<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: TechnoXander</title>
    <description>The latest articles on DEV Community by TechnoXander (@technoxander).</description>
    <link>https://dev.to/technoxander</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4193861%2F6f018e05-6bd4-4db7-a20c-e5c22688e631.png</url>
      <title>DEV Community: TechnoXander</title>
      <link>https://dev.to/technoxander</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/technoxander"/>
    <language>en</language>
    <item>
      <title>Designing a name-matching service people can argue with</title>
      <dc:creator>TechnoXander</dc:creator>
      <pubDate>Fri, 13 Nov 2026 12:33:28 +0000</pubDate>
      <link>https://dev.to/technoxander/designing-a-name-matching-service-people-can-argue-with-bc6</link>
      <guid>https://dev.to/technoxander/designing-a-name-matching-service-people-can-argue-with-bc6</guid>
      <description>&lt;p&gt;Name matching looks like a solved problem until you deploy it in a market where names carry particles, transliterations, honorifics and business trading names. Then it becomes a product decision disguised as an algorithm.&lt;/p&gt;

&lt;p&gt;Here is the design I would defend in a review.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Separate the score from the outcome.&lt;/strong&gt; The matcher returns a score and a reason. The scheme mapping turns that into Match, Close Match or No Match. Keeping these apart means you can retune the thresholds without retraining the model, and you can explain a result without exposing the model.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Make the reason a first-class field.&lt;/strong&gt; Not a debug log. A field that travels with the response, appears in the portal, and can be shown to a customer. Something like &lt;code&gt;reason: TRANSLITERATION_VARIANT&lt;/code&gt; or &lt;code&gt;reason: BUSINESS_TRADING_NAME&lt;/code&gt;. This is the single highest-value change you can make, because it turns an unexplainable answer into a reviewable one. The matching considerations are set out in &lt;a href="https://technoxander.com/importance-of-matching-algorithms-in-vop" rel="noopener noreferrer"&gt;matching algorithms in VoP&lt;/a&gt;, and the reason field is what makes them operable.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Version everything.&lt;/strong&gt; Matcher version, threshold set version, normalisation rules version. Store the version on every response. When match rates shift in six months, you will need to know whether the population changed or the logic did.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Normalise explicitly, not magically.&lt;/strong&gt; A written order of operations, applied in a fixed sequence:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;1. Unicode NFKC
2. Trim, collapse whitespace
3. Case fold
4. Remove punctuation except hyphens and apostrophes
5. Expand known honorifics (Mr, Mrs, Dr, Mx)
6. Handle particles (van, von, de, bin, al) as tokens, not noise
7. Transliterate to a canonical script
8. Preserve the original string alongside the canonical one
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Step 8 is the one teams skip, and it is the one that saves you in a complaint. Never destroy the input.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Do not normalise away meaning.&lt;/strong&gt; "Mohammed" and "Muhammad" may be the same person. "Al Rahman Trading" and "Al-Rahman Trading Co" may be the same entity. "Smith" and "Smyth" may not be. A matcher tuned for recall will merge the last pair and create a false positive that a human then has to unwind. The &lt;a href="https://technoxander.com/vop-name-matching-quiz" rel="noopener noreferrer"&gt;VoP name matching quiz&lt;/a&gt; is a useful calibration exercise for a product team, because it forces the argument about which errors you would rather make.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Treat business names as a different problem.&lt;/strong&gt; Legal name, trading name, group name and brand name are four different strings. A personal-name algorithm applied to companies produces Close Match noise at scale. If your institution serves corporates, this is where the exceptions queue comes from.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Log the decision, not just the request.&lt;/strong&gt; Every response should record: input string, canonical string, matcher version, score, threshold band, outcome, and the reason. That record is what you produce when someone asks why a payment was allowed.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Test with the population you have.&lt;/strong&gt; A test suite of synthetic European names will pass and tell you nothing. Sample real name pairs from your own exceptions queue, with permission and with the data minimised, and use those. The distribution is the product.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Publish the limits.&lt;/strong&gt; Tell the consumer of the API what the matcher does not do. It does not verify identity. It does not confirm ownership. It does not resolve transliteration across every script. It gives you a signal about a string. Teams that overstate the signal end up in disputes about what the check meant.&lt;/p&gt;

&lt;p&gt;Two operational notes.&lt;/p&gt;

&lt;p&gt;First, the &lt;a href="https://technoxander.com/routing-and-verification-mechanism-in-vop" rel="noopener noreferrer"&gt;routing and verification mechanism&lt;/a&gt; determines which institution you are asking, and a correct answer from the wrong institution is still wrong. Log the route alongside the result.&lt;/p&gt;

