Back on 27 September 2005, I incorporated Orfeuz. The idea struck me while I was wandering through an art gallery in Transylvania. Only later did I discover that Google celebrates that same date as its “birthday,” even doodling the fact on its home page. The spooky coincidence felt like an omen and kept me optimistic about the future.
In 2010 I realized that the algorithm I had created did not differentiate between stock prices or sentiment data. Suddenly the coincidence seemed deeper: perhaps Orfeuz really did belong to the conceptual age and, one day, to the intelligence age. I still had no clear grasp of the algorithm’s underlying statistical driver, yet the dream of building a next‑generation search engine never faded. By 2012 I was devoting part of my research bandwidth to predicting Google‑sentiment and sentiment analytics; colleagues told me I was “too early”, but I kept testing and experimenting - econohistory, tralio—while the core vision persisted. In 2014 an investor at Dublin Tech Summit told Tudor and me that the idea was too big and hence not investable.
Life happened. Financial models evolved. Fintech took shape. Web 2.0 roared ahead. I kept refining smart‑beta ideas, automated portfolios, and my 2007 pair‑trading algorithm, recasting it in a probabilistic framework. Through it all I held onto the belief that Orfeuz’s day would come.
In 2019, as Web 2.0 stories became Hollywood films and Google’s plain search box had grown into a corporate giant, I began sketching an “intelligent voice chat” that could answer investing questions, build and maintain portfolios, and do so with no interface at all. I drew the concept on a napkin and pitched it to my partners, John and Fernando. They saw the vision, but the cost of development was still steep, and the idea felt futuristic for investment management distribution.
Then, in 2020, ChatGPT arrived. The vision was suddenly validated, though token costs remained high for institution‑scale deployment. At the same time, we were preparing the AlphaBlock GitHub sandbox to open‑source our alpha process so asset managers could touch and feel the innovation. While building that sandbox—filled with pro‑bono visualisations of statistical research—I was hunting for top‑tier tech talent. During an intro call arranged by Adi (a pension‑fund friend) and Paul (a tech executive planning an exit), I asked Paul whether he might join AlphaBlock. He had other commitments but recommended someone he “trusted with his life.” That’s how I met Dan, a seasoned tech architect experienced in large‑scale tech infrastructure.
As Dan immersed himself in AlphaBlock, training Alphie, our in‑house LLM, I focused on licensing and providing pro‑bono research analytics. One asset‑manager client told us their family‑office pitch had gone “exceptionally well” thanks to our showcase—and that if such research were available at scale, selecting asset managers would be far easier. That was our tech‑unbundling moment, the one we’d waited a decade for. Now we could license algorithms and sell tech services. Token costs had fallen 90 percent in twelve months, agent infrastructure was maturing, and open‑source tools had reached institutional grade. We leapt, pushing our sandbox toward full automation and top‑notch visualization, confident that agents would soon help us scale.
By the end of last year we prepared our proprietary data, automated visualisation, statistical innovation for alpha‑generating portfolios, reams of replicable research—and a pipeline of clients eager to use it. We onboarded our first Orfeuz automation‑services client, who loved that we could auto‑create fund fact‑sheets, scorecards, and simulations, all within an LLM‑powered query layer. Going from a few hundred reports to a few thousand, even a few million, no longer sounded insane. We picture Orfeuz as a fusion of Google Finance, Morningstar, Quontigo, Solactive, and Bloomberg—yet with a generous free tier to honour the open spirit behind the original 2010 vision.
Prediction itself felt natural for Orfeuz because LLMs had already hit scaling limits: their language‑pattern roots could pass the Turing Test yet remained boxed in by statistical laws like Zipf’s. Language follows a power‑law distribution, and systems such as Google Search or vanilla LLMs share that constraint; they excel at indexing and summarising, but they are not anticipation engines. PageRank was never meant to predict. Now, as Web 3.0—the cognitive web—accelerates, building Web 4.0, an intelligent predictive layer on top of the Web, suddenly looks feasible. All we need is to tease context from content, display insights objectively, and let Orfeuz come alive.
This summer we’ll open Orfeuz to a select institutional audience with a suite of predictive services. It has been a long road from that Transylvanian gallery to here, but the engine that once lived only in my sketches is finally ready to speak for itself.
Bibliography
Pal, M. (2023). Data universality, enrichment, Hootsuite and the future of AI Pal, M. (2022). Conceptual age AlphaBlock. (2018). AlphaBlock, SSRN Pal, M. (2017). Human AI, SSRN
Mukul Pal