<?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: Mohamed Ibrahim </title>
    <description>The latest articles on DEV Community by Mohamed Ibrahim  (@rafmo20d).</description>
    <link>https://dev.to/rafmo20d</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%2F1135540%2F9ba3545f-f7b4-4993-a8a5-65b83397d095.jpeg</url>
      <title>DEV Community: Mohamed Ibrahim </title>
      <link>https://dev.to/rafmo20d</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/rafmo20d"/>
    <language>en</language>
    <item>
      <title>Why our finance AI is not allowed to produce a single number</title>
      <dc:creator>Mohamed Ibrahim </dc:creator>
      <pubDate>Thu, 19 Nov 2026 23:12:16 +0000</pubDate>
      <link>https://dev.to/rafmo20d/why-our-finance-ai-is-not-allowed-to-produce-a-single-number-395j</link>
      <guid>https://dev.to/rafmo20d/why-our-finance-ai-is-not-allowed-to-produce-a-single-number-395j</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%2Ff90c1o9urdzheqseoapp.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%2Ff90c1o9urdzheqseoapp.png" alt=" " width="800" height="479"&gt;&lt;/a&gt;---&lt;br&gt;
title: "Why our finance AI is not allowed to produce a single number"&lt;br&gt;
published: false&lt;br&gt;
description: "How we built OfinTech Desk so every figure traces back to its source, with deterministic math, a CPU-only classifier and an LLM that only explains."&lt;/p&gt;

&lt;h2&gt;
  
  
  tags: opensource, ai, fintech, python
&lt;/h2&gt;

&lt;p&gt;Every finance team I have worked with asks the same question about every report: where did this number come from?&lt;/p&gt;

&lt;p&gt;When AI enters the picture, that question gets harder to answer, not easier. A language model can write a fluent month-end summary, but if it also produced the figures in it, nobody can say where they came from. In finance, a number you cannot trace is a number you cannot use.&lt;/p&gt;

&lt;p&gt;So when we built OfinTech Desk, we started from one rule: &lt;strong&gt;every figure must have a source&lt;/strong&gt;. This post explains the architecture that rule forced on us.&lt;/p&gt;

&lt;h2&gt;
  
  
  The rule, in practice
&lt;/h2&gt;

&lt;p&gt;In OfinTech Desk you can click any number and open the chain behind it:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The computation that produced it&lt;/li&gt;
&lt;li&gt;The ledger rows that fed that computation&lt;/li&gt;
&lt;li&gt;The source document&lt;/li&gt;
&lt;li&gt;The exact page in that document&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;If a figure cannot show that chain, it does not appear.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three layers, three jobs
&lt;/h2&gt;

&lt;p&gt;The key decision was to split the work into three layers and to be strict about what each one is allowed to do.&lt;/p&gt;

&lt;h3&gt;
  
  
  1. Deterministic code does the math
&lt;/h3&gt;

&lt;p&gt;All arithmetic happens in plain, deterministic code: statements, month-end close, KPIs, VAT return figures. Same input, same answer, every time. There is no sampling, no temperature and no "usually right".&lt;/p&gt;

&lt;p&gt;This also makes the system testable in the normal way. A KPI is a function, and a function has unit tests.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. A classifier makes closed decisions
&lt;/h3&gt;

&lt;p&gt;Some steps are not arithmetic but judgement, for example matching a bank line to a payment provider payout. These are closed decisions: the answer comes from a fixed set of options.&lt;/p&gt;

&lt;p&gt;For those we use a small trained classifier we call the Ofin model. It returns a choice together with a measured probability, so the system can tell a confident match from a doubtful one, and a person can set the threshold above which matches are accepted automatically.&lt;/p&gt;

&lt;p&gt;It learns from the team's own accept and reject decisions, so the second month is more accurate than the first.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. The LLM only explains
&lt;/h3&gt;

&lt;p&gt;The language model sits at the end of the pipeline. It reads the figures and their sources and explains them in plain language, in Arabic or English. It never invents a number.&lt;/p&gt;

&lt;p&gt;Because the model is not trusted with the math, the choice of provider matters much less. You can use any LLM provider, a local model, or no model at all, and every figure stays the same.&lt;/p&gt;

&lt;h2&gt;
  
  
  No GPU required
&lt;/h2&gt;

&lt;p&gt;We wanted this to run where finance data already lives, often on a modest server inside the company. So the classifier is deliberately small.&lt;/p&gt;