&lt;p&gt;Second, the scheme rules are the contract. &lt;a href="https://technoxander.com/epc-vop-rulebook" rel="noopener noreferrer"&gt;The EPC VoP rulebook&lt;/a&gt; defines the message set and the expected behaviour, including what to do when no response arrives. Build the timeout path before the happy path. It is the part that runs during an incident.&lt;/p&gt;

&lt;p&gt;The same argument applies to investigation: &lt;a href="https://technoxander.com/aml-ai-investigation-throughput-scaling/" rel="noopener noreferrer"&gt;retrieval, not judgement, is the bottleneck&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;The conclusion is simple. A matcher people can argue with is worth more than a matcher that is 2 percent more accurate, because the arguable one gets used, monitored and improved. The accurate one gets switched off after the first complaint nobody could explain.&lt;/p&gt;

</description>
      <category>api</category>
      <category>architecture</category>
      <category>compliance</category>
      <category>payments</category>
    </item>
    <item>
      <title>What 'grounded' actually means in an AML summary</title>
      <dc:creator>TechnoXander</dc:creator>
      <pubDate>Fri, 13 Nov 2026 12:16:33 +0000</pubDate>
      <link>https://dev.to/technoxander/what-grounded-actually-means-in-an-aml-summary-5fna</link>
      <guid>https://dev.to/technoxander/what-grounded-actually-means-in-an-aml-summary-5fna</guid>
      <description>&lt;p&gt;Every vendor in this category says their summary is grounded. Almost none of them will tell you what that means operationally, and the ones who will usually mean "we put citations in the output".&lt;/p&gt;

&lt;p&gt;Citations are the easy part. Grounding is a property of the whole pipeline: what was retrieved, what was excluded, what the model was allowed to conclude from it, and whether a reviewer can reconstruct the path six months later when the case is challenged. If any of those four is missing, the summary is not grounded. It is plausible.&lt;/p&gt;

&lt;p&gt;Here is how I would test it.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;&lt;p&gt;Retrieval coverage. Take a closed case with a known outcome. Ask the system to produce a summary and then ask it, explicitly, which sources it did not consult. A grounded system can answer that, because retrieval is a separate step with a defined source list. A system that cannot answer is generating from a prompt, not from evidence. The design question is spelled out in &lt;a href="https://technoxander.com/rag-ai-aml-investigation-explainable" rel="noopener noreferrer"&gt;RAG in AML investigation&lt;/a&gt;, and the test is simple: if the source list is not a first-class object, you cannot audit the output.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Exclusion logic. Most real investigations are shaped by what was left out. The transaction that looked odd but had a documented explanation. The prior case that was closed with a clear rationale. A summary that cites everything and weighs nothing is not analysis. Ask the system how it handled a specific exclusion. If the answer is a sentence rather than a reference, you have a writing assistant.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Claim-to-source mapping. Every factual sentence in the summary should map to a source record, and every inference should be labelled as an inference. This sounds pedantic until the first time a supervisor asks why the summary says "unusual counterparty jurisdiction" when the customer had declared that corridor. That single error, repeated across a portfolio, is how a model risk finding starts.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Temporal integrity. AML cases are time-bound. A summary written today about activity from fourteen months ago must not import today's screening result as if it were known then. Systems that rebuild the customer profile at query time rather than at case time will silently rewrite history. This is the least discussed failure mode and the most damaging in a SAR context.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Regeneration behaviour. Ask the same question twice. If the summary changes materially with no new evidence, you do not have a controlled process. Some variation is fine in prose. Variation in the findings is not.&lt;/p&gt;&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Now the harder point.&lt;/p&gt;

&lt;p&gt;Grounding is not only a retrieval problem. It is a memory problem. An investigator reaching a conclusion needs to know what the institution decided in similar cases, not just what the transaction data says. That is &lt;a href="https://technoxander.com/organisational-memory-aml-ai-investigations" rel="noopener noreferrer"&gt;investigation organisational memory&lt;/a&gt;, and it is usually the last thing built because it is the hardest to model. Cases are messy, rationales are inconsistent, and the older records are often PDFs.&lt;/p&gt;

&lt;p&gt;If you skip it, you get a system that is excellent at describing a single case and useless at explaining why this case is different from the last twenty. Analysts notice within a week. They start opening the old cases in a second window, and the productivity gain quietly disappears.&lt;/p&gt;

&lt;p&gt;There is a regulatory dimension too. Model governance expectations do not stop at "the model works". They extend to whether you can explain the output, evidence the oversight, and reproduce the decision. &lt;a href="https://technoxander.com/fca-aml-ai-model-governance-expectations" rel="noopener noreferrer"&gt;FCA expectations for AML AI governance&lt;/a&gt; is the practical frame here: documentation, validation, monitoring and a named owner. A citation in a PDF is not a control. A retrievable evidence chain with a human sign-off is.&lt;/p&gt;

