Bruno Wu

Generic large language models fail at reasoning with operations data. Not marginally. Fundamentally. The failure is architectural. Tokenization destroys temporal signal in time-series metrics and flattens topology into text, stripping away the relationships that give operational data meaning. The model sees numbers and words, not systems. HelixGPT was built to solve this. Its Mixture of Experts architecture uses specialized experts for metrics, logs, topology, and change, preserving causality across observability and ITSM. The performance gap is not subtle. On standardized reasoning benchmarks, HelixGPT experts achieve RougeL scores above 0.85, compared to roughly 0.20 for vanilla LLMs, with an order of magnitude fewer parameters. That difference separates plausible output from operationally useful, actionable intelligence.

147

7 Comments

🚀 Rethinking Agentic RL for LLMs: Driving Real-World Autonomy As a GenAI AI Architect, I see the industry moving rapidly from static, single-shot AI models to agents capable of multi-turn, autonomous problem solving. But what does it *actually* take to turn large language models (LLMs) into robust, multi-turn agents in open-ended environments? 📄 "A Practitioner's Guide to Multi-Turn Agentic Reinforcement Learning" offers actionable clarity. The authors dissect the chaos of current approaches and propose a practical framework grounded in real experimentation, not just theory. ✨ At its core, their recipe consists of: - 🏗️ Environment: Building up agent skills through progressively complex scenarios (think: spatial, object, and solution scale). Training on easier tasks still yields transferable skills—an insight that should shape our context engineering strategies. - 🎯 Reward: Structuring feedback matters. Dense milestone rewards accelerate training, but only if well-aligned with your RL algorithm. Too sparse, and learning stagnates; too frequent, and signals can get noisy. - 🧩 Policy: Smart initialization with demonstration-based priors (SFT) paired with online RL makes agents sample-efficient and more generalizable. The headline? It's the multi-turn design—not just algorithm choice—that really drives success. 🧪 Their benchmarks across TextWorld, ALFWorld, and SWE-Gym show LLMs growing beyond scripted responses toward true agentic behavior—adapting, generalizing, and operating in the wild. 🔍 For GenAI leaders, the biggest friction isn't always about model size—it's about crafting environments, rewards, and policies that unlock real autonomy. This work brings much-needed rigor and recipe-driven guidance, cutting through the hype with insights we can act on. 🤔 Which pillar do you find most challenging in agentic AI development: environment complexity, reward structure, or policy optimization? #AIResearch #GenAI #LLM #MachineLearning #DeepLearning #AIEngineering #ReinforcementLearning #AgenticAI

8

Inference at scale is evolving. The future is composable systems where offline batch, online batch (near-realtime), and realtime traffic can all be served from a single, unified inference endpoint for most use-cases. This simplifies your architecture, reduces operational overhead, and increases throughput. Our new tutorial series will show you how to build it step by sep. In the first installment, Erik Saarenvirta walks you through creating a scalable image classification system on GKE that can be adapted to a variety of use-cases. Leave us feedback in comments or ask questions we are happy to answer. Kent Hua Ishmeet Mehta Erik Saarenvirta https://lnkd.in/dVNiUhwE

16

History is repeating itself. 10 years ago the recsys industry was migrating from linear model recsys to sparse neural network recsys. Back in 2016, many people in the field were skeptical that deep learning would deliver positive ROI for recsys, the fear being rooted in the perceived higher training and inference cost. I see people repeat the same fear, uncertainty, and doubt about this year's migration from sparse neural network recsys to LLM-native recsys. Those who do not lean into the LLM-recsys migration this year will fall behind.

78

1 Comment

💰 Built an End-to-End Cash Flow Forecasting & Liquidity Risk ML System Over the past few weeks, I designed and implemented a production-style analytics system to help finance teams proactively manage liquidity risk. 🔍 Problem: Finance teams often react to cash shortfalls instead of predicting them early. 🎯 Solution: I built an end-to-end machine learning pipeline that: • Forecasts monthly cash flow using time-series modeling • Classifies liquidity risk (High vs Low) using logistic regression • Generates business KPIs (avg cash flow, peak months, risk periods) • Visualizes insights through an interactive Streamlit dashboard • Is version-controlled and published on GitHub 🛠 Tech Stack Python | Pandas | Scikit-learn | Time Series Modeling Logistic Regression | Matplotlib Streamlit (Dashboard) Git & GitHub 📊 What the System Delivers ✅ Forecasted future liquidity trends ✅ Identified high-risk periods before shortfall ✅ Created a decision-support dashboard for finance leaders ✅ Built modular pipeline structure (data → features → ML → deployment) This project strengthened my understanding of: • Financial data modeling • ML pipeline structuring • Business-focused analytics • Deploying data products, not just models 🔗 GitHub Repository: https://lnkd.in/eXYe2egZ Always open to feedback and conversations around Data Science, Financial Analytics, and ML-driven decision systems. #DataScience #MachineLearning #FinancialAnalytics #Python #Streamlit #OpenToWork