&lt;p&gt;On a 2-core CPU it labels a bank line in about &lt;strong&gt;0.03 ms&lt;/strong&gt;. A 500-page bank statement takes under a second, and retraining on a team's own decisions takes about 0.3 seconds. No GPU anywhere.&lt;/p&gt;

&lt;h2&gt;
  
  
  Read-only by design
&lt;/h2&gt;

&lt;p&gt;OfinTech Desk does not replace your accounting system. It connects read-only to QuickBooks Online, Xero, Zoho Books and Odoo, or imports a general ledger from Excel or CSV, and sits beside the system you already run.&lt;/p&gt;

&lt;p&gt;Read-only is a feature, not a limitation. An assistant that cannot write to your books or move money is one a risk officer can actually approve. Everything runs on your own servers, so the data never leaves your network.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I learned
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Decide what the model is not allowed to do.&lt;/strong&gt; The most useful design decision was a restriction, not a capability.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Measured confidence beats fluent confidence.&lt;/strong&gt; A probability you can set a threshold on is worth more than a well-written paragraph.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Small models go a long way.&lt;/strong&gt; For closed decisions, a tiny CPU model was faster, cheaper and easier to audit than a large one.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;p&gt;The core is free and open source under Apache-2.0. There is a paid tier for payment settlement reconciliation, which is how we fund the project.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;GitHub: &lt;a href="https://github.com/OfinTech/ofinTech-Dask" rel="noopener noreferrer"&gt;https://github.com/OfinTech/ofinTech-Dask&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Walkthrough video: &lt;a href="https://youtu.be/kAX4ZPZ1NFk" rel="noopener noreferrer"&gt;https://youtu.be/kAX4ZPZ1NFk&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Website: &lt;a href="https://ofintech.ai" rel="noopener noreferrer"&gt;https://ofintech.ai&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Contributions are welcome, and I would love feedback on the provenance model, especially on where you think it breaks down.&lt;/p&gt;

</description>
      <category>opensource</category>
      <category>ai</category>
      <category>fintech</category>
      <category>python</category>
    </item>
    <item>
      <title>OfinTech Desk: an open-source AI workspace for finance where every number has a source</title>
      <dc:creator>Mohamed Ibrahim </dc:creator>
      <pubDate>Wed, 28 Oct 2026 01:14:37 +0000</pubDate>
      <link>https://dev.to/rafmo20d/ofintech-desk-an-open-source-ai-workspace-for-finance-where-every-number-has-a-source-4066</link>
      <guid>https://dev.to/rafmo20d/ofintech-desk-an-open-source-ai-workspace-for-finance-where-every-number-has-a-source-4066</guid>
      <description>&lt;p&gt;&lt;a href="https://www.youtube.com/watch?v=jfSBOAi2N0U" rel="noopener noreferrer"&gt;&lt;/a&gt;Most "AI for finance" demos have the same flaw: the model writes a confident paragraph full of numbers, and nobody can tell which of them came from the ledger and which the model made up.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;&lt;a href="https://github.com/OfinTech/ofinTech-Dask" rel="noopener noreferrer"&gt;OfinTech Desk&lt;/a&gt;&lt;/strong&gt; is our answer to that, and as of &lt;strong&gt;v1.0.0 (27 Oct 2026)&lt;/strong&gt; it is fully open source under &lt;strong&gt;Apache-2.0&lt;/strong&gt;. It's a desktop app (plus an optional self-hosted Team Server) that reads your ledger, spreadsheets, documents and inbox, does the arithmetic &lt;em&gt;in code&lt;/em&gt;, fills in your own Excel and Word files with a source for every figure, and asks a human before it sends or changes anything.&lt;/p&gt;

&lt;h2&gt;
  