&lt;p&gt;Two implementation notes, from having watched this go wrong.&lt;/p&gt;

&lt;p&gt;First, log the retrieval, not just the generation. Store the query, the source list, the version of each source and the timestamp. Storage is cheap. Reconstructing a case from memory is not.&lt;/p&gt;

&lt;p&gt;Second, make the summary rejectable. If the reviewer can only accept or edit, you cannot measure whether the summary was any good. Give them a one-click "not grounded" flag with a reason, and review those weekly. That single field will tell you more about model performance than any benchmark, because it captures the cases where the model was confidently wrong.&lt;/p&gt;

&lt;p&gt;The conclusion is unglamorous. Grounded means reproducible. Reproducibility is also the throughput argument: &lt;a href="https://technoxander.com/aml-ai-investigation-throughput-scaling/" rel="noopener noreferrer"&gt;reproducible output is a throughput argument&lt;/a&gt; is the commercial version of the same point. If a different reviewer, six months later, cannot follow the same evidence to the same conclusion, the system is generating text about a case rather than investigating it. Everything else, including the demo, is decoration.&lt;/p&gt;

</description>
      <category>aml</category>
      <category>ai</category>
      <category>llm</category>
      <category>discuss</category>
    </item>
    <item>
      <title>Bug Report: Analyst Capacity Leaks Into Retrieval</title>
      <dc:creator>TechnoXander</dc:creator>
      <pubDate>Wed, 11 Nov 2026 14:14:22 +0000</pubDate>
      <link>https://dev.to/technoxander/bug-report-analyst-capacity-leaks-into-retrieval-1007</link>
      <guid>https://dev.to/technoxander/bug-report-analyst-capacity-leaks-into-retrieval-1007</guid>
      <description>&lt;p&gt;Engineers know how to write a good bug report. Compliance rarely gets one. So here is one, for the most expensive defect in many AML operations.&lt;/p&gt;

&lt;p&gt;Title: Analyst capacity drops sharply when retrieval is manual&lt;br&gt;
Severity: high, but invisible on every dashboard&lt;br&gt;
Environment: an investigation team with several systems and no single case view&lt;/p&gt;

&lt;p&gt;Steps to reproduce&lt;/p&gt;

&lt;p&gt;Open an alert.&lt;br&gt;
Try to find the customer’s earlier cases.&lt;br&gt;
Try to find linked accounts.&lt;br&gt;
Try to assemble a timeline.&lt;br&gt;
Observe how long it takes before you can start thinking.&lt;/p&gt;

&lt;p&gt;Expected: the case opens with history, links and a draft chronology.&lt;br&gt;
Actual: the analyst assembles them by hand, then decides.&lt;/p&gt;

&lt;p&gt;Workaround: ask a senior colleague. This works until that colleague leaves.&lt;/p&gt;

&lt;p&gt;Suggested fix: retrieve before the analyst opens the case, show sources, keep the person in charge of the decision. See the idea behind &lt;a href="https://technoxander.com/rag-ai-aml-investigation-explainable/" rel="noopener noreferrer"&gt;RAG-powered AI that generates evidence&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Notes for triage&lt;/p&gt;

&lt;p&gt;This is not a model-accuracy bug. It’s a workflow bug.&lt;br&gt;
Measure it before fixing. A one-week log is enough: &lt;a href="https://www.pearltrees.com/technoxander47/item838554012" rel="noopener noreferrer"&gt;https://www.pearltrees.com/technoxander47/item838554012&lt;/a&gt; shows how to sort which tasks to fix first.&lt;br&gt;
Reproduced in a hypothetical day: &lt;a href="https://verificationofpayee.blogspot.com/2026/11/day-in-two-aml-workflows.html" rel="noopener noreferrer"&gt;https://verificationofpayee.blogspot.com/2026/11/day-in-two-aml-workflows.html&lt;/a&gt;.&lt;br&gt;
Context: &lt;a href="https://technoxander.com/aml-ai-investigation-throughput-scaling/" rel="noopener noreferrer"&gt;AML investigation capacity&lt;/a&gt; and &lt;a href="https://technoxander.com/organisational-memory-aml-ai-investigations/" rel="noopener noreferrer"&gt;organisational memory in AML&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Over to you&lt;/p&gt;

&lt;p&gt;How do you label tickets that cost time but never break anything?&lt;br&gt;
Would you ship a fix that moves work from analysts to a retrieval layer, even if accuracy needs monitoring?&lt;br&gt;
Hot take: the worst bugs are the ones with no error message. Agree, or fight me?&lt;/p&gt;

