Wojciech Galuba

Very proud to have worked on developing the first open source modelling suite for Marketing Mix Models (MMM), which we at Mutinex hope will lay the groundwork for a more democratic and transparent MMM landscape. The new validation suite enables practitioners to rigorously test and compare MMM models in a consistent, open framework. It's a step toward greater accountability, better standards, and shared progress in marketing science. We’re excited to share it with the community - contributions and feedback are more than welcome! https://lnkd.in/dmWgKUd2 https://lnkd.in/dWJPtr2m

10

2 Comments

Been tinkering with synthetic data for our llm pipelines at Nurau, where one of our biggest challenges has always been making it feel more human- not just statistically plausible and in avoiding uncanny valleys, but psychologically and linguistically aligned with the real people using our tools. I recently stumbled upon PILOT, a framework for steering synthetic data generation in llms using structured psycholinguistic profiles. It kicks off by translating natural language personas like "anxious introvert" or "confident expert" into multidimensional profiles covering traits like emotional tone, readability, and lexical diversity. From there, it guides the model via three methods: plain persona prompts, schema-based steering (think rigid templates for consistency), or a hybrid that mixes both. The result? Schema steering slashed repetitive phrasing by up to 40% and boosted coherence scores, while the hybrid nailed a sweet spot for diverse but controlled responses. Expert evals yielded high marks on quality, though it does trade some conciseness for richer vocab (there ought to be slider for pedantry 😅 ). In a world where synthetic data is dodging regulatory headaches and data scarcity (it's exploding in healthcare sims and financial modelling), this means more reliable training fodder that mirrors human norms, with correlations to gold-standard human judgments hitting 0.9. But, here's the thing: while PILOT may shine on behavioural control, without human oversight, we risk outputs that sound right but miss cultural nuances or ethical edges. Its implementation evidently still needs solid governance to build trust. At Nurau, we're exploring how this folds into continuous learning loops for production llms, making synthetic data a true partner in ethical AI scaling. you can check out the article here: https://lnkd.in/eeTqkQz2

34

2 Comments

Document-centric tasks sit at the core of many enterprise, business, and government workflows. Search is important, but the real challenge is going beyond retrieval, enabling systems to reason across documents, verify facts, and handle multi-step information tasks. Great to see a new model from Databricks moving in this direction. I expect we’ll see many more models, including open-weight ones, designed specifically for document-driven workflows, an area with huge potential across enterprise and government use cases.

23

1 Comment

#MeWriting The fundamental issue driving RAM availability and prices right now is an unwarranted belief that applying more compute to LLM training or inference makes them better. We are past any reasonable point of diminishing returns on scale up or scale out with LLMs. But we're pretending like those returns will be super-linear well into the future. For two years now, going back to the availability of Mistral 7B Instruct v0.2 and the PrivateGPT stack to run it, I have said that small LLMs feel as knowledgeable as the big ones, they just aren't as annoyingly loquacious. Gemma 4B as a daily driver first stop, running on a meager Raspberry Pi 5 without a GPU, does what I need in 19/20 text cuing cases. Probably because I know what LLMs work well for and don't work well for. If I'm running that Pi 24/7 at full CPU inferencing power, it costs me $4/month at ridiculously high SoCal standard residential electricity rates. Cheaper now that I'm in Nevada and have solar during the day. A completion ("answer") spits out faster than I can read along. Do I need two pages generated in 5 seconds? Here is the point. When our demand for LLM compute pares back to what we actually need to run the algorithms and give useful results — when we are actually "efficient" — we won't need so much cloud infrastructure and centralized stored of RAM, GPU, storage, and the like. Prices of the pieces will collapse. Ironically, Grok helped me illustrate this concept using $1M worth of cloud compute hardware for a few microseconds.

1

Excited to see LMEval, a new open-source framework for accurately and efficiently comparing large models. Giskard benchmark's Phare (phare.giskard.ai) leverages LMEval for safety and security. #AI #LLM #OSS Elie Bursztein Matteo Dora Alex Combessie 🐢 Kurt Thomas Guillaume Sibout

4

I went down a bit of a rabbit hole on AI spreadsheet ingestion this month, as I continue improving our data pipelines at Valae. There’s a lot of research and products in this space that don’t match what we see inside companies every day. Papers like Microsoft Research’s SpreadsheetLLM achieve high metrics for table extraction, but there’s always a catch. The success is measured on spreadsheets where the data is already pretty clear. Real company spreadsheets aren’t so easy. Sometimes they have three tables crammed into one tab, column headers that are someone's initials, and cell B7 containing "check with Marco re: pricing (NOT final)". In my research, I also dug into the survey "Toward Real-World Table Agents" (Tian et al., 2025) which confirms what I found after reading dozens of these papers: almost all LLM table research uses clean academic datasets. The messy reality of business data is barely ever addressed. That gap is where my team at Valae lives. We're perfecting proprietary processes for turning the messiest possible spreadsheets into clean, machine-readable knowledge, automatically. Despite SpreadsheetLLM, despite LlamaIndex launching LlamaSheets, despite years of research… nothing on the market truly handles the gap between what a spreadsheet contains and what it means. Encoding is a solved-ish problem. Interpretation isn't. And that's the one that matters. https://lnkd.in/er7PdDBP

16

1 Comment

Regarding the AI disruption of SW freakout - I am interested in knowing if there is demand/supply based market analysis of this? If so, please do comment and let me know. The current repricing of SW seems to assume SW is demand constrained?? - i.e. the same amount of SW can be generated much cheaper, and thus software vendors will face pricing pressure and competition. But what if SW is supply constrained? Why not a scenario where 10 times as much software can be sold for 2 times as much revenue using the same number of software engineers? Considering the premium multiple that SW companies have commanded, and the high salaries of SW engineers, I would guess that the market is more likely to be supply constrained.

22

4 Comments

Max Ashton-Lelliott wrote a brilliant paper on the power of structured data as an input into LLMs and agentic AI. I've covered this elsewhere but it's worth re-iterating, without adequate structured inputs you can't really get these tools into compliance workflows - not while keeping the economics of it all reasonable. And I may be speaking out of turn, but this isn't even about RegGenome's data (the best) - it's just madness to hope that a) these models will somehow be super-efficient while trying to digest what are essentially image files and b) regulators the world over will one day collectively start publishing everything in neatly packaged XML. This is where people like us come in.

18

3 Comments

We’ve been experimenting with something interesting in one of our products recently. Instead of going down the usual Vector DB + embeddings route for retrieval, we’re exploring a page indexing approach. The idea is simple: rather than embedding everything and relying purely on similarity search, we index information in a more structured way at the page level and retrieve it deterministically when needed. If you are someone working in this domain or building something interesting around PageIndexing , would love to have a chat and connect with you. My DMs are always open and you can also write to us at contact@nodevbuild.com Let's Build! Ashish Kumar Faiz Anwar

15

There’s a lot of momentum around agentic AI right now. 📌 In practice, many systems labeled “agents” are really workflow automation with LLMs and tool calling. That’s often the right solution - workflows are predictable, testable, and deliver value quickly. Problems show up when we deploy these systems expecting agent-level behavior. Workflows are optimized for correctness on known paths. Agents exist to recover when the path is wrong. If a system can’t recognize failure, update state, and re-plan its approach, autonomy won’t magically appear - no matter how many tools or prompt chains we add. I wrote a deeper breakdown (with diagrams + concrete examples) on Substack 👉 Curious how others are drawing this line in production systems.

29

4 Comments