  
  This post is not a feature tour. It's a walk through the engineering decisions inside the engine, the places where we think the design is interesting, and the places where we'd love your help.
&lt;/h2&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;TL;DR for the impatient&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Python engine (Starlette + SQLite), React/TypeScript UI, Electron shell. ~29k lines.&lt;/li&gt;
&lt;li&gt;Bring any model: OpenAI, Anthropic, Gemini, DeepSeek, Mistral, Groq, OpenRouter, xAI, Azure, or local via Ollama / LM Studio.&lt;/li&gt;
&lt;li&gt;Our own agent loop (no framework) so runs can pause for approval and resume after a restart.&lt;/li&gt;
&lt;li&gt;A &lt;em&gt;figure check&lt;/em&gt; flags every number in an answer that doesn't appear in a tool result.&lt;/li&gt;
&lt;li&gt;A tiny formula language, parsed (no &lt;code&gt;eval&lt;/code&gt;), compiled to Excel, and evaluated with Excel's own rules — verified cell-by-cell against LibreOffice.&lt;/li&gt;
&lt;li&gt;Tool selection per request: &lt;strong&gt;54% smaller requests&lt;/strong&gt; on a simple question.&lt;/li&gt;
&lt;li&gt;99 engine tests, a browser smoke test, UI in English, Arabic (RTL), Spanish and French.&lt;/li&gt;
&lt;/ul&gt;
&lt;/blockquote&gt;




&lt;h2&gt;
  
  
  The architecture in one picture
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt; Desktop app (Electron)                               Team Server (Docker)
 ┌───────────────────────────────┐                    ┌──────────────────────────────┐
 │ Renderer: React UI (sandboxed)│                    │ Caddy (HTTPS) ─► engine       │
 │   ▲ token via preload bridge  │   or, in team mode │   server mode, accounts,      │
 │ Main: starts engine, keeps    │ ─────────────────► │   invitations, webhooks       │
 │   master key in OS storage    │   loads server UI  │   (WhatsApp, Telegram)        │
 │ Engine (Python) on 127.0.0.1  │                    └──────────────────────────────┘
 └───────────────────────────────┘
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;One engine, two modes.&lt;/strong&gt; The same Python engine runs as a single-user sidecar on loopback inside the desktop app, or as a multi-user server behind Caddy. In desktop mode it creates a master key, stores it in the OS keychain (Keychain / DPAPI / libsecret), generates a fresh access token on every launch, and rejects foreign &lt;code&gt;Host&lt;/code&gt; headers to defeat DNS rebinding.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;SQLite everywhere.&lt;/strong&gt; WAL mode, one connection per thread, serialised writes, FTS5 for search over documents, conversations and memory. It needs zero administration on a laptop and is plenty for a team-sized server. (PostgreSQL is on the roadmap — see "Where you can help".)&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Few, permissive dependencies.&lt;/strong&gt; Starlette, uvicorn, pydantic, httpx, mcp, openpyxl, cryptography, PyYAML. Charts, the Markdown renderer and the workflow editor are written in the project, so there's no runtime CDN or font download.&lt;/p&gt;




&lt;h2&gt;
  
  
  1. Why we wrote our own agent loop
&lt;/h2&gt;

&lt;p&gt;We looked at agent frameworks and kept bumping into one requirement: &lt;strong&gt;a run must be able to stop in the middle, wait hours for a human, and resume — even after the process restarts.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;So the loop in &lt;code&gt;engine/desk_engine/agent/harness.py&lt;/code&gt; is deliberately boring:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;model call → tool calls → results → model call … until the model answers (max 16 steps)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Every tool declares an &lt;code&gt;effect&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="nc"&gt;Effect&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;Literal&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"read"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"write"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;

&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;needs_approval&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nb"&gt;self&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;-&amp;gt;&lt;/span&gt; &lt;span class="kt"&gt;bool&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="nb"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;policy&lt;/span&gt; &lt;span class="k"&gt;is&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="kc"&gt;None&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;          &lt;span class="c1"&gt;# MCP tools: auto | ask | off&lt;/span&gt;
        &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nb"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;policy&lt;/span&gt; &lt;span class="o"&gt;!=&lt;/span&gt; &lt;span class="s2"&gt;"auto"&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nb"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;effect&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="s2"&gt;"write"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;read&lt;/code&gt; tools (calculators, queries, document search) run immediately.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;write&lt;/code&gt; tools (send an e-mail, post a Telegram message, run a workflow) don't run at all. The loop records an approval with the &lt;strong&gt;exact arguments&lt;/strong&gt;, persists the pending calls into the run's &lt;code&gt;state&lt;/code&gt;, sets the run to &lt;code&gt;waiting_approval&lt;/code&gt;, and returns.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;When an editor approves or rejects in the inbox, &lt;code&gt;_resolve_pending&lt;/code&gt; replays the stored calls. A rejection is fed back to the model as a tool result — &lt;em&gt;"The user rejected this action. Do not retry it unless they ask; offer an alternative."&lt;/em&gt; — so the conversation keeps going instead of dead-ending.&lt;/p&gt;