&lt;p&gt;From the team at &lt;a href="https://technoxander.com/" rel="noopener noreferrer"&gt;TechnoXander&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>productivity</category>
      <category>rag</category>
      <category>software</category>
    </item>
    <item>
      <title>Reviewing a Closed Case Like a Pull Request</title>
      <dc:creator>TechnoXander</dc:creator>
      <pubDate>Thu, 05 Nov 2026 06:49:22 +0000</pubDate>
      <link>https://dev.to/technoxander/reviewing-a-closed-case-like-a-pull-request-2dk3</link>
      <guid>https://dev.to/technoxander/reviewing-a-closed-case-like-a-pull-request-2dk3</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Flu6gsom2kdcyjpijm44d.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Flu6gsom2kdcyjpijm44d.png" alt=" " width="800" height="335"&gt;&lt;/a&gt;&lt;br&gt;
Engineers have a ritual for changes that matter: someone else reads the diff before it ships. Nobody does that for a closed AML case, which is odd, because a closed case becomes a reference for every similar one that follows.&lt;/p&gt;

&lt;p&gt;Let’s run the ritual anyway. The “PR” is case-4821. The commit message reads: “Closed: consistent with profile.”&lt;/p&gt;

&lt;p&gt;The review comments&lt;/p&gt;

&lt;p&gt;Why? The message states a verdict, not a reason. What changed my mind from “unusual” to “fine”?&lt;br&gt;
Evidence? Which documents or earlier cases supported it? I can’t see links.&lt;br&gt;
Scope? Does “consistent” cover this counterparty only, or the whole pattern?&lt;br&gt;
Expiry? Is this still true if the business model changes?&lt;br&gt;
Related PRs? Did this reference earlier cases, and are they linked?&lt;/p&gt;

&lt;p&gt;A commit-message standard for closures&lt;/p&gt;

&lt;p&gt;Treat it like conventional commits. A title with the verdict. A body with the evidence and what you ruled out. A footer with scope, date, review-by and links to related cases.&lt;/p&gt;

&lt;p&gt;That’s the shape of a good &lt;a href="https://technoxander.com/organisational-memory-aml-ai-investigations/" rel="noopener noreferrer"&gt;AML case memory&lt;/a&gt; entry, and it’s the data model behind retrieval you can trust, as we explain in &lt;a href="https://technoxander.com/rag-ai-aml-investigation-explainable/" rel="noopener noreferrer"&gt;RAG-powered AI that generates evidence&lt;/a&gt;. For where this sits in the stack, see &lt;a href="https://technoxander.com/aml-technology-stack-investigation-layer-missing/" rel="noopener noreferrer"&gt;the missing layer of AML technology&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;A diagram of when to reuse a past decision: &lt;a href="https://www.tumblr.com/vop25/832421899455905792/when-can-yesterdays-judgement-help-todays-alert?source=share" rel="noopener noreferrer"&gt;Diagram&lt;/a&gt;. A walk-through of how a reviewer samples cases: &lt;a href="https://verificationofpayee.blogspot.com/2026/11/reviewer-consistency-test-aml-case-memory.html" rel="noopener noreferrer"&gt;Blogspot walk through&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Over to you&lt;/p&gt;

&lt;p&gt;What would your commit message for “Closed: consistent with profile” say?&lt;br&gt;
Should a closure need a second reader, like a PR does, or is that overkill?&lt;br&gt;
Hot take: a closure with no stated scope is a bug. Agree, or fight me?&lt;/p&gt;

&lt;p&gt;Brought to you by &lt;a href="https://technoxander.com/" rel="noopener noreferrer"&gt;TechnoXander&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>fintech</category>
      <category>aml</category>
      <category>discuss</category>
      <category>ai</category>
    </item>
    <item>
      <title>Using RAG in AML: Why Retrieval Quality Matters More Than Similarity</title>
      <dc:creator>TechnoXander</dc:creator>
      <pubDate>Wed, 04 Nov 2026 13:52:41 +0000</pubDate>
      <link>https://dev.to/technoxander/using-rag-in-aml-why-retrieval-quality-matters-more-than-similarity-2e0o</link>
      <guid>https://dev.to/technoxander/using-rag-in-aml-why-retrieval-quality-matters-more-than-similarity-2e0o</guid>
      <description>&lt;p&gt;Retrieval-augmented generation (RAG) can help AML teams access relevant information from large collections of case records, policies and investigative material. But retrieving a similar document is not the same as retrieving useful evidence.&lt;/p&gt;

&lt;p&gt;This distinction matters in financial crime investigations. A system might find a previous case with comparable transaction activity, but that case may involve a different customer profile, payment purpose or set of circumstances. The &lt;a href="https://technoxander.com/rag-ai-aml-investigation-explainable/" rel="noopener noreferrer"&gt;use of RAG in explainable AML investigations must account for those differences&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;A practical RAG workflow should consider three questions:&lt;/p&gt;

