Prasad Velaga, PhD

Over the holiday break, I spent some time revisiting some of the foundational work shaping the field of scientific AI. One paper that stands out—published back in 2021—was co-authored by Damien Guého, now a Senior R&D Engineer at Geminus. The paper tackles a timeless challenge: how do you uncover the true physics of a system when all you have is sparse, noisy, imperfect data? Damien and his co-author's approach is both clever and elegant. Instead of a brute force deep learning strategy, they leveraged clever math to uncover a physics-grounded structure with interpretable dynamics—from messy, nonlinear signals. The result is a robust, reliable and generalizable solution that uses a fraction of the computational effort. They first solved the problem using a deep learning approach that required 7,720 parameters. They then applied their method that used just 3 terms and was 10,000× more accurate. This mindset is core to Geminus' mission to simplify engineering and operational decision making in complex systems where uncertainty abounds and data is sparse and messy. Work like that of Damien and his colleagues continues to inspire where we’re headed next. Proud of the team at Geminus Labs—and excited for what’s coming. Check out his paper here. https://lnkd.in/eHt-5tXX Watch this space - be sure to follow Geminus to be part of the conversation.

32

Metis Analytics just benchmarked Railroad Memory against LongMemEval, the ICLR 2025 standard for long-term memory in AI agents. Current SOTA on LongMemEval is 82.4% (EmergenceMem, using frontier models and complex RAG pipelines). Railroad scored 61.2%. That's 74% of SOTA. GPT-4.1 with its million-token context window scored 56.7%. That's 69% of SOTA. Railroad Memory with Llama 3.3 70B and YAML files outperforms OpenAI's largest context window on the academic benchmark for long-term memory. Let that sink in. An open-source model. Text file storage. No vector database. No knowledge graph. No embeddings. Beating a million-token context window. Some other findings: 90% accuracy on abstention. This means Railroad correctly says "I don't know" when it doesn't have the information. Most AI systems hallucinate confidently. This is the difference between a tool you can trust and one you can't. 70.8% on knowledge updates. Railroad tracks changes over time better than raw conversation history. When the budget changed from $75K to $50K, Railroad remembers both values and which is current. Railroad Memory won 4 of 7 categories against Zep, which uses GPT-4o and a Neo4j knowledge graph. The takeaway: bigger context windows don't solve memory. Structured extraction does. The benchmark results, raw outputs, and reproduction scripts are all in the repo. Everything is verifiable. https://lnkd.in/gchcA-Cj

2

Skewed targets killed OLS until np.log1p(df['target'])linearized relationships—always plot residuals (plt.scatter(preds, residuals)) to validate post-fit. Ladder of powers (log, sqrt, boxcox) unlocks non-linear truths for linear models. Boosted R² from 0.4 to 0.82 overnight. Curve-fitting fixes that worked wonders?

1

From my new book AI-Powered Digital Twins: *Model Upgrades* AI models degrade over time through concept drift—the statistical properties of input data shift as the world changes, causing model accuracy to decay. A predictive maintenance model trained on equipment operating under normal conditions may perform poorly when supply chain disruptions force operation outside typical parameters. Continuous maintenance requires monitoring model performance in production, detecting degradation, and triggering retraining workflows maintaining prediction quality. * Retraining strategies* vary by model type and operational constraints. Some models retrain continuously on streaming data, adapting to evolving patterns in near real-time. Others retrain on schedules—nightly, weekly, monthly—balancing freshness against computational costs. Critical models may maintain shadow deployments where new versions process production data without affecting outputs, enabling performance comparison before cutover. A/B testing exposes portions of traffic to candidate models, measuring improvement before full deployment. *Model tuning* optimizes hyper-parameters and architectures for changing requirements. As computational budgets evolve, models can be scaled up for accuracy or down for efficiency. As new data modalities become available, models can incorporate additional inputs. As understanding deepens about what matters in predictions, model architectures can be refined to emphasize relevant features. This tuning represents continuous optimization along accuracy-efficiency-interpretability tradeoffs, adapting to shifting organizational priorities and technological capabilities. *Model replacement* becomes necessary when architectural limitations prevent further optimization. A classical machine learning model may hit accuracy ceilings that deep learning surpasses. A deep learning model may lack the interpretability required for regulatory compliance, necessitating replacement with inherently transparent alternatives. Foundation model advances may make custom training obsolete for applications adequately served by prompted or fine-tuned general models. Replacement strategies must account for validation requirements—proving new models match or exceed predecessors—and rollback plans enabling quick reversion if replacements underperform or introduce unforeseen issues. Photo today: Then and Now: * Then- I knew only AI theory, foundational and super important. * Now- I know integration, deployment, and business. 'Then' alone always felt like something big was missing... Not anymore ;) Follow for more... #digitaltwins #ai #models #upgrades #newbook O'Reilly Wiley