&lt;p&gt;Because state lives in SQLite rather than in a Python coroutine, a restart is just another resume. The loop speaks two wire protocols — OpenAI Chat Completions and Anthropic Messages — which together cover every provider we list.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The model is treated as untrusted input.&lt;/strong&gt; It never gets SQL, a shell, or code execution. Its output is rendered from Markdown into React elements, never as HTML. Tools return text for the model &lt;em&gt;plus&lt;/em&gt; typed blocks (&lt;code&gt;table&lt;/code&gt;, &lt;code&gt;kpis&lt;/code&gt;, &lt;code&gt;link&lt;/code&gt;, &lt;code&gt;skill&lt;/code&gt;) that the UI draws itself.&lt;/p&gt;




&lt;h2&gt;
  
  
  2. The figure check: catching numbers that came from nowhere
&lt;/h2&gt;

&lt;p&gt;After the final answer, the harness collects every tool result, every user message and every consulted memory as &lt;strong&gt;evidence&lt;/strong&gt;, then extracts all numbers from the answer and checks each one against it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;unverified_figures&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;answer&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;str&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;evidence&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;list&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="kt"&gt;str&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;-&amp;gt;&lt;/span&gt; &lt;span class="kt"&gt;list&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="kt"&gt;str&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;known&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;v&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;text&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;evidence&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;_&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;v&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="nf"&gt;_numbers&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
    &lt;span class="n"&gt;flagged&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
    &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;raw&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;v&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="nf"&gt;_numbers&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;answer&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="n"&gt;a&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nb"&gt;abs&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;v&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;a&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&lt;/span&gt; &lt;span class="mi"&gt;100&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="s2"&gt;"."&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;raw&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="s2"&gt;"%"&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;raw&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
            &lt;span class="k"&gt;continue&lt;/span&gt;  &lt;span class="c1"&gt;# small counts, list numbers, days&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="mi"&gt;1900&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;=&lt;/span&gt; &lt;span class="n"&gt;a&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;=&lt;/span&gt; &lt;span class="mi"&gt;2100&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="s2"&gt;"."&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;raw&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="s2"&gt;","&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;raw&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
            &lt;span class="k"&gt;continue&lt;/span&gt;  &lt;span class="c1"&gt;# years&lt;/span&gt;
        &lt;span class="n"&gt;ok&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nb"&gt;any&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nb"&gt;abs&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nb"&gt;abs&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;cand&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;a&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;=&lt;/span&gt; &lt;span class="nb"&gt;max&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nb"&gt;abs&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;cand&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mf"&gt;0.005&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mf"&gt;0.011&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
                 &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;k&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;known&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;cand&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;k&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;k&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;100&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;k&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;100&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="n"&gt;ok&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="n"&gt;raw&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;flagged&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
            &lt;span class="n"&gt;flagged&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;append&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;raw&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;flagged&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;20&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A few details that took iteration:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The parser understands &lt;code&gt;1,234.5&lt;/code&gt;, &lt;code&gt;(1,200)&lt;/code&gt; as negative, the Unicode minus &lt;code&gt;−&lt;/code&gt;, and suffixes like &lt;code&gt;k&lt;/code&gt;, &lt;code&gt;m&lt;/code&gt;, &lt;code&gt;bn&lt;/code&gt;, &lt;code&gt;million&lt;/code&gt;, &lt;code&gt;%&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;k * 100&lt;/code&gt; and &lt;code&gt;k / 100&lt;/code&gt; let &lt;code&gt;0.153&lt;/code&gt; in a tool result verify &lt;code&gt;15.3%&lt;/code&gt; in the answer.&lt;/li&gt;
&lt;li&gt;A 0.5% relative tolerance lets the model round &lt;code&gt;1,234,567&lt;/code&gt; to &lt;code&gt;1.23m&lt;/code&gt; without being flagged.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;It's a heuristic, not a proof — and we say so in the UI. But it turns "trust me" into a short list a reviewer can actually check. &lt;strong&gt;Ideas for making it smarter (units, currencies, derived figures) are very welcome.&lt;/strong&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  3. A formula language that agrees with Excel
&lt;/h2&gt;

&lt;p&gt;When the assistant builds a workbook, computed columns and totals aren't pasted values — they're &lt;strong&gt;live Excel formulas&lt;/strong&gt;. That raises a hard problem: the value we show in the app must equal what Excel computes when you open the file.&lt;/p&gt;

