Announcing LIISA! Our Legal Identity Intelligent Suffix Adjudicator
Today we are announcing the launch of LIISA, our Legal Identity Intelligent Suffix Adjudicator, which enables users to receive more accurate results without adding false positives. LIISA looks at every entity – companies, organizations, or groups – in our database and adjudicates what is important contextually and what is not. LIISA enriches entity profiles with new, additional aliases that are not present on government watchlists. With LIISA, users have the ability to screen hundreds of thousands of relevant aliases in our global risk database, while also having the flexibility to turn these aliases on or off through our Weak AKAs classification.
What does LIISA Do?
From day one, Castellum.AI’s calling card has been finding THE needle in the haystack, so imagine our surprise when users said their top request is that we show more needles (i.e. results). This wasn’t an issue of false negatives, rather it was the difference between use cases: transaction and client screening v.s. ad-hoc screening.
In transaction and client screening input is generally standardized and comprehensive. For example, a corporation’s full legal name, address, registration number and more are screened. Ad-hoc screening is the opposite, users’ input is generally only the most memorable part of a name, often omitting over 70% of the total characters in an entity’s full legal name.
Real examples that our users shared with us, that we now solve for, include:
The above results include searches across all watchlists — sanctions, politically exposed persons, export control and more.
How LIISA Compares to Competitors’ Models
The problem isn’t how to show a user what they’re searching for, it's how to do so without overwhelming them with false positives even when less than 1/3 of the legal name is entered. Some companies have tried solving this through vintage algorithms, some of which date from the 1960s, such as Soundex, Levenshtein or Jaro-Winkler. These dated approaches use arithmetic to compare character input against total name character lengths.
Others have used an even less refined approach by taking each word of a search term, and returning every result that includes the search term. This is the approach used by the official UK HMT OFSIs website. For example, if you search Bank Otkritie in the UK government database, you get 494 results, including completely irrelevant results like Commercial Bank of Syria and Musa Kalim Hawala, a terrorist in Afghanistan whose street address includes the word “Bank.”
By comparison, if you search UK HMT sanctions on Castellum.AI, you get one result for Bank Otkritie. Castellum.AI is 99.8% more accurate than the UK HMT sanctions search! If the search is expanded to cover Castellum.AI’s global sanctions database, there are 10 results, all relevant sanctions against that bank from other authorities such as Switzerland SECO or Australia DFAT, with zero false positive results.
Try sanctions screening with LIISA
The difference between our approach and that of governments and competitors is our focus on context, not arithmetic. Rather than approaching the question of “how do we show users similar results to their search, we focus on “how do we show users what they want to see.” This is in large part driven by our open and free screening platform which enables us to gather user testimony and clear data on how people search and what is important to users.
LIISA Example Enrichments
Bank of Kunlun Co Ltd and Kunlun Holding Company Ltd.: Both of these entities are sanctioned by the US OFAC, and LIISA created different aliases for each:
Bank of Kunlun CO LTD → Bank Kunlun
Kunlun Holding Company Ltd.→ Kunlun
Although the most unique term for both entities is Kunlun, “Bank” also provides contextual value. “Holding Company Ltd.” and “Co Ltd” do not. Using LIISA, users searching for “Bank of Kunlun CO LTD” can input any variation of Kunlun + Bank, and users searching for “Kunlun Holding Company Ltd.” can simply search “Kunlun” to receive accurate results.
Another example is Subsidiary Bank Sberbank of Russia Public Joint Stock Company. LIISA generates the following relevant alias:
Subsidiary Bank Sberbank of Russia Public Joint Stock Company → Sberbank
LIISA does not create an alias of “Bank Sberbank” because it understands that the first “Bank” does not add value, and based on user input, is not how people search this term.
LIISA’s ability to create contextually-derived aliases reduces false positives. A user searching “Kunlun Bank” or “Bank of Kunlun” (using default match score setting of 85) will see the result for “Bank of Kunlun Co Ltd” but not “Kunlun Holding Company Ltd.” Likewise, a user searching “Kunlun” will see “Kunlun Holding Company Ltd.” but not “Bank of Kunlun Co Ltd.”
Notably, LIISA does not alter our Jgram search algorithm. By focusing on the creation of new, relevant aliases based on watchlist entry names, LIISA responds to what users have indicated they want to see.
How To Work With LIISA
Even though we think LIISA is incredible, it is crucial that you as a user retain flexibility in your search, and as such, all of the new aliases created by LIISA are marked as Weak AKAs. Thus you can turn off these new aliases by shifting the Weak AKAs toggle in our Advanced Search, as seen below:
Our official recommendation is that Weak AKAs be disabled for screening where you have detailed client information, like in transaction and client screening, but enabled for ad-hoc screening or situations where you believe you might be missing significant portions of relevant information.
Try Out LIISA’s Enrichments
Want to see more about LIISA’s impact? Create a free account or log in and try these searches:
Aeroflot
Hapjanggang
Sanam
TURQUESA
About Castellum.AI
Castellum.AI automates compliance screening by providing watchlist screening solutions through online platform, API, and bulk data subscriptions.
Castellum.AI obtains global sanctions information directly from authorities issuing sanctions, and then proceeds to standardize, clean and enrich the data, extracting key information like IDs and addresses from text blobs. Castellum.AI enriches as many as fifteen separate items per entry.
The database consists of over 1,000 watchlists, covering over 200 countries and eight different categories (sanctions, export control, law enforcement most wanted, contract debarment, politically exposed persons, regulatory enforcement, delisted, and elevated risk). Castellum.AI checks for watchlist updates every five minutes directly from issuing authorities.