14

4 years at Uber today. My most valuable lesson: grow people’s Data Science and TPM skills to supercharge collabs. This is how I did it 👇 HOW: - Composed a 100+ page document called “Data Science For Everyone”, which is equipped with step-by-step tutorials, videos and notes that accelerate and orientate practical skill development for real problems - Maintain the Guide current with regular updates, engaging visualizations (where possible) and notes that complement Source of Truth documentation elsewhere -Circulated passively and engaged comments/suggestions inline and via slack, encouraging others to share if they found it valuable - Made myself available to chat, zoom and present IMPACT: - 1.2k+ users across the business - 50+ teams, involving professionals such as: MLE, SWE, Data Scientists, Data Engineers and TPMs/PMs across Safety, Risk, Legal, Strategy & Ops and many others - 400+ distinct users per month for the last few months PURPOSE: - Grass-root tool to onboard new hires - Teaches how to use GenAI tools - Architectural designs for experimenting with simple LLM use to multi-agentic RAG workflows - Generic code bases for people to get started - Standard project directory designs that illustrates best practices for scalable and efficient growth - Guidance and code snippets for how to use Python for big data projects, including how to build automation and classic ML solutions CIRCULATION: - Organic - Met a need shared across professionals across all levels. NEXT STEPS: - I hope to collaborate with others to compose a Data Science course curriculum for K-12 students - In the meantime, I’ll be working on general educational resources A high tide rises all boats and builds a meaningful network of collaborative ships!

80

10 Comments

A new benchmark dropped this week (AMA-Bench, Zhao et al.) showing that most agent memory systems (RAG, GraphRAG, summarization, embedding retrieval) underperform just cramming the full trajectory into a long context window. The benchmark tests memory over agent trajectories: tool calls, API responses, state changes, not chat history. The distinction here is important because agent traces are causal chains, not documents. Step n explains step n+10, and cosine similarity doesn't capture that relationship. Summarization can entirely lose that. And errors can compound, because every irrelevant memory retrieval pollutes the LLM's context. The paper proposes a causality graph approach that helps by preserving explicit causal links between actions and outcomes instead of lossy compression. I think it's an interesting direction that's emerging. Agent memory isn't a retrieval problem, it's a causal reasoning problem. Full visual breakdown in comments. #AgenticMemory #LLMs #Evaluation #AMABench

71

6 Comments

There’s one idea I keep finding useful, whether working on ML problems or making everyday decisions: second-order thinking. First-order thinking asks: “What happens next?” Second-order thinking asks: “And then what?” As the Farnam Street blog puts it, it’s about understanding “the consequences of the consequences”—looking beyond the immediate outcome to how a decision plays out over time. It’s the difference between optimizing a metric and understanding the system behind it. In ML, a first-order decision might be: “ This model is directly optimized for revenue—ship it.  ” Second-order thinking asks: “How does this change user trust, content diversity, content bias, or long-term retention?” A real-world parallel is quick-commerce platforms that optimized aggressively for short-term conversion and revenue using so-called “dark patterns.” While these tactics delivered immediate gains, they triggered user backlash and regulatory scrutiny, ultimately undermining trust.  In organizations, first-order thinking pushes for speed. Second-order thinking considers culture, talent compounding, and the behaviors we unintentionally reward. Second-order thinking is harder because: - It’s less obvious - It’s harder to measure - It often conflicts with short-term wins But as highlighted in Farnam Street’s writing on mental models, this is exactly where long-term, durable advantages are built. When designing systems or products or sometimes in everyday work, it helps to pause and ask one extra question: “If this works, what happens next—and after that?”

62

5 Comments

Bias in Causal Inference | Framing the problem When you can’t run an A/B test and rely on observational data to understand the impact of initiatives (e.g., promotions or new feature introductions), careful problem framing and awareness of bias are essential. Two common biases in causal inference are selection bias and confounding bias. Here’s a summary on the former, selection bias: Selection Bias | Conditioning on an Effect (Collider Bias) Selection bias from conditioning on an effect arises when we restrict the analysis to a variable influenced by treatment. For example, measuring a promotion’s impact on cancellations only among purchasers introduces bias. The promotion increases purchases, and purchase intent affects both purchases and cancellations, leading to misstated impact estimates. Problem framing depends on the outcome to influence. For example, to reduce overall churn, measure the total effect across all exposed users, including non-purchasers, capturing funnel-level influence. For post-purchase interventions, measure the conditional effect among purchasers. MVP causal representation: Promotion → Purchase ← Purchase Intent (unobserved) Purchase → Cancellation TLDR: Measure total cancellation impact for broad churn prevention strategies OR conditional impact among purchasers for tactical post-purchase interventions. Selection Bias | Conditioning on a Mediator Selection bias from conditioning on a mediator occurs when we control for a variable on the causal path from treatment to outcome. For example, notifications influence engagement, which impacts retention. Controlling for engagement removes part of the effect, giving an incomplete estimate. For attributing retention impact from notifications, focus on the total effect, capturing both direct impact and that flowing through engagement. To understand how much retention is driven by engagement, measure the mediated effect, which helps guide targeted engagement-driving initiatives without conflating total impact. MVP causal representation: Notification → Engagement (Mediator) → Retention Notification → Retention (direct) TLDR: Measure total impact on retention for measuring the success of feature introduction (e.g., notifications) OR the mediated impact from engagement to retention for targeted engagement strategies. Note: All opinions are my own and shared for data science community discussion.

99