Christine Lam

🧠 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

We all hate opening a full fridge and feeling like there's "nothing to eat," so I tried to fix that with a little Python. 🐍🥘 Meet Patatri's Kitchen—my personal AI assistant that tracks my pantry and plans meals based on what I actually have in stock. ✨ I used this project as a playground to dive deep into Agentic AI and Google’s Agent Development Kit (ADK). 🧠🤖 Instead of a simple chatbot, I built an intelligent "Kitchen Orchestrator" powered by Gemini 2.5 Flash. It autonomously handles the logic to: ✅ Analyze live inventory 📊 ✅ Prioritize existing ingredients to reduce waste 📉 ✅ Auto-generate shopping lists for missing items 🛒 ✅ Auto-purchase required items for cooking 🛍️ Just a fun build to experiment with orchestrating AI agents! 🛠️🚀 Check out the demo. 👇 #Python #GenerativeAI #GoogleGemini #AgenticAI #MachineLearning #GoogleADK #Coding #SideProject

22

2 Comments

Last week, I attended SVGenAI’s “Halloween Horror Stories” tech meetup in the South Bay, an evening that mixed real-life lessons, community spirit and some spooky tales from the world of AI! Huge thanks to the organizers, Shermie Miller-Zhang and Sujee Maniyam, for making the event so engaging. Learned a lot from fantastic speakers: Saul L. (AI agent protocols), Kiran Karunakaran (manufacturing & the butterfly effect), Dennis Yap (AI fact-checking) and Vishakha Gupta (agentic memory & data schemas). From protocols for AI communication and industrial automation mishaps to the challenge of fact-checking machine generated text and solving data chaos, every speaker shared practical takeaways and reminded us why reliable engineering is key for AI success. Top lessons: - Tiny data errors can trigger major failures, especially in manufacturing. - AI agents need standardized, reliable protocols to communicate (think AMTP for multi-agent systems). - Building trustworthy AI means grounding models in verifiable facts, not just scaling up data. - Robust data management is vital to make AI pipelines work in the real world. It was also inspiring to meet new friends like Peter Corless: our chat about LLM perplexity was a highlight. I got to network with other amazing folks as well. If you’d like to dive deeper into the stories, learnings and practical tips from the speakers, check out my detailed recap in the blog! (Link attached below) https://lnkd.in/eU6NhzSH #AI #TechEvent #GenAI #AgenticAI #SouthBay #TechCommunity #SVGenAI

48

12 Comments

Recently I gave a talk at SF Ruby about a library we built (and recently open-sourced!) at AngelList called Zaxcel. Why'd we build it? Excel is the lingua franca of finance, but most programmatic Excel generation libraries force you to think in coordinates and formula strings. We took a different approach: model the workbook as a directed graph where data flows through named objects. Let the library handle coordinate resolution and formula composition. The result is a DSL that actually makes sense: you write formulas as Ruby expressions, reference things by name, and the type checker catches mistakes at compile time. It's revolutionized how we build financial statements. What used to take weeks now takes hours (and with AI agents - only minutes!), empowering thousands of workbook deliveries a quarter. Read how we went from nightmare refactors to 100x productivity: https://lnkd.in/es2aTttR Open-sourced repo (you can use it today!): https://lnkd.in/ecKRwcaW

91

8 Comments

I recently stumbled across cocoindex-code on GitHub(https://lnkd.in/e3nJUWAq), and my reaction was… complicated. Part excitement at the “great minds think alike” moment, part quiet acknowledgment of my own bandwidth limits as a solo dev. Those who’ve followed my work know I’ve been tinkering with my own codebase-rag project last year. The core idea maps almost exactly onto what cocoindex-code does: AST-based semantic indexing, a local-first embedding design, precise retrieval to trim the context an Agent needs to consume. Seeing it executed at this level — Rust-powered engine, clean MCP integration, solid incremental indexing — my first thought was: the direction was right. My second thought was: one person really does have finite hours in a day. Over the past several months, I’ve been convinced of something that I think doesn’t get discussed enough in AI engineering circles: Agents have two hard physical constraints — context window capacity and attention bandwidth. When we ask a general-purpose Agent to reason over a 100,000-line codebase, we’re fighting both at once. My codebase-rag experiment was an attempt to work around those constraints rather than through them. cocoindex-code took the same thesis and productionized it with harder tooling. That gap got me thinking about a few broader patterns: The insight arrives before the implementation. In the LLM era, identifying the right architectural direction — like cognitive offloading — is the entry ticket. If you notice that the problem you’re independently working on is also being pursued by YC-backed teams or prolific open-source contributors, that’s a signal you’re standing on the right ground, not a reason to stop. Implementation quality determines the ceiling. I used scripts to validate the concept. The best projects are using Rust to solve production-grade performance anxiety. There’s a meaningful gap between “this works” and “this scales” — and closing that gap is where most of the craft lives. The solo developer’s real constraint isn’t ideas. It’s the constant negotiation between validating a direction and polishing an execution with finite time. Sometimes seeing a more complete implementation of your own idea isn’t discouraging — it’s a kind of release. It confirms the direction was sound, and it frees you to focus your attention on the next meaningful “how.” The Agent paradigm is shifting from instruction mode to service mode. The principle I keep coming back to: “The best context is no context.” Good AI-native applications shouldn’t make Agents heavier — they should make them lighter and more precise through well-designed external state and domain infrastructure. Respect to the open-source teams pushing this forward. And onward for the rest of us working on the architecture modernization side. #AI #AgentNative #SoftwareArchitecture #RAG #EmbeddedSystems #EngineeringMindset #OpenSource