&lt;p&gt;So &lt;code&gt;engine/desk_engine/files/expr.py&lt;/code&gt; implements a tiny language:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;variance            = actual - budget
variance_pct        = (actual - budget) / budget
gross margin        = [pl:Gross profit:amount] / [pl:Revenue:amount]
same, dotted form   = pl.total.amount
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Only &lt;code&gt;ABS, ROUND, MIN, MAX, SUM, AVERAGE&lt;/code&gt;, the four operators and parentheses. Each expression is tokenised and parsed into a small AST of frozen dataclasses (&lt;code&gt;Num&lt;/code&gt;, &lt;code&gt;Col&lt;/code&gt;, &lt;code&gt;Ref&lt;/code&gt;, &lt;code&gt;Neg&lt;/code&gt;, &lt;code&gt;Bin&lt;/code&gt;, &lt;code&gt;Call&lt;/code&gt;, &lt;code&gt;Span&lt;/code&gt;). That single AST is then either:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;compiled&lt;/strong&gt; to an Excel formula with real cell references, or&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;evaluated&lt;/strong&gt; in Python with Excel's semantics: blank as zero, division by zero shown blank, and — the fun one — rounding.
&lt;/li&gt;
&lt;/ol&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="c1"&gt;# ROUND(7250.5 * 1.15, 2) is 8338.08 as in Excel, not 8338.07.&lt;/span&gt;
&lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="kt"&gt;float&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;Decimal&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;f&lt;/span&gt;&lt;span class="s2"&gt;"{x:.15g}"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;quantize&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;q&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;rounding&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="no"&gt;ROUND_HALF_UP&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Excel rounds half away from zero at 15 significant digits; naive float rounding disagrees on values like this one. No &lt;code&gt;eval&lt;/code&gt; anywhere.&lt;/p&gt;

&lt;p&gt;The test suite then does something we're quite proud of: &lt;strong&gt;every formula cell in a generated workbook is recalculated by LibreOffice and compared with the engine's value.&lt;/strong&gt; If the two disagree, the build fails.&lt;/p&gt;

&lt;p&gt;The same "resolve once" idea applies to the whole file: a definition is resolved into a model where every cell has value, number format, live formula, readable formula, source and address. The Excel writer, the Word writer, the in-app viewer and the reviewer all read that one model, so they &lt;em&gt;can't&lt;/em&gt; disagree.&lt;/p&gt;

&lt;h3&gt;
  
  
  Maker and checker
&lt;/h3&gt;

&lt;p&gt;Every generated file is annotated by an &lt;strong&gt;independent reviewer&lt;/strong&gt;: a separate model call with a fresh context containing only the file, its sources and read-only tools. It never sees the drafting conversation, so it can't be talked into agreeing with it. Findings land as notes on specific cells and a Review sheet.&lt;/p&gt;




&lt;h2&gt;
  
  
  4. Shipping fewer tokens: tool selection and cache-friendly prompts
&lt;/h2&gt;

&lt;p&gt;An agent re-sends its whole context on every step, and tool schemas are the biggest chunk (~32,000 characters for all built-in tools). &lt;code&gt;agent/context.py&lt;/code&gt; keeps requests small:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="no"&gt;CORE&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="s2"&gt;"finance"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"knowledge"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"memory"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"skills"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"meta"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"mcp"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;

&lt;span class="no"&gt;TRIGGERS&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="s2"&gt;"files"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"workbook"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"excel"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"xlsx"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; …&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"مصنف"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"تقرير"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; …&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"libro"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"informe"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; …&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"classeur"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; …&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="s2"&gt;"accounting"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"invoice"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"ledger"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"p&amp;amp;l"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; …&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"فاتورة"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"دفتر"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; …&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    …
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ul&gt;
&lt;li&gt;Every request carries the core groups, plus groups whose trigger words appear in the last three user messages (in &lt;strong&gt;four languages&lt;/strong&gt;), plus groups already used in the thread.&lt;/li&gt;
&lt;li&gt;If the model needs something that wasn't sent, it calls &lt;code&gt;more_tools&lt;/code&gt; and gets the group on the next step. A tool that wasn't sent still runs if called by name — &lt;strong&gt;selection never blocks work&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Tools are sorted by &lt;code&gt;(category, name)&lt;/code&gt;, and the system prompt is split into a &lt;strong&gt;stable prefix&lt;/strong&gt; and a small &lt;strong&gt;changing part&lt;/strong&gt; (date, language, memories for this message), so provider prompt caches stay valid. Anthropic cache breakpoints are set after the tools, after the stable system block, and on the latest message.&lt;/li&gt;
&lt;li&gt;Tool results older than two turns are cut to 600 characters &lt;em&gt;in what is resent&lt;/em&gt; — never in what is stored, so they still count as evidence for the figure check.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Result, reproduced by &lt;code&gt;tests/test_tokens.py&lt;/code&gt;: a simple calculation question went from &lt;strong&gt;32,783 to 14,973 characters (54% smaller), 27 tools to 16.&lt;/strong&gt; Set &lt;code&gt;DESK_TOOL_SELECTION=off&lt;/code&gt; if a model misbehaves.&lt;/p&gt;