&lt;p&gt;Is the source relevant? The system should identify why a previous record has been retrieved, rather than relying only on keyword overlap.&lt;/p&gt;

&lt;p&gt;Is the information reliable? The investigator should be able to inspect the source, its date and its limitations.&lt;/p&gt;

&lt;p&gt;Does it support the current assessment? Retrieved information should help frame further questions, not automatically determine the outcome.&lt;/p&gt;

&lt;p&gt;This is especially important when historical case records contain outdated assumptions or incomplete reasoning. A previous decision should not become a permanent instruction for future cases.&lt;/p&gt;

&lt;p&gt;The quality of the underlying records also matters. If firms want AI to retrieve meaningful investigative context, they need structured and consistently documented information. This connects directly with the wider challenge of &lt;a href="https://technoxander.com/organisational-memory-aml-ai-investigations/" rel="noopener noreferrer"&gt;organisational memory in AML investigations&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;RAG should therefore be evaluated on more than response speed. Firms should test whether it retrieves appropriate evidence, identifies uncertainty and allows investigators to verify the material used.&lt;/p&gt;

&lt;p&gt;These are important considerations when assessing an &lt;a href="https://technoxander.com/aml-investigation-platform/" rel="noopener noreferrer"&gt;AML investigation platform&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;AI can make information easier to access, but the quality of the investigation still depends on how that information is assessed. To discuss the practical application, visit &lt;a href="https://technoxander.com/contact/" rel="noopener noreferrer"&gt;TechnoXander&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>rag</category>
    </item>
    <item>
      <title>Dear Product Manager Writing the AML AI Requirements Doc</title>
      <dc:creator>TechnoXander</dc:creator>
      <pubDate>Fri, 30 Oct 2026 07:38:32 +0000</pubDate>
      <link>https://dev.to/technoxander/dear-product-manager-writing-the-aml-ai-requirements-doc-17c9</link>
      <guid>https://dev.to/technoxander/dear-product-manager-writing-the-aml-ai-requirements-doc-17c9</guid>
      <description>&lt;p&gt;Dear product manager,&lt;/p&gt;

&lt;p&gt;You’ve been asked to write requirements for “AI in AML”. It’s an easy brief to get wrong, because the phrase hides two different products.&lt;/p&gt;

&lt;p&gt;The first is detection AI. It sits upstream and decides which transactions get flagged. The second is investigation AI, which sits where an analyst opens the alert and decides what to do. The distinction is the core of our piece on the &lt;a href="https://technoxander.com/ai-aml-investigations-hype-vs-value/" rel="noopener noreferrer"&gt;two types of AI in AML&lt;/a&gt;. Most vendors only sell the first.&lt;/p&gt;

&lt;p&gt;Write requirements for the second one too&lt;/p&gt;

&lt;p&gt;First screen. The analyst sees customer history, prior dispositions and linked accounts without opening other tabs. That’s the gap behind &lt;a href="https://technoxander.com/aml-alerts-missing-context-financial-crime/" rel="noopener noreferrer"&gt;alerts that arrive without context&lt;/a&gt;.&lt;br&gt;
Traceability. Every sentence in a summary points to a record. Read why &lt;a href="https://technoxander.com/rag-ai-aml-investigation-explainable/" rel="noopener noreferrer"&gt;evidence-grounded AI beats answer-generating AI&lt;/a&gt;.&lt;br&gt;
Human decision. The tool recommends, a named person decides, and the log shows who.&lt;br&gt;
Memory. When a case closes, reasoning is kept in a form the next case can use.&lt;br&gt;
Data hygiene. FATF warns that poor or biased data fed into a model can be learned by it, so say how data quality is checked. &lt;br&gt;
sanctionscanner&lt;/p&gt;

&lt;p&gt;One requirement to resist&lt;/p&gt;

&lt;p&gt;Don’t write “reduce false positives” as the only success measure. It describes detection. If the goal is better decisions, measure what analysts see and how well they can explain the result. That is the argument of &lt;a href="https://technoxander.com/future-aml-investigation-beyond-detection/" rel="noopener noreferrer"&gt;the next phase of AML is better decisions&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;We work on the investigation layer at &lt;a href="https://technoxander.com/" rel="noopener noreferrer"&gt;TechnoXander&lt;/a&gt;, and you can see how at the &lt;a href="https://technoxander.com/aml-investigation-platform/" rel="noopener noreferrer"&gt;platform page&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Over to you&lt;/p&gt;

&lt;p&gt;What is the first line of your AI requirements doc, and does it mention the analyst?&lt;br&gt;
Should traceability be a hard requirement or a nice-to-have?&lt;br&gt;
Hot take: a requirement that can’t be tested in a demo isn’t a requirement. Agree, or fight me?&lt;/p&gt;