28

3 Comments

Sharing a list of the latest packages for using large language models in #Rprogramming. The packages let you incorporate #AI into data models while remaining within the #R environment. There can be a cost for token access through the #API, depending on the model -- Anthropic charges $5 for token access. But the cost is minimal, and many are free. This gives you an opportunity to explore prompts that can improve your decisions and analysis with #data model results.

6

1 Comment

Wright’s law states that as the production volume of an item increases, its cost falls and quality improves. These cost reductions and process improvements have a name. The Learning Curve. Computing’s exponential growth is powered by a derivative of Wright’s Law — Moore’s Law. Dr. Gordon Moore, founder of Intel, observed that the number of transistors on a computer chip doubles every 2 years or so. Wright focused on experience; Moore focused on time. You know what else follows a learning curve? Renewable Energy. In 1976, the price of a solar system, adjusted for inflation, was $106 per watt. Over the past 50 years, the price of solar systems has plummeted from $106/watt to $0.38/watt, a price decline of 99.6%. You know what does not benefit from Wright’s Law? Fossil fuels. Take a look here: https://ed.gr/eewd4 Why does renewable energy keep getting cheaper while fossil fuels remain stagnant? The key determinant here is the cost of the actual fuel. The cost of coal on avg makes up 40% of the total costs associated with burning coal to generate electricity. This means that no matter how efficient we make coal plants, there’s always going to be a significantly higher fixed cost element where Wright’s Law doesn’t apply. With a low-cost floor, as exists with renewables, there is exponentially more room for growth, allowing Wright’s Law to work. That’s the power of learning. (8/8)

8

1 Comment

The buildout of the ecosystem to support AI computing is clearly showing up in the producer price index data. Two charts below showing year-over-year percent changes. Thoughts: •The top chart shows the yearly change in the price for domestically manufactured electrical equipment (NAICS 3353). Prices were up 8.6% year-over-year as of January. For comparison, prices were up just 2.9% at their peak rate of increase in 2018. Last year, YoY prices were +4.3% in January 2025. As such, I suspect some of the additional upward pricing pressure is due to tariffs affecting imported intermediate inputs as well as raising the cost of domestic inputs like aluminum and steel. •The bottom chart shows yearly changes in the price for domestically made navigation, medical, and control instruments. Prices showed a massive upward spike in January such that year-over-year percent change was 10.6% in January. For comparison, during the worst of the 2021-2022 inflationary cycle, prices rose 6.4% year-over-year at their peak (August 2022). •Why would navigation, medical, and control instruments be affected? My speculative answer: the worsening memory chip shortage. These products require memory chips, and as Bloomberg has written, the world is seeing a worsening shortage of memory chips. To quote from one article (https://lnkd.in/ggCgVcjH), “The tariffs and pandemic crisis seem a joke compared to this [current memory shortage]. A separate Bloomberg article has written about large price increases for memory chips so far this year (https://lnkd.in/g5a4ZYqn). Tariffs on intermediate inputs may also play a role in 2025's increases but can't explain the huge January 2026 jump. Implication: the two wildcards for inflation in 2026 are (i) a US war against Iran and possible impacts on crude oil prices and (ii) worsening shortages due to the incredible demands being placed by the buildout of the ecosystem to support AI computing. One key finding is that firms experiencing shortages will tend to sharply raise prices (https://lnkd.in/gw_zNCzf). When you then consider that energy prices declined in 2025, the message is clear: hold interest rates where they are unless the labor market significantly deteriorates. #supplychain #supplychainmanagement #markets #economics #freight  

33

1 Comment

Throughput Thursday: Are Your Bottlenecks Hiding in Plain Sight? Most manufacturers blame throughput issues on labor or equipment. But lately, we’re seeing more hidden constraints hiding in plain sight. Consider these real-life examples- • Inaccurate BOMs • Inaccurate routings • Poor demand signals (or no demand signals, other than orders!) • Poor inventory record accuracy • Lack of alignment between scheduling, procurement, and real-time capacity • Poor layout We helped one client increased throughput 12% just by reducing setup time. We did this by putting all his supplies on a cart, so he didn’t have to walk to the supply room to get what he wanted, and we organized the work process better. The result? An easy 50% reduction in set up time! Remember: Throughput doesn’t just happen on the floor—it starts with what (and how) you feed the system.