4

Last night, I had the opportunity to return to UCLA to speak in the Philosophy Department’s Being in the World series. It was a full-circle moment. These were the same halls where I first began studying computer science and imagining a career pivot into engineering. We discussed career paths, job markets, AI’s impact on education, and how students can break into tech. But one question stood out: How do you think an undergraduate degree in philosophy has benefitted you both generally and in your field? I’ve thought about that question for years. Generally, philosophy has given me a disciplined way of orienting myself in the world. We are living through profound shifts:  artificial intelligence, ecological destabilization, geopolitical realignment, social fragmentation. Philosophy trains you to examine systems in their totality rather than reducing them to a single domain like economics, politics, or engineering. It provides a scaffolding for moving between disciplines without losing coherence - a space of expansive thought that transcends any particular topic. In my field, philosophy has been equally practical. Software engineering is not fundamentally about writing code. It is about navigating ambiguity, clarifying assumptions, defining terms, and reconciling human intention with technical constraint. At higher levels, engineering becomes the design of logical systems in response to shifting organizational and cultural pressures. The tension between vision and implementation mirrors much older philosophical tensions:  perception and reality, reason and passion, abstraction and instantiation. The questions are ancient. The tooling is new. It has been years since I’ve written on Descartes or Russell, but I remain a philosopher in spirit. Grateful to UCLA Department of Philosophy for the invitation, and especially grateful to the students asking serious questions about their futures.

11

$100k can be low income. Even for a single-person household. Of course this isn’t referring to a small town in the countryside. According to the California Department of Housing and Community Development’s 2025 income levels (linked below): In Marin, San Francisco, and San Mateo counties <$109700 is low income for single-person households. In Santa Cruz county <$111100. And in Santa Clara county <$111700. It’s pretty well known that Silicon Valley and surrounding areas are fairly expensive. And for many people the salaries in tech make up for and reflect that. Still, when the median income in Santa Cruz county (e.g. $92950 for a single-person household) is lower than the low income threshold… It’s pretty astonishing to think about. Accounting for income, is the SF bay area the most costly city in the world for the average resident?

32

7 Comments

Really excited to share our journey participating in the AgentX competition by Berkeley RDI over the past two months! 🎉 We've been tackling some of the biggest challenges facing AI assistants today, and it's been an intense but rewarding experience. Like many of you, we've experienced the frustration of AI assistants that just can't seem to handle complex or novel requests. They often hit you with the dreaded "I can't help you with that" because they're limited by rigid programming and a lack of adaptability. (Did you know a recent study found that nearly 76% of chatbot users report frustration with existing solutions?!) And then there's the cost! Existing AI systems often rely on expensive models, even for simple tasks, leading to inefficient resource management and limiting accessibility. It's like using a sledgehammer to crack a nut! That's why we set out to build something different: an AI agent that's both intelligent and efficient. Here's a glimpse into our approach and what we built for the AgentX competition: 1. Intelligent Task Decomposition: We focused on breaking down complex tasks into smaller, more manageable sub-problems, allowing for more targeted solutions. 2. Adaptive Strategy: The agent dynamically adjusts its strategy based on the user's goals, the complexity of the task, and the available budget. 3. Model Orchestration: We experimented with using a mix of models, including smaller, more specialized models like DeepSeek-R1, Llama 3.2, Mistral, Gemini, etc. for specific tasks. This allows us to optimize for both performance and cost. 4. Autonomous Tool Use: We integrated tools like Google Search to enable the agent to gather real-time information and stay grounded. The agent can also create tools it needs on-the-fly. 5. Budget-Aware Design: Cost was a central consideration throughout the entire development process. We wanted to prove that you don't need to break the bank to get powerful AI assistance. One of the biggest challenges was finding the right balance between performance and cost. We spent a lot of time experimenting with different model configurations and optimization techniques. It was a fascinating learning experience! We're now opening up a waitlist for early access to our Pro version and would love to get your feedback. If you're also passionate about building efficient AI agents or have experience with agentic AI, we'd love to hear from you! 👇 🎯 Waitlist Link: https://lnkd.in/ghxnadtS A huge thank you to Hugging Face, Lambda and Groq for sponsoring this project, and my partners Parth Shah, Harshil Patel and Saisha Shetty, couldn’t have done it without you! What are some of the biggest frustrations you've experienced with current AI assistants? Share your thoughts in the comments! 👇 #LLM #AutonomousAgents #AI #MachineLearning #OpenSourceAI #RAG #Mistral #Gemini #Llama #DeepSeek #AIInfra #AgentX #Competition #ProjectUpdate #AIEfficiency

33