&lt;p&gt;Warm regards,&lt;br&gt;
Sonal&lt;/p&gt;

</description>
      <category>ai</category>
      <category>fintech</category>
      <category>discuss</category>
      <category>banking</category>
    </item>
    <item>
      <title>Dear Engineer Building the Alert Screen</title>
      <dc:creator>TechnoXander</dc:creator>
      <pubDate>Tue, 27 Oct 2026 07:04:24 +0000</pubDate>
      <link>https://dev.to/technoxander/dear-engineer-building-the-alert-screen-127i</link>
      <guid>https://dev.to/technoxander/dear-engineer-building-the-alert-screen-127i</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fp7m2kv9b96pap9x9mww0.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fp7m2kv9b96pap9x9mww0.png" alt=" " width="800" height="337"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Dear engineer,&lt;/p&gt;

&lt;p&gt;Somebody just walked over with a request that sounds simple: “Analysts need more context when an alert opens. Can you add it to the screen?”&lt;/p&gt;

&lt;p&gt;I’m writing because I’ve seen how this goes. It starts with good intentions and ends with a screen that scrolls for a page and a half. So here are a few things I wish someone told every team at the start.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;First, what decision does this screen support?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;An AML analyst opening an alert is trying to decide one thing: is this activity explained, or does it need to go further? Every item on the screen should help with that decision. If a field wouldn’t change what the analyst does, it’s decoration.&lt;/p&gt;

&lt;p&gt;Ask three analysts to cross out the fields they never use. You’ll be surprised how little survives.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Second, “nothing found” has two meanings&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;This is the one that bites developers hardest. Imagine the screen shows “Previous cases: 0”.&lt;/p&gt;

&lt;p&gt;Does that mean there are none? Or that the search couldn’t find them, because the earlier case was filed under a trading name and your query used the legal name?&lt;/p&gt;

&lt;p&gt;A human reading “0” will assume the first. So the screen quietly teaches analysts that no history exists, when it only means none was found.&lt;/p&gt;

&lt;p&gt;My suggestion: give every section three honest states: found, none found and could not check. Make “could not check” visible and boring, a small grey label. It takes ten minutes to build and prevents a lot of false comfort.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Third, show how sure you are&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Entity matching is messy. Two records might be the same customer, or they might not. Don’t merge them silently, and don’t hide them either. Show “possible match”, show the evidence behind it (same address, same phone number), and let the analyst confirm or reject.&lt;/p&gt;

&lt;p&gt;Their answer is data too. Store it, so the next analyst doesn’t redo the same check.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Fourth, order by consequence, not by database&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Put what changes decisions at the top: how this activity compares with the customer’s normal pattern, the outcome of the last case, any open questions. Keep raw records one click away.&lt;/p&gt;

&lt;p&gt;Think in three layers: the verdict frame, the supporting evidence, the drill-down. If a field lives in the top layer, it should earn its place.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Why I care about this&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;When context is hard to see, analysts do what anyone would. They open other tabs, search again, copy identifiers and rebuild a picture the institution already holds. Better detection doesn’t fix that. A better first screen does.&lt;/p&gt;

&lt;p&gt;I wrote about the compliance side of the problem here: &lt;a href="https://technoxander.com/aml-alerts-missing-context-financial-crime/" rel="noopener noreferrer"&gt;why alerts arrive without context.&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Over to you&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;I’d really like to hear how you’d approach this, especially if you’ve built something similar.&lt;/p&gt;

&lt;p&gt;Have you built an internal tool where “no data” was ambiguous? How did you handle it?&lt;br&gt;
If you could show only five things at the top of an investigation screen, what would they be?&lt;br&gt;
Hot take: a screen that shows less but is trusted beats a screen that shows everything. Agree, or fight me?&lt;/p&gt;

&lt;p&gt;Warm regards,&lt;br&gt;
Sonal&lt;/p&gt;

</description>
      <category>fintech</category>
      <category>ux</category>
      <category>compilance</category>
      <category>discuss</category>
    </item>
    <item>
      <title>Dear MLRO, Put Case Ownership Before Closure Counts</title>
      <dc:creator>TechnoXander</dc:creator>
      <pubDate>Mon, 19 Oct 2026 13:35:39 +0000</pubDate>
      <link>https://dev.to/technoxander/dear-mlro-put-case-ownership-before-closure-counts-47p5</link>
      <guid>https://dev.to/technoxander/dear-mlro-put-case-ownership-before-closure-counts-47p5</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F5rjjbubcrfvfcbwjsgx0.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F5rjjbubcrfvfcbwjsgx0.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;br&gt;
At 2.15 pm, an investigator opens a case that looks straightforward. The customer profile describes one type of business activity, but the available transaction history suggests another. A previous review contains a useful observation, buried in an older case file.&lt;/p&gt;