7

8 Comments

Boundless Opportunity in Florida (3rd largest state) for Construction Equipment OEMs, Dealers and Rental Companies Based on BiltData.ai's analysis of 100M+ transactions, BiltData has examined the diverging paths inside the U.S. construction equipment and rental industries, and shows OEMs, dealers and rental companies where opportunity resides. Florida is powerhouse, the 3rd largest among U.S. states in total construction spending by 2030, trailing only Texas and California. The state currently represents ~7% of all USA construction spending in 2025, and is projected to grow to 8% by 2030 driven by Florida's rapid population growth, infrastructure needs, and data center projections. By example, Florida is poised for substantial data center growth between 2025 and 2030, transitioning from a secondary market to one of the fastest growing U.S. hubs. This surge is fueled by AI hyperscale demand, Florida’s sales tax exemptions on data center equipment and electricity (for facilities with >100 MW). Most importantly, utilities are investing billions in hardening and new power generation. Florida's Top 5 MSAs (metropolitan statistical areas) in 2030 account for ~80% of the state's construction spending: 1. Miami-Fort Lauderdale-West Palm Beach 2. Tampa-St. Petersburg-Clearwater 3. Orlando-Kissimmee-Sanford 4. Jacksonville 5. North Port-Bradenton-Sarasota Florida's powerhouse economy directly boosts demand for construction equipment, particularly earthmoving. Biltdata’s predictive buyer intelligence has identified significant construction equipment sales, rental and product support opportunities identifying a small percentage of contractors who will will drive ~45% of demand. BiltData.ai's predictive buyer ops turns raw data into actionable intel for construction equipment OEMs, dealers, and rental companies, optimizing marketing, inventory, and sales for growth in high-demand sectors. Our proprietary BiltReady Opportunities provides intel with a 100% view of actionable opportunities: Sales, Rental, Parts & Service. Behind BiltData is a proprietary AI engine trained on 100M+ transactions. No Opportunity. Left Behind.

14

Greg Lambert just described my product without knowing it exists. His piece in The Texas Lawbook today, "After the Claude Crash," nails the core shift: the moat moved from search to verification. Agents can draft. The bottleneck is proving the output is trustworthy. He's right. Where I disagree is the default architecture implied by most "agentic workflow" talk: cloud pipes. MCP connectors feeding workflows through Google Drive, Microsoft 365, Slack, Box. That's not my architecture. Here's what happened this week while he was writing that article: On Feb. 10, Judge Rakoff ruled from the bench in United States v. Heppner that 31 documents a defendant generated using Anthropic's Claude and later shared with defense counsel were not protected by attorney-client privilege or work product. The AI tool's own terms did not treat inputs or outputs as confidential. Sending those documents to your lawyer after the fact did not fix it. That ruling is about a consumer AI tool. Now ask the harder question: how many of the cloud pipes in your workflow have terms that clearly and enforceably commit to confidentiality for every document your agent touches? How many have you actually read? Every tool in the workflow that lacks enforceable confidentiality terms is a surface where privilege can break. Governance without infrastructure integrity is a policy document, not a solution. I built CloseVector to solve the layer his article is missing. Air-gapped eDiscovery that deploys on hardware your firm owns outright. No cloud. No external APIs. No third-party data access. Citation-gated retrieval, full audit trails, and attorney-ready evidence packaging designed to support ABA Formal Opinion 512 duties and defensible FRCP 26(g) certification. Lambert diagnosed the problem. I built the infrastructure to address it. If you're building your firm's future on pipes you don't own, you're just waiting for your own Heppner moment. Let's talk. DM me or book time directly. Links in first comment.

4

3 Comments

A new curriculum on ChapterPal: Must read papers on variational autoencoders Variational Autoencoders (VAEs) are a class of generative models that learn compact, meaningful representations of data while enabling the creation of new, realistic samples. By combining principles from deep learning and probabilistic modeling, VAEs provide a structured framework for understanding complex data distributions, making them essential for tasks such as image generation, anomaly detection, representation learning, and latent space exploration. This curriculum provides a foundational and advanced understanding of Variational Autoencoders, equipping the learner with the critical knowledge to build and innovate upon this powerful class of generative models. Learners will trace the evolution of VAEs from their inception to their state-of-the-art applications, mastering the core principles and key advancements. Learn from the best on ChapterPal: https://lnkd.in/exUhcfbc

50

4 Comments