building a scalable, empty restaurant while others feed the market. We sacrifice the very things that make AI products win "Perfection kills momentum. Over-engineering stalls learning". A Phased, Balanced Menu: Launch the Snack Cart (Next 3 to 6 Months) - Goal: Learn what dishes customers truly crave. Prioritise speed and customer validation. - Action: Use ready made batters: Cloud APIs, pre trained models, RAG. - Watch out: Do not mistake early excitement for product market fit. Build the Industrial Kitchen (Months 4 to 9) - Goal: Scale efficiently without losing flavour. Introduce versioning, observability, and governance while keeping innovation alive. - Action: Standardise core MLOps and data pipelines. - Watch out: Avoid process overreach that chokes experimentation. Scale the Udupi Chain (Month 10 and Beyond) - Goal: Deliver a complete, scalable AI experience. - Action: Integrate proven wins into end to end workflows with fine tuned models. - Watch out: Preserve agility even as systems mature. Guard fiercely against the bureaucracy and power struggles that often accompany scale. The Leadership Tension: We’re serving crispy dosas at the cart today while designing the kitchen for tomorrow. This friction isn’t dysfunction. It’s the rhythm of every modern AI company choosing to evolve rather than stagnate and every leader today faces this trade-off. Are you feeding customers today, or just stocking the pantry for tomorrow? #AI #TechLeadership #MachineLearning #MLOps #ProductManagement #ScalingAI #AIProducts #AIGovernance #EngineeringLeadership

20

NVIDIA just cracked turn-taking in Voice AI. Meet PersonaPlex-7B — an open-source, full-duplex conversational model. Most voice agents are half-duplex (talk → listen → talk). PersonaPlex listens and speaks at the same time, enabling real interruptions and backchanneling (“uh-huh”, “right”). Why it matters • Dual-stream transformer operating directly on audio tokens (no separate ASR/TTS) • Parallel prediction of text + audio for ultra-low latency • Zero-shot persona control: swap personality (text) and voice (audio) at runtime • Built on Moshi, trained on 2,250+ hours of conversational data For real-time support bots and roleplay agents, the ASR → LLM → TTS cascade is obsolete. #VoiceAI #ConversationalAI #GenerativeAI #NVIDIA #OpenSource #AIResearch #LLMs #RealtimeAI

41

1 Comment

We had an interesting session with our FasterDocs NED yesterday. When you’re this close to something, living and breathing it every day, you can miss what’s obvious. It forces you to go back to your why. The reason FasterDocs exists is actually very simple. Ralph Hardwick spent 20+ years in agencies and kept seeing the same thing, margin being lost at the scoping stage. Our NED saw exactly the same when building, scaling and exiting his agency. And I’ve seen it from the other side, recruiting into client services teams. No one joins an agency to be stuck in admin. But when scopes aren’t properly defined, things get missed. When change requests aren’t tracked and billed, work gets absorbed. When amends go through without sign-off, someone pays for it. And it’s always your bottom line. Every agency knows this happens. The question is, what are you doing about it? PM tools manage delivery, but they don’t protect margin. By the time work reaches delivery, the damage is already done. That’s why we built FasterDocs. To fix the commercial layer before the work even starts, and give client services teams more time to focus on work that is billable and work they actually enjoy. We’re starting to use it on live campaigns with agencies now. If you want to see how it works, drop me a message. #agencylife #aiautomation

11

2 Comments

The latest episode of "Where AI Works" just aired. I was honored to discuss AI, technology, and business in an interview by Serguei Netessine, Sr. Vice Dean of Innovation at The Wharton School. We covered topics including strategies for new AI business models, shifts to AI platforms, and the cultural divide between AI adopters and skeptics. You can find the full podcast here: https://lnkd.in/gCjQnRhU #AI #Strategy #Businessmodels #innovation #Wharton  

30

4 Comments

We fine-tuned a 2B vision-language model to parse bank statements at production scale. The full write-up from Shubham Tiwari is on our engineering blog. Fine-tuning is no longer the hard part. You have the GPU, the dataset, the model. The training script is a few hundred lines. The hard part is everything that happens after. The post covers how we worked through GPU bottlenecks when bank statements pushed output tokens past our context window and resource constraints, tuned the VLM-specific parameters that defaults get wrong, fine-tuned training prompts to fix specific field regressions, and a few other things that mattered more than hyperparameters. Why does this matter? Frontier APIs handle document AI well, until you're processing tens of thousands of pages a day. At that point you're paying thousands a month, inheriting rate limits, and can't fix specific failure modes. A purpose-built 2B model on a single GPU changes the economics. Really proud of the work Shubham and the team put into this. https://lnkd.in/g3qkuQvd

118

5 Comments

Interesting point from Larry Fink - Can AI help increase GDP with a diminishing population? This reminded me of the questions posed by Prof Jhunjhunwala at #GlobalAIConclave Also a great point made by Shri Ambani ji: - There are limited opportunities for scale around the globe where 100s of billions of dollars can be invested meaningfully. This is where India presents multiple opportunities at this scale #GlobalDialogues Larry Fink Mukesh Ambani

