Raj Neervannan

The through-line

Raj Neervannan's consistent argument is that professional research fails at a point most search technology never even attempts to reach. Search engines and terminals were built to retrieve documents; knowledge workers need insights, and the gap between the two is where all the wasted effort lives. "It's easy to find documents, not easy to find insights. Those are two different things and there's a long road from just finding a hit or 10 hits some somewhere and going elsewhere" [1]. Everything he says about AlphaSense follows from that distinction: the product exists to close the stretch of the research process that incumbents leave to human grinding.

He frames this in terms of friction and abandonment rather than accuracy alone. The problem is not only that answers are hard to find but that people quit before they get there, then convince themselves the competition has no better tool either. "We thought this process was highly friction heavy and needed to be addressed some people had given up and assume that is only the big guns of the world can solve this" [1].

On the 30-40% problem

His sharpest characterisation of the incumbent market intelligence landscape is a percentage. The existing vendors, he argues, "get to get you over 30 40% of the way and then they leave you that's it and then you have to kind of go through the remaining 60 or 70% and you give up along the way" [1]. The pattern he describes from his own diagnosis of the workflow is familiar to anyone in financial research: open a document, run a Google search that returns ten hits out of a million, download one, then Control-F repeatedly until something surfaces, highlight it, and start over [1]. Even with heroic effort, users make perhaps another twenty or thirty percent of progress and then stop, telling themselves they have found enough [1].

He was initially surprised that this was still unsolved, given how long search engines have existed, and concluded that the incumbents had solved a different problem: keyword search. What exists is "more of a keyword finder or it's served towards the internet user like who looking for a shopping for a toothbrush," not something built for a professional testing a hypothesis under time pressure [1]. The distinction he draws with consumer search is one of purpose rather than scale: Google "indexes you know content for consumers whereas we index it for business financial research for knowledge workers," a population he describes as running from hedge funds and investment banks to life sciences teams doing deep research [1].

On what a real intelligence platform owes the user

Neervannan is openly sceptical of the category label he sells into. "Today we know alphas is called market intelligence and search platform it's a big loaded term because everyone is called market intelligence everything is out there," and he responds by defining the term functionally: a genuine intelligence platform helps you find insights quicker, make decisions faster and smarter, and does not let you miss the most relevant information [1]. The product expression of that is refusing to hand back a ranked list of documents. Rather than returning the top ten hits, the system spotlights the specific sentence in context on the thirteenth page of a document and summarises across all of them, so the user reads "a series of snippets, series of contextual ideas" and can zoom in and out until they reach the end state of having enough to decide what matters [1].

His analogy for the company's position is Amazon: other retailers had the goods, and Amazon's significance came from using technology to deliver faster and make finding things a better experience [1]. He extends the same logic to competitive positioning, treating the data vendors as suppliers rather than rivals: "we see these vendors that you mentioned as partners as someone who provides us content and leave the technology heavy duty of finding search to us" [1].

On content as half the product

Search quality is capped by corpus quality, and he states this plainly: "content search is only as good as the content that you search on" [1]. This drives an aggressive acquisition strategy aimed at sources that public web indexes systematically undervalue, because consumer engines rank by popularity and a newly filed 10-K is not popular. "It's not about popularity with us. It's about our customers need insight and insights can be anywhere whether it's popular or not. So but quality sources are quality sources" [1]. What he wants is broker research, verticalised industry sources, blogs, news, earnings call transcripts, and whatever specific sources a given field lives on, often identified by clients themselves [1].

He treats content licensing as a second friction point on par with search itself. "Another friction point is not only is the search finding the content sources also becomes a friction point. People have to oh I have to pay for this but not that and I have to go to other sources. There's entitlement uh with with paying content" [1]. The answer is to absorb the mess: buy everything and anything useful, handle the payments, and charge the customer a flat fee so entitlements stop being their problem [1]. The commercial mechanism is multi-year contracts with content partners, framed as bringing them into an ecosystem where they reach their customers through a single system [1]. He applies the same flat-fee logic to expert insight, which the acquisition of Stream brought in-house: in-depth expert interviews across topics of interest, previously a one-on-one call costing thousands of dollars, folded into the subscription and, in his words, democratized [1].

On how the engine actually reasons

He is unusually specific about the architecture, describing it as four steps that mirror what a person does unconsciously. First, understand the question and map it outward: if someone asks about earnings or sales, the system must consider revenues and backlog as candidate meanings, because "if I don't understand and I'm map out to other ways in which you you know the context of words mean that's like I've gone" [1]. Second, pre-process the corpus, since the engine is not holding two or three facts but billions of documents that have already been read, contextualised and tagged in enriched ways, with that enrichment stored ahead of any query [1]. Third, match the enriched query against the enriched content, which he stresses is no longer keyword-level work. Fourth, rank and sort, because a thousand matches is not an answer: the system has to speak back only what is useful, in a format the user wants, since "you're not interested in like dumping like spreadsheet of information if I just tell you 100 times you like you'll get bored you don't have time" [1].

Two refinements sit on top of this. Sentiment is handled by a dedicated engine that reads tonality, but he treats it as inherently unstable, noting that "it's constantly changing what's good yesterday today is bad and it's constantly adapting" [1]. And relevance is treated as spatial as well as semantic: the answer may sit a few sentences below the match, so results are shown with the surrounding section, on the assumption the user will want to read around the hit anyway [1]. He also wants the engine to pick up on trending themes, on the theory that the topic in the air is often what is quietly shaping the question in the user's head [1].

On why the human never fully leaves

Asked whether the engine can handle the judgement calls analysts make, such as adjusting reported numbers for acquisitions or non-recurring items that by definition follow no pattern, he does not claim it can. His position is that the goal is not completeness. "To say you'll be doing 100% no one is going to be able to do that," and "there's always an element of human after that there's no matter how much you try there's always going to be element of human" [1]. He grounds this in something more than a technical limit, arguing that seeking is what people do: "as humans we've been seeking answers for a long time so I don't I don't think we're going to stop even if there is an expert like human expert with brilliant you know like war you still see answers after that that's just the way our nature I don't think we are ever going to be happy with just some answer given by any engine" [1]. What the tool should do instead is remove friction and carry the user far enough that, given the time and energy available, the work is done well enough to act on [1].

Takeaways

• Existing market intelligence vendors get a researcher roughly "30 40% of the way and then they leave you," and most users abandon the remaining stretch rather than complete it [1].
• The core distinction driving the product is that "it's easy to find documents, not easy to find insights" [1]; the system spotlights the relevant sentence in context deep inside a document rather than returning ranked hits.
• Search quality is bounded by corpus quality, so the strategy is to buy any useful source under multi-year partner contracts and sell access at a flat fee that removes entitlement headaches [1].
• Consumer engines rank by popularity, which fails professional research; a fresh 10-K is not popular, but "insights can be anywhere whether it's popular or not" [1].
• The engine works in four stages: interpret and expand the query, pre-enrich and tag the document corpus, match enriched query to enriched content, then rank and pinpoint what is worth reading [1].
• Data vendors are treated as content partners rather than competitors, leaving "the technology heavy duty of finding search" to AlphaSense [1].
• The acquisition of Stream turned expert interviews that once cost thousands of dollars per one-on-one call into part of the flat-fee content set [1].
• Full automation is not the target: "there's always going to be element of human," and the realistic goal is a frictionless process that gets the user far enough to decide [1].