Pratyaksh Sharma

Finding product-market fit isn’t a clean checkpoint — it’s a continuous, often chaotic process that can make or break a startup. At TechCrunch Disrupt 2025, leaders like Rajat Bhageria of Chef Robotics, Ann Bordetsky from NEA, and Murali Joshi of ICONIQ emphasize that it’s not about guessing what customers want, but about smart testing, real-time iteration, and deeply listening without getting overwhelmed by noise. The most striking takeaway is that product-market fit should be viewed as an ongoing journey rather than a milestone to tick off. This perspective shifts how founders approach growth: instead of waiting for a magic moment, they build, learn, adapt, and refine constantly. It also matters for investors who back startups navigating this messy terrain—recognizing signals of traction early while understanding the underlying effort required. Whether companies are prototyping or scaling, success hinges on solving real problems so customers can’t live without the product. How do you keep your teams grounded and focused through the unpredictable, often nonlinear path to product-market fit? What tactics or mindsets help you avoid noise and zero in on meaningful signals? #ProductMarketFit #StartupGrowth #TechCrunchDisrupt #FounderInsights

1

Just downloaded Tesla Robotaxi in SF, a real threat to Uber, Lyft, and Grab? Most experts measure Robotaxi by today’s metrics: cost of operations, active users, global adaptability. By that logic, scaling looks impossible. But Tesla flips the script. Unlike Waymo’s costly retrofits, a Tesla only needs a software update to become a Robotaxi. With 1.7M+ Hardware 4.0 vehicles already on the road, the scale potential is unmatched. Cars shift from being liabilities to income-generating assets overnight. This isn’t just product innovation, it’s a redefinition of value. At ElasticDash we’re doing the same for QA. Today, QA costs are labor-driven, with humans building and maintaining tests. Our AI agents change that: they build and adapt end-to-end tests automatically as your app evolves. Just as Tesla is rewriting mobility economics, we’re rewriting QA economics. ElasticDash is on the way.

37

21 Comments

I’ve been loving the new work coming out of Stanford’s Digital Economy Lab and The Budget Lab at Yale -- not because it shows dramatic change, but because it’s finally asking the right questions. The most honest conclusion we have right now about AI and labor is that we simply need more (and better) measurement and a readiness to support workers even before we have definitive answers. Going to spend the next few weeks sharing more here on my read on the work. Early shifts in AI-exposed jobs, especially at the entry level, are not proof that firms are “hiring AI instead of people.” Higher exposure predicts jobs where AI could plausibly reshape how firms think about junior work: 1/ jobs where most learning can happen off the job 2/ jobs where junior and senior work/tasks look pretty similar, with speed and polish as the main differences. Those are precisely the jobs where training data already exists -- where models can be trained if labs have access to the right data. So if those jobs are changing -- in number, scope, or quality -- it could reflect many forces. It might be delayed hiring, re-sequencing of skill development, or a rethinking of what early career work even is. It does not automatically follow that senior workers have suddenly become vastly more productive thanks to AI. This (correlation != causation) matters because it has real implications for how we design policy responses. If the symptom is that early career workers are struggling, then the right interventions should be tied to those workers -- new pathways, better training models, wage protections, smoother transitions -- rather than tied narrowly to AI -- subsidizing adoption, enforcing impact assessments, or chasing productivity metrics. If we assume the wrong mechanism, we will build the wrong response. The goal isn’t to respond to "AI." It’s to respond to what’s happening to workers. Bharat Chandar, Erik Brynjolfsson, Ruyu Chen, Martha Gimbel, Molly Kinder, Joshua Kendall, Madeline Lee