2

🚀 Incredible Fireside Chat with Jensen Huang & Aravind Srinivas Today, we had the privilege of participating an extraordinary conversation between Jensen Huang (Founder, President & CEO of NVIDIA) and Aravind Srinivas (CEO of Perplexity) as part of our Perplexity Business Fellowship program. 🔥 Key Highlights: 1️⃣ On Pivoting & Resilience: - Jensen shared NVIDIA's near-death experience with NV1/NV2, when they had to completely pivot from curved surfaces to triangles, and forward texture mapping to inverse mapping 💡 "Just because we don't know how to compete doesn't justify doing it the wrong way" - a lesson in confronting hard truths 2️⃣ On Leadership Philosophy: - "I don't have any goals. The journey IS the goal" - Jensen's refreshing take on focusing on daily excellence over numerical targets - NVIDIA doesn't have annual plans or tape-out deadlines - they focus on building every chip perfectly 💡 "I have plenty of time" - inspired by a Japanese gardener's mindful approach to work 3️⃣ On Learning & AI: - Pre-AI: Jensen would buy textbooks at Fry's to learn 3D graphics - Post-AI: "This is way better!" - now uses Perplexity to read ArXiv papers 💡 "Productivity isn't about answers anymore, it's about questions" - the art of context engineering 4️⃣ On Physical AI: - Robots in warehouses/factories happening "approximately now" - Multimodal reasoning models can now be told "how would you drive?" and reason through it 💡 The next 2-3 years will focus on massive simulation and digital twins 5️⃣ On Inspiration: - Roger the dishwasher at Denny's who showed extraordinary pride in his work - His mother who taught him English without knowing English herself 💡 "Learn from everybody" - finding lessons in unexpected places 6️⃣ On Competition & Collaboration: - NVIDIA collaborates with competitors, shares roadmaps openly - Focus on finding your unique position rather than beating others 💡 "We're inventing new computing platforms that don't exist today" 🌟 The Bottom Line: This conversation beautifully illustrated how technical excellence, authentic leadership, and human wisdom intersect in building transformative companies. What resonated most with you? The pivot stories, the leadership philosophy, or the vision for physical AI? #PerplexityFellowship #Leadership #AI #NVIDIA #Entrepreneurship #TechLeadership

49

1 Comment

I recently interviewed Harshit Omar, Co-Founder & CTO of FluidCloud, and asked him three direct questions about what recent cloud outages really mean for enterprises. His answers cut through the noise — on:
• Why “multi-cloud” often isn’t truly resilient
• Where most enterprises are still exposed to risk
• What real cloud portability actually looks like in practice Great read!

34

1 Comment

Had a great conversation with Ravit Jain about what’s happening in enterprise AI.   Instead of getting caught up in benchmark-driven narratives around #superintelligence, we stayed focused on addressing real challenges like the 95% failure rate for AI pilots (https://lnkd.in/gSsD_K_W), and the very concerning 13% decline in early career employment for U.S. workers in roles most exposed to AI (https://lnkd.in/gpDmav5D). Delivering on the promise of AI will need a focus on operational success and workforce adaptation — not just headline-grabbing demos

62

5 Comments

At our #AgenticAI Learning Event in San Jose, we were fortunate to have Amit Singhal — Founder of Sitare University and former SVP of Search at Google — share a perspective that reframed how we should think about AI. His core thesis: every technology wave democratized something. Mainframes democratized compute. The internet democratized connectivity. Mobile democratized attention. Cloud democratized scale. AI will democratize judgment — the ability to interpret information, make decisions, and take action under constraints. But here’s where it gets important for enterprises: judgment at scale is only valuable if it’s consistent and responsible. Consumers want experiences. Enterprises need outcomes — efficiency, control, risk reduction, decision advantage. Horizontal AI tools fail enterprises precisely because they have no ownership model, no liability assignment, no domain accountability. No trust. Amit made clear why this matters deeply for trade compliance. At KlearNow Corp , AI isn’t just about automation — it’s about human-level consistency and accountability at scale, where every agent decision is traceable to data, rules, and a named human escalation path. In a trade environment as complex and legally consequential as today’s, that’s not a nice-to-have. It’s the cornerstone. As Amit closed: we’re living through an industrial revolution-level shift. Hire the best machine makers — and demand accountability at scale. Thank you, Amit, for bringing this clarity to our CAB event. #AgenticAI #TradeCompliance #KlearNow #EnterpriseAI #Accountability

20

3 Comments

This was a great panel. From Harness Engineering to System-3 thinking for coding agents, to ASI for specific areas arriving before AGI, we covered a lot of ground.

16

Insights from StateStreet's EVP & CIO, Kartikesh Herur, during the "Autonomous AI Agents in Software" panel discussion at the Imagination in Action Summit at Stanford University. "...you have to start thinking about that evolution of that agent, because now it's building based on the learnings that it has...but over time it's going to learn and it's going to understand the different factors coming into play that don't exist in today's world."

18

1 Comment

Been listening to this amazing conversation with Andrej Karpathy on Dwarkesh Patel podcast: Andrej Karpathy on The Cognitive Revolution: https://lnkd.in/gQ-kzGCt It’s a long one, but totally worth it — Andrej is such a fantastic educator. I’ve only made it through the first 1.5 hours, but there are already so many gems: 🧠 On Reinforcement Learning: RL sucks — it’s like "sucking supervision through a straw.” But it's the best we have. Such a good analogy and evocative imagery. 💡 On the future of models: The best future models will learn to extract the cognitive core of intelligence — separating reasoning and generalization from the rote memorization that defines current LLMs. You can already see glimpses of that shift as smaller models begin outperforming larger ones. 🤖 On Agentic AI: He compares agentic AI to compilers — tools that make humans better programmers of thought. Will we get to AGI? No definitive answer, but I like the framing. I'm still only 70% through the episode, but it’s packed with insights like these. I’ll share more takeaways as I go — this one’s definitely worth a listen.

7

1 Comment

❤️ Loved this from Harshit Omar Cloud outages aren’t anomalies; they’re warnings. Most enterprises are architected as if hyperscalers are “too big to fail,” but recent outages have proven the opposite. Resilience needs a new baseline: Can you run somewhere else when your provider falters? That’s the promise of #CloudCloning and true #multicloud optionality. Big thanks to MR Rangaswami Sir for the conversation. Worth the read.

30

1 Comment