&lt;p&gt;Before the investigation can progress, the analyst must find the missing information, reconcile the discrepancies and establish what requires further review.&lt;/p&gt;

&lt;p&gt;None of that work is visible in a closure target.&lt;/p&gt;

&lt;h3&gt;
  
  
  Dear MLRO,
&lt;/h3&gt;

&lt;p&gt;Your investigators are expected to make decisions that withstand scrutiny. But the quality of those decisions depends on more than individual judgement. It also depends on how information is organised, how responsibility is assigned and how the investigation process supports the people doing the work.&lt;/p&gt;

&lt;p&gt;Clearer instructions and additional training have value. They cannot, however, solve problems caused by fragmented information, unclear ownership or inefficient escalation.&lt;/p&gt;

&lt;p&gt;The question is whether your operating model gives investigators the conditions they need to make sound decisions.&lt;/p&gt;

&lt;h3&gt;
  
  
  Give investigators the right information
&lt;/h3&gt;

&lt;p&gt;An alert is a reason to investigate, not a complete account of customer activity.&lt;/p&gt;

&lt;p&gt;Analysts may need transaction history, due diligence records, screening results and previous case decisions to assess what has happened. When these sources are disconnected, investigators spend time reconstructing the case before they can analyse it.&lt;/p&gt;

&lt;p&gt;This creates delays and increases the risk that relevant information will be overlooked or previous decisions misunderstood.&lt;/p&gt;

&lt;p&gt;Measure more than the time taken to close an alert. Consider how much effort is required to assemble the necessary evidence and whether the information available supports a defensible decision.&lt;/p&gt;

&lt;p&gt;The objective is not to provide more information indiscriminately. It is to make relevant evidence accessible and explain its significance and limitations.&lt;/p&gt;

&lt;h3&gt;
  
  
  Make ownership and escalation clear
&lt;/h3&gt;

&lt;p&gt;An open case should have a clear next action, an identified owner and an understandable reason for any delay.&lt;/p&gt;

&lt;p&gt;Responsibility may pass between investigators, supervisors and specialist teams. What matters is that the reasoning behind the current position travels with the case. The next person should not have to reconstruct what has already been examined or which conclusions remain provisional.&lt;/p&gt;

&lt;p&gt;Escalation also needs defined boundaries. Analysts should know which decisions they can make independently, when additional review is required and who has authority to decide.&lt;/p&gt;

&lt;p&gt;Without clear criteria, routine uncertainty can generate unnecessary referrals, while cases needing specialist input may remain with the wrong person.&lt;/p&gt;

&lt;h3&gt;
  
  
  Use quality assurance to improve the process
&lt;/h3&gt;

&lt;p&gt;When a case is returned for additional work, correcting the case is only the first step. You should also understand why the problem occurred.&lt;/p&gt;

&lt;p&gt;Was evidence difficult to find? Were internal requirements unclear? Did the recorded rationale fail to support the decision?&lt;/p&gt;

&lt;p&gt;Track recurring causes of rework and respond accordingly. Improve information access where evidence is missing, clarify standards where expectations differ, and use targeted coaching where analytical skills need development.&lt;/p&gt;

&lt;p&gt;Investigators remain accountable for their work. Management is equally responsible for the tools, standards and processes that support it.&lt;/p&gt;

&lt;h3&gt;
  
  
  Make technology reduce friction
&lt;/h3&gt;

&lt;p&gt;Technology should help investigators find relevant evidence, understand case history and identify unresolved questions without replacing professional judgement.&lt;/p&gt;

&lt;p&gt;Automation can organise records, prepare summaries and highlight relationships for review. Material statements should remain traceable to their sources, and conflicting or missing information should be clearly identified.&lt;/p&gt;

&lt;p&gt;The &lt;a href="https://technoxander.com/aml-investigation-platform/" rel="noopener noreferrer"&gt;TechnoXander AML investigation platform&lt;/a&gt; provides a reference point when evaluating whether a more structured investigation workflow could address these operational problems. Any solution should fit your firm's policies, data controls and oversight requirements.&lt;/p&gt;

&lt;h3&gt;
  
  
  The question I would ask your team
&lt;/h3&gt;

&lt;p&gt;Do your analysts spend most of their time assessing evidence, or finding it, reconciling it and explaining why information is missing?&lt;/p&gt;

&lt;p&gt;The answer should determine what you improve first. Fragmented information, delayed approvals, knowledge gaps and repeated case corrections require different interventions.&lt;/p&gt;

&lt;p&gt;Review one difficult investigation from the first alert to the final decision. Record every point where the investigator has to search, wait, reconstruct or seek clarification.&lt;/p&gt;

&lt;p&gt;Those findings will give you a practical starting point for improving investigation quality, rather than simply asking your team to close cases faster.&lt;/p&gt;

</description>
      <category>aml</category>
      <category>compilance</category>
      <category>regtech</category>
      <category>mlro</category>
    </item>
    <item>
      <title>Customer Context Is Part of Every AML Alert</title>
      <dc:creator>TechnoXander</dc:creator>
      <pubDate>Mon, 19 Oct 2026 13:24:57 +0000</pubDate>
      <link>https://dev.to/technoxander/customer-context-is-part-of-every-aml-alert-2gib</link>
      <guid>https://dev.to/technoxander/customer-context-is-part-of-every-aml-alert-2gib</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Faqkt001xzpa26a286w2a.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Faqkt001xzpa26a286w2a.png" alt=" " width="800" height="450"&gt;&lt;/a&gt;&lt;br&gt;
An alert arrives near the end of a shift. The amount is unusual, the payment reference is incomplete, and the customer's transaction history is difficult to explain.&lt;/p&gt;

&lt;p&gt;The analyst must decide whether to search other systems, request information from another team or make a preliminary assessment.&lt;/p&gt;

&lt;p&gt;This is where transaction data quality becomes a practical AML investigation problem. Missing information slows reviews, creates repeated requests and weakens the reasoning behind the final decision.&lt;/p&gt;

&lt;h2&gt;
  
  
  &lt;strong&gt;What goes wrong&lt;/strong&gt;
&lt;/h2&gt;

&lt;p&gt;The analyst sees a transaction but lacks enough context to interpret it. Counterparty details may be missing, expected customer behaviour may be unclear, or previous alerts may sit in another case-management system.&lt;/p&gt;

&lt;p&gt;Each gap forces the analyst to reconstruct the picture manually.&lt;/p&gt;

&lt;p&gt;Without consistent information, different analysts may reach different conclusions about similar activity. Reviewers may also struggle to understand why a case was closed if the supporting evidence was never recorded.&lt;/p&gt;

&lt;p&gt;An alert can therefore be completed without fully resolving the uncertainty that made it difficult.&lt;/p&gt;

&lt;h2&gt;
  
  
  &lt;strong&gt;Why it keeps happening&lt;/strong&gt;
&lt;/h2&gt;

&lt;p&gt;Customer profiles, payment records and previous investigation decisions often sit across systems with different purposes. Each may work correctly in isolation, yet investigators still have to assemble the full picture themselves.&lt;/p&gt;

&lt;p&gt;Ownership is another challenge. When data fails to match or transfer correctly, responsibility may be unclear. Analysts work around the problem to clear their queues, while the underlying defect remains.&lt;/p&gt;

&lt;p&gt;Operations reports may track alert volumes and closures without showing how much time investigators spend searching for basic information.&lt;/p&gt;

&lt;p&gt;The &lt;a href="https://handbook.fca.org.uk/handbook/fcg3" rel="noopener noreferrer"&gt;FCA's Financial Crime Guide&lt;/a&gt; highlights the importance of accurate monitoring data, understanding customer relationships and recognising monitoring-system limitations.&lt;/p&gt;

&lt;h2&gt;
  
  
  &lt;strong&gt;How to fix it&lt;/strong&gt;
&lt;/h2&gt;

&lt;p&gt;Start by identifying the minimum context analysts need to make sound decisions. Depending on the alert, this may include customer profiles, expected activity, transaction details, counterparty information and previous related activity.&lt;/p&gt;

&lt;p&gt;Then identify the cause of each data gap. Is the information not collected, incorrectly mapped, missing between systems or simply difficult to find? Each problem needs a different remedy.&lt;/p&gt;

&lt;p&gt;Assign ownership for material defects, assess their impact on investigations and retest the data after remediation. Confirm that the information reaches the analyst as expected, rather than relying on an engineering ticket being marked complete.&lt;/p&gt;

&lt;p&gt;One practical test is to ask an analyst to explain a recently closed alert using only the retained case record. If the reasoning depends on memory, private notes or undocumented conversations, the record is incomplete.&lt;/p&gt;

&lt;p&gt;An effective &lt;a href="https://technoxander.com/aml-investigation-platform/" rel="noopener noreferrer"&gt;AML investigation workflow&lt;/a&gt; should help analysts connect information, assess anomalies and preserve the reasoning behind decisions.&lt;/p&gt;

&lt;p&gt;For your next team review, choose one difficult alert and identify every point at which the analyst had to leave the case file to find essential information.&lt;/p&gt;

&lt;p&gt;Fixing that friction may improve investigative quality more than simply asking the team to work faster.&lt;/p&gt;

</description>
      <category>aml</category>
      <category>fincrime</category>
      <category>uk</category>
      <category>regtech</category>
    </item>
  </channel>
</rss>