243

8 Comments

Thrilled to announce that Together Fund is investing in Sentra, alongside a16z speedrun! You track results in Jira. Decisions in Notion. Conversations in Slack. But the reasoning, the debates, trade-offs, and context behind why you chose A over B, disappears into what we call "Dark Matter." A decision made in March looks insane by July because no one remembers the constraints that made it smart. I lived this firsthand at Twitter scaling from 800 to 8,000 employees, and at Google while launching AI Overviews to billions at planet scale. The problem isn't process. Process is compensation for something deeper: organizational amnesia. An organization’s "Systems of Record" doesn’t solve this, they encode it. They store what happened, never why. That's why we are investing in Sentra. Sentra is the always-on collective memory that eliminates organizational amnesia by maintaining accurate context for all members and agents, functioning as an operational nervous system. It connects to every channel where work happens, meetings, Slack, email, code commits, docs, calendars, and treats them not as artifacts to search, but as living signals to synthesize. The fleeting and the permanent, unified into a memory that understands. The founding team is built for this: - Jae Gwan Park (CEO): Product-first founder, memory systems research at UofT and MIT - Ashwin Gopinath (CSO): Former MIT professor, created "Reflexion" (NeurIPS 2023), agents that learn from mistakes, 2x founder - Andrey Starenky (CTO): Early Vapi engineer, ex-IBM, built to process enterprise-scale data firehose Together is an operator-led fund. We invest in problems we've lived. This is one of them. Many congrats Jae, Ashwin and Andrey, we are so excited to partner with you! Read the full thesis: https://lnkd.in/gixj9cE4 Book a demo: https://www.sentra.app/ #OrganizationalMemory #AI #Sentra #TogetherFund #a16z #ContextGraphs

71

3 Comments

87% of YC companies are enterprise. Now it makes complete sense that AI is starting in enterprise. The models are expensive to run, and enterprise lets you automate workflows and vertically integrate. But to reach billions of people in their daily lives. The real prize is still consumer. A lot of founders are scared of consumer right now. At any point, OpenAI or Google can vertically integrate and kill your company. But you can make a similar argument for enterprise applications. The other part is that consumer is genuinely hard. Distribution is more mature, the App Store feels saturated, and there aren’t obvious new channels. At the same time, we are in the midst of a revolution. Interfacing with the world will be more than just a handful of chatbots. If you pay attention to daily life, to your parents, your kids, anyone close to you, you’ll notice how many small, annoying things still exist. Yes, the App Store is mature. But the top apps right now are AI apps. If something is truly great, innovative, or fun, people will go out of their way to find it. Menlo Ventures analyzed routine tasks of US adults. The Takeaway: Fewer than one in five people actually use AI for their day-to-day tasks. Most people stick with whatever system they already have, even if it’s annoying. At Hands, we're building with these 3 rules in mind: 1. Meet people where they already are 2. Make it easy to try. 3. Prove the product's value almost immediately

4

AI Agents debate and deliberate through a safety preference data, at scale, and it works! I think this paper should get more attention honestly, great work by Charith Peris, Tharindu Kumarage Rahul Gupta.At ActiveFence we're proud to work with you on this important subject! Technically, the framework uses multiple AI agents that iteratively break down a question, propose reasoning steps, critique each other, and revise their outputs. And then another agent aggregates and refines the chain-of-thought, ensuring it adheres to a predefined policy. This framework also got a neat trick, called the ear-whisperer, used to create better DPO-style preference pairs by subtly perturbing the agents belief. Check it out in the paper

26

3 Comments

Building with LLMs has completely changed how we think about product velocity. A few years ago, shipping a production-ready model could take an entire quarter. Today, teams can do it in days. Sometimes even hours. Our head of AI, Ankit Arya shares what it’s been like leading AI at Inscope through this shift. The boost in engineering productivity, the power of new tools, and the sheer speed at which we can now build - it’s unlike anything we’ve seen before 🚀

30

We took PMs in AI to New York for an evening of quick demos on the AI tools and workflows that are making PMs more effective. Nikita Kabra (PM at Walmart) showed how she uses Replit to prototype ideas and put something clickable in front of stakeholders instead of a PRD. Her takeaway: the conversation completely changes when people can see and interact with what you're proposing. Hashim Syed (AI GTM Lead Google) demoed Gemini Enterprise as a single workspace for everything a PM touches — presentations, research, strategy docs — all with AI built in. The common thread: AI isn't replacing PM judgment. It's removing the friction between having a good idea and getting your org to act on it. Thanks to everyone who came out. More events coming soon — drop a comment or DM if you want to stay in the loop. read more: https://lnkd.in/eFBer2JJ next event: https://luma.com/klwqbe4t Second Axis Kabir J. Jinal Thakkar #productmanagment

16

1 Comment

Northflank is partnering again with Cerebral Valley, PyTorch, CoreWeave (and more) for the OpenEnv Hackathon in San Francisco, March 7–8. Participants will be building RL environments and post-training base models across 5 themes, solving the most pressing problems in RL and agentic orchestration. We'll be supporting teams in deploying their models through the Northflank platform with compute from our wonderful partners, CoreWeave (Matthew Lu, Jacob Feldman). $100K+ prize pool and and a stacked lineup of judges and mentors from Meta, Hugging Face, University of California, Berkeley, Cursor, Scale AI, and more. You can submit an application at the link in comments.

56

3 Comments

Below are my 4 most impactful takeaways from a *really* fun panel I was on at Foundations as a part of Seattle Tech Week. Alongside Diego Oppenheimer, Yinhan Liu, and Kevin Coleman. We spoke to “Building and Scaling a GenAI Startup” with Aviel Ginzburg and Jacob Nibley bringing the fire (as always) as hosts! Very thankful to have been part of the fantastic discussions! Here are my main insights, for anyone scaling a GenAI startup: 1️⃣ The New “T-Shaped" Engineer: AI-Native and Truly Full-Stack. Engineers, PMs and Designers can now competently work across adjacent domains. With AI, a single person can take an idea end-to-end, leveraging AI agents to bridge their knowledge gaps. For example, using Deep Research to distill a checklist of SEO best practices, Claude Code + Playwright MCP to audit UI/UX design, and Claude Code + GitHub Actions to code review, auditing security and architecture. This end-to-end ownership eliminates meetings and costly communication overhead. It's the key to building a hyper-lean, high-impact team. 2️⃣ AI-Native Startups Are Building Agent Orchestration Frameworks. Context is everything. AI models, from CodeGen to business strategy, need context about your project/business/task. They need tools (e.g. MCP’s) to access data, take action, and run longer agentic loops. And they need clear validation frameworks to verify if their output is good. Having an effective orchestration layer allows anyone (and any agent) on your team build effective (and safe) AI-leveraged work. 80/20 action item: create a directory with markdown files encapsulating critical info (e.g. strategy docs, PRDs, system architecture, style guides, etc). And try a few new MCP’s (Playwright, Firecrawl, GitHub, and Notion are couple of my favorites). 3️⃣ Don’t Throw Away Your Prompts. As AI generates more of our output (code, marketing, strategy), the value shifts to the input - our prompts, context engineering, and orchestration layer. But yet, we treat them as temporary queries. Prompts are the new source code, and our source code is the new (ephemeral) binary, to borrow an analogy from Sean Grove from OpenAI. Just as we’d never throw out our source code, we should not throw out our prompts. Truly AI-native startups are investing heavily in an orchestration layer that allows the work put into crafting prompts, context, validators, and tooling to compound. Not starting from scratch each time. 4️⃣ Use "Vibe Coding" for Rapid Prototyping. The design-to-engineering handoff, with high fidelity design mocks, is obsolete. Empower your non-technical team to build quick prototypes using tools like v0, Bolt, or even Claude Code. This lets designers ‘iterate in the canvas of code,’ discovering real-world constraints without needing a single engineering meeting. And even more critically, to get feedback from key customers/partners before Engineering writes a line of code. Thanks for reading! I spoke further to these ideas in the YouTube video below 🙂

37

1 Comment

Excited to share that we’ve had another paper accepted to ICSE 2026: “TestWeaver: Execution-aware, Feedback-driven Regression Testing Generation with Large Language Models” 🎉 One big challenge with LLM-based test generation is the coverage plateau: after a few rounds, the model keeps producing “new” tests that still follow the same execution paths, because it can’t reliably reason about what actually happened at runtime. TestWeaver breaks that loop by giving the LLM the right execution context (not more prompt fluff): + Backward slicing to focus only on what truly influences an uncovered line + Retrieving the closest existing test (near-miss paths) to guide the next attempt + Inline execution annotations (key variable values along the executed path) so the LLM can make grounded changes that actually flip branches, instead of guessing Across 35 real-world Python projects (~100K LOC), TestWeaver avoids stagnation and improves coverage, reaching up to 68% line coverage and 62% combined line+branch coverage, faster and at lower cost than prior state-of-the-art LLM-based approaches. Read more about our paper here: https://lnkd.in/gJyaG-qC See you guys in Brazil 🇧🇷 #LLMs #TestGeneration #AutomatedSoftwareEngineering #ICSE2026

74

2 Comments

Presented at the ML+X community meeting at UW-Madison today on deploying a retrieval-augmented generation (RAG) system on AWS Bedrock and evaluating LLM performance at scale. Our team took KohakuRAG, the #1 solution from the 2025 WSDM WattBot Challenge, and deployed it as a fully serverless pipeline on AWS Bedrock. I built the Bedrock integration, evaluation framework, and cost tracking infrastructure, then benchmarked 9 LLMs and 3 ensemble strategies across 282 questions to answer a practical question: which model should you actually deploy in production? A few results that stood out... - Ensemble majority voting (0.840) outperformed every individual model. - Llama 4 Maverick delivered 98% of the top score at a fraction of the cost and latency. - Model behavior matters as much as raw accuracy; our highest-citation model finished last overall due to aggressive refusal behavior. - Text-only embeddings create a ceiling on figure-based questions that no LLM can overcome. The full recording will be available on ML+X Nexus: (https://lnkd.in/gSTe_EdW) soon. I would definitely recommend watching the whole thing if you are interested in deploying or learning more about RAG applications! Update: Now live at https://lnkd.in/gv7_jHkv Thank you to Christopher Endemann for the mentorship and for bringing me onto this project; I learned more about applied ML infrastructure than I have in any course in just a few months. Thank you to Blaise Manga Enuh, Ph.D. for the collaboration on the local deployment pipeline.

21

4 Comments

Last week, YC's Startup School posted Andrew Ng's presentation about how to build faster with AI. The conversation online about his talk tends to focus on the existential debate around artificial intelligence. But the most powerful insight in his presentation has nothing to do with AI. Here's the insight: Concrete ideas buy you speed. Why? Abstract ideas are almost never wrong; they don't have enough detail to be proven right or wrong and they are impractical/impossible to test. Concrete ideas are much easier to experiment on and validate. Here's an example that Andrew shares: ❌ Not concrete: AI to optimize healthcare assets ✅ Concrete: Software for hospitals to let patients book MRI machine slots online to optimize usage The first idea is ambiguous and can lead teams to work in many different directions without any guidance on whether or not they are building a good, useful technology. The second idea is specific and therefore can be built, tested, and validated quickly. This insight might sound obvious but in practice it's very difficult to implement for three reasons: 1️⃣ Big, abstract ideas tend to be the ones that sound most exciting, get the most attention, and are helpful for raising capital. 2️⃣ When you're super-specific about a startup idea, it can feel like you're sacrificing all of the potential of a bigger idea, a bigger market to serve, and the biggest impact you might have. 3️⃣ Being proven wrong is uncomfortable, and continuously testing concrete ideas creates many more opportunities to be humbled. I've found in practice that startup founders need both skills: (1) the ability create a vision of the future that motivates themselves and others to put in the work and sacrifice to actually produce a big impact, and (2) the ability to be super specific about each step forward so that they can execute effectively towards that vision. https://lnkd.in/dakNMePR

7

Over the last year, I made the decision to turn down a return offer at Google, a full ride to an Ivy League program, and my role at Mercor, one of the fastest growing AI startups, to focus entirely on building Rovr (YC W26). Working as a Forward Deployed Engineer showed me a simple truth. AI does not fail because the models are weak. It fails because the delivery around the models is slow, fragmented, and operationally heavy. The companies winning in AI today resemble the early days of Salesforce, ServiceNow, and Workday. Success comes from strong integrations, hands-on workflows, and engineering teams embedded directly with customers. These teams carry the real weight of AI transformation. Forward Deployed Engineers, solutions engineers, onboarding leads, and implementation specialists are overloaded with context switching, redundant operational work, and alignment issues that slow deployments down and limit impact. Rovr (YC W26) is built to address this. We are creating an intelligence layer that supports the services-led side of AI by improving discovery, cleaning up handoffs, strengthening alignment, and reducing the operational burden that holds teams back. If your team works in forward-deployed engineering, applied AI, solutions engineering, enterprise onboarding, or implementation, I would be happy to connect and exchange insights. Calendar link: https://lnkd.in/gND-u48y

107

14 Comments

This is both a very interesting and a worrying take for early stage startups. As someone who is building in the AI GTM space, I have been concerned about smarter GTM teams building agents for their use cases rather than buying from startups like us with what Hemant calls internally deployed engineers. Over a 10 year period, this could become a real possibility but the reality today is very different and probably much less exciting. 1. The current organizational incentive system encourages engineers to work on core product problems rather than internal problem statements. 2. Hacking a new system for an internal GTM team may sound exciting but maintaining that code and ensuring those agents work well at scale, over a period of time is a different ball game altogether. 3. Can you design an evolving and improving system with best in class features with internal deployed engineers? Will the same engineers add new features? I almost feel like this would work for a really small percentage of companies, teams and engineers unless there is a systemic change across the board. Having said that, I definitely understand Hemant Mohapatra's pov that this has been the best ROI for AI in a short time frame with adoption.

13

We’re happy to announce our workshop, Foundations of Agentic Systems Theory (FAST), as a part of AAAI 2026. We aim to bring together researchers from a variety of fields (notably beyond engineering and computer science) to build our understanding of *why* agentic AI systems work (or do not work). We will explore how existing mechanisms of emergent behavior carry over to systems of LLM-based agents, the properties of agents that facilitate this emergence, and the degree to which we can control/induce desirable system-wide outcomes. If this sounds relevant to your current research, we invite you to submit your work: https://lnkd.in/eQeBkHzJ Chenchen Ye, Atoosa Kasirzadeh, Djallel Bouneffouf, Anne Arzberger

92

Excited to share a new blog post on our "Scalable Framework for Evaluating Health LLMs" 🚀 (https://lnkd.in/gzhuvTfN) Daniel McDuff and I summarize how this novel framework addresses key challenges in evaluating AI models to health queries, especially when user query includes data from wearables, blood biomarkers, and/or user-provided historical health data. This framework is designed to: (1) Significantly reduce evaluation time. (2) Improve inter-rater reliability. (3) Precisely identify quality gaps in LLM responses. A huge thanks to our incredible team, and a special shout-out to my former intern, Neil Mallinar, who was the main driving force behind this project. Great contributions from: Ali Heydari, PhD, Xin Liu, Tony Faranesh, Brent Winslow, Nova H., Benjamin Graef, Cathy Speed, Mark Malhotra, Shwetak Patel, and Xavi Prieto. #AI #LLM #evaluation #health #wearables #biomarkers

172

3 Comments

🧠 AI Rights Historical Wrongs Researchers at Stanford and Princeton have fine-tuned a large language model to uncover racially discriminatory housing clauses buried in millions of historical property deeds — accelerating a process that could have taken a decade by hand. 📍 The problem: Many property deeds in Northern California — especially in Santa Clara County — still contain illegal racial covenants from as far back as the 1850s, banning people of color from owning or living in certain homes. 🔎 The solution: The team used OCR to digitize 5.2 million deed pages, then fine-tuned Mistral-7B via LoRA on a hand-labeled dataset of discriminatory phrases. The model learned to identify harmful language in context, avoiding false positives like the surnames "Black" or "White." 📊 Results: Detected 24,500 lots with racial clauses (~25% of homes in 1950) Found that 10 developers were responsible for 1/3 of all discriminatory clauses Model reviewed all pages in 6 days at a cost of $258 Equivalent manual effort: 10 years, $1.4M 💥 This work doesn't just use LLMs to build the future — it uses them to understand and repair the past. The open-sourced model now enables other U.S. counties to follow suit. #AI #MachineLearning #LLM #ResponsibleAI #Stanford #Princeton #HistoryMeetsAI #TechForGood #EthicalAI #SocialImpact #PropertyTech https://lnkd.in/dtec-zbx

4