&lt;h2&gt;
  
  
  5. Memory you can audit
&lt;/h2&gt;

&lt;p&gt;Long-term memory is where assistants get creepy, so ours is conservative:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The assistant only &lt;strong&gt;proposes&lt;/strong&gt; a fact (a materiality threshold, an account mapping, a close deadline). Nothing is saved until a person approves it.&lt;/li&gt;
&lt;li&gt;Every entry shows its origin, who saved it and how often it was used; answers list which memories they consulted.&lt;/li&gt;
&lt;li&gt;Secrets are refused by pattern: passwords (in four languages), API keys, IBANs, full account numbers, and card numbers validated with a Luhn check.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  6. Extending it
&lt;/h2&gt;

&lt;p&gt;There are four well-defined extension points, and each is a good first contribution:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Skills&lt;/strong&gt; — plain &lt;code&gt;SKILL.md&lt;/code&gt; files with YAML front matter. They're instructions only; nothing in them is executed. Fifteen built-in finance skills ship today (month-end close, bank reconciliation, 13-week cash forecast, IFRS 15 / ASC 606…). If you know a finance process well, you can contribute without writing Python.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The Open Accounting Connector&lt;/strong&gt; — QuickBooks, Xero, Zoho Books and Odoo are built in. Any other ledger plugs in through an OpenAPI 3.1 contract with five endpoints, in pull mode (a reference adapter in &lt;code&gt;examples/accounting-adapter/adapter.py&lt;/code&gt;, standard library only) or push mode (an ingest API with a hashed key).&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nv"&gt;ADAPTER_KEY&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;change-me python examples/accounting-adapter/adapter.py
curl -H &lt;span class="s2"&gt;"Authorization: Bearer change-me"&lt;/span&gt; \
  &lt;span class="s2"&gt;"http://127.0.0.1:8787/v1/profit-and-loss?start=2026-01-01&amp;amp;end=2026-03-31"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;MCP, both directions&lt;/strong&gt; — connect external MCP servers as agent tools with an &lt;code&gt;ask/auto/off&lt;/code&gt; policy per tool (and &lt;code&gt;auto&lt;/code&gt; is downgraded for write tools), or expose the workspace as a read-only MCP server at &lt;code&gt;/mcp&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Languages&lt;/strong&gt; — UI strings are written in English inside &lt;code&gt;tr("…")&lt;/code&gt; and translated in &lt;code&gt;ui/src/i18n/{ar,es,fr}.ts&lt;/code&gt;. &lt;code&gt;npm run i18n:check&lt;/code&gt; reports coverage (currently 899/899). Arabic flips the layout with logical CSS properties.&lt;/p&gt;




&lt;h2&gt;
  
  
  Run it in five minutes
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/OfinTech/ofinTech-Dask.git &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="nb"&gt;cd&lt;/span&gt; ofinTech-Dask
pip install -r engine/requirements.txt
&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="nb"&gt;cd&lt;/span&gt; ui &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; npm install &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; node build.mjs&lt;span class="o"&gt;)&lt;/span&gt;

&lt;span class="c"&gt;# engine + UI in a browser&lt;/span&gt;
&lt;span class="nb"&gt;cd&lt;/span&gt; engine
&lt;span class="nv"&gt;DESK_DEV&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;1 &lt;span class="nv"&gt;DESK_UI_DIR&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;../ui/dist &lt;span class="nv"&gt;DESK_LOCAL_TOKEN&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nv"&gt;$(&lt;/span&gt;python -c &lt;span class="s2"&gt;"print('x'*40)"&lt;/span&gt;&lt;span class="o"&gt;)&lt;/span&gt; python -m desk_engine
&lt;span class="c"&gt;# open http://localhost:47821/#token=xxxxxxxx… (same token)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Point it at a local model through Ollama if you'd rather not use an API key. For the full desktop app: &lt;code&gt;(cd desktop &amp;amp;&amp;amp; npm install &amp;amp;&amp;amp; npm start)&lt;/code&gt;. For a team: &lt;code&gt;cd server &amp;amp;&amp;amp; cp .env.example .env &amp;amp;&amp;amp; docker compose up -d --build&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;Tests:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;cd&lt;/span&gt; engine &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; python -m unittest discover -s tests -t .   &lt;span class="c"&gt;# 99 tests, ~11 s&lt;/span&gt;
&lt;span class="nb"&gt;cd&lt;/span&gt; ui &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; npx tsc --noEmit &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; npm run i18n:check
python e2e/ui_smoke.py                                   &lt;span class="c"&gt;# real UI in Chromium + a fake model&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;External services are replaced by an HTTP mock transport shaped like each provider's documented responses; the MCP tests run a real MCP server process.&lt;/p&gt;




&lt;h2&gt;
  
  
  What we don't claim
&lt;/h2&gt;

&lt;p&gt;We think honesty about limits is part of open source, so the README has a section for it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Tax and accounting output is general guidance. &lt;strong&gt;No&lt;/strong&gt; certification or compliance is claimed (ZATCA, PDPL, SOC 2, SAMA or anyone else).&lt;/li&gt;
&lt;li&gt;Accounting connectors are &lt;strong&gt;read-only&lt;/strong&gt;. There is no live bank feed and no live FX feed.&lt;/li&gt;
&lt;li&gt;The reviewer is a second model check, not an audit.&lt;/li&gt;
&lt;li&gt;Connectors were tested against mocks of documented APIs, not live accounts.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Where you can help
&lt;/h2&gt;

&lt;p&gt;Here's what's on the roadmap and where a contributor could make a real dent:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;🐘 &lt;strong&gt;PostgreSQL&lt;/strong&gt; support for larger Team Servers.&lt;/li&gt;
&lt;li&gt;🔌 &lt;strong&gt;New connectors&lt;/strong&gt;: Sage, NetSuite, FreshBooks, MYOB — or an adapter for &lt;em&gt;your&lt;/em&gt; ERP via the Open Accounting Connector.&lt;/li&gt;
&lt;li&gt;✍️ &lt;strong&gt;Write-back&lt;/strong&gt; to accounting systems (draft bills and journals, behind approvals).&lt;/li&gt;
&lt;li&gt;🔐 &lt;strong&gt;SAML SSO and TOTP.&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;📊 Keeping charts and images when editing a user's own workbook; scheduled dashboard e-mails and PDF export.&lt;/li&gt;
&lt;li&gt;🌍 &lt;strong&gt;Translations&lt;/strong&gt;: native-speaker review of Arabic, Spanish and French, or a new language.&lt;/li&gt;
&lt;li&gt;🧪 &lt;strong&gt;Live testing&lt;/strong&gt;: try the connectors against a QuickBooks sandbox, a Xero demo company or an Odoo trial and tell us what breaks.&lt;/li&gt;
&lt;li&gt;📚 &lt;strong&gt;Skills&lt;/strong&gt;: write a &lt;code&gt;SKILL.md&lt;/code&gt; for a finance process you know.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The contribution rules are short and opinionated, because the project handles other people's money: no secrets or real data in fixtures, no vendor lock-in, the model is untrusted, no compliance claims, and &lt;strong&gt;every figure stays traceable&lt;/strong&gt;. Changes touching auth, permissions, encryption, approvals or agent tools get a second reviewer.&lt;/p&gt;




&lt;p&gt;⭐ &lt;strong&gt;Repo:&lt;/strong&gt; &lt;a href="https://github.com/OfinTech/OfinTech-ofinTech-Dask.git" rel="noopener noreferrer"&gt;https://github.com/OfinTech/OfinTech-ofinTech-Dask.git&lt;/a&gt;&lt;br&gt;
📄 &lt;strong&gt;Docs:&lt;/strong&gt; &lt;code&gt;docs/ARCHITECTURE.md&lt;/code&gt;, &lt;code&gt;docs/SECURITY.md&lt;/code&gt;, &lt;code&gt;docs/COSTS.md&lt;/code&gt;, &lt;code&gt;docs/FILES.md&lt;/code&gt;&lt;br&gt;
🤝 &lt;strong&gt;Contributing:&lt;/strong&gt; &lt;code&gt;CONTRIBUTING.md&lt;/code&gt; · &lt;strong&gt;Security reports:&lt;/strong&gt; &lt;code&gt;SECURITY.md&lt;/code&gt;&lt;/p&gt;

&lt;p&gt;If you've built agents that touch money, I'd especially like to hear how you handle the "where did this number come from?" problem. Drop a comment, open an issue, or send a PR.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>python</category>
      <category>agents</category>
      <category>deepseek</category>
    </item>
    <item>
      <title>🔐 Excited to share a demo of our Encryption-as-a-Service (EaaS)</title>
      <dc:creator>Mohamed Ibrahim </dc:creator>
      <pubDate>Tue, 02 Jan 2024 21:15:39 +0000</pubDate>
      <link>https://dev.to/rafmo20d/excited-to-share-a-demo-of-our-encryption-as-a-service-eaas-22fp</link>
      <guid>https://dev.to/rafmo20d/excited-to-share-a-demo-of-our-encryption-as-a-service-eaas-22fp</guid>
      <description>&lt;p&gt;🔐 Excited to share a demo of our Encryption-as-a-Service (EaaS) using HashiCorp Vault! 🚀 Check out the GitHub repository for all the details: Encryption-as-a-Service Vault Demo&lt;/p&gt;

&lt;p&gt;🤔 What's in the Demo?&lt;/p&gt;

&lt;p&gt;Go-based Encryption-as-a-Service application.&lt;br&gt;
HashiCorp Vault's Transit Secrets Engine for secure and scalable encryption.&lt;br&gt;
Simple RESTful API for encrypting and decrypting data.&lt;br&gt;
🚀 Key Features:&lt;/p&gt;

&lt;p&gt;Easy setup with clear step-by-step instructions.&lt;br&gt;
Seamless integration with HashiCorp Vault for robust security.&lt;br&gt;
Go application showcasing best practices for EaaS implementation.&lt;br&gt;
🔧 How to Get Started:&lt;/p&gt;

&lt;p&gt;Clone the repository: git clone &lt;a href="https://github.com/RafMo20D/encryption-as-a-service-vault-demo.git"&gt;https://github.com/RafMo20D/encryption-as-a-service-vault-demo.git&lt;/a&gt;&lt;br&gt;
Follow the setup steps in the README.&lt;br&gt;
Run the Go application and test the encryption and decryption endpoints.&lt;br&gt;
👩‍💻 Why EaaS?&lt;/p&gt;

&lt;p&gt;Centralized encryption management.&lt;br&gt;
Enhanced security with Vault's key management.&lt;br&gt;
Scalable and efficient for diverse application needs.&lt;br&gt;
🚨 Troubleshooting Tips:&lt;/p&gt;

&lt;p&gt;Ensure your Vault server is properly configured.&lt;br&gt;
Check Vault token permissions for encryption operations.&lt;br&gt;
Join the community discussions for additional support.&lt;br&gt;
👏 Contributions Welcome:&lt;/p&gt;

&lt;p&gt;Open issues, provide feedback, or submit pull requests.&lt;br&gt;
Let's build a robust Encryption-as-a-Service solution together!&lt;br&gt;
📚 Learn more and contribute: GitHub Repository&lt;/p&gt;

&lt;h1&gt;
  
  
  Encryption #HashiCorpVault #GoLang #CyberSecurity #OpenSource #DeveloperCommunity
&lt;/h1&gt;

</description>
    </item>
    <item>
      <title>Encryption as a Service (EaaS) and Secure Data Vaults: Revolutionizing Data Security in the Digital Age</title>
      <dc:creator>Mohamed Ibrahim </dc:creator>
      <pubDate>Sun, 10 Dec 2023 15:54:24 +0000</pubDate>
      <link>https://dev.to/rafmo20d/encryption-as-a-service-eaas-and-secure-data-vaults-revolutionizing-data-security-in-the-digital-age-6ob</link>
      <guid>https://dev.to/rafmo20d/encryption-as-a-service-eaas-and-secure-data-vaults-revolutionizing-data-security-in-the-digital-age-6ob</guid>
      <description>&lt;p&gt;&lt;a href="https://res.cloudinary.com/practicaldev/image/fetch/s--70TNQHR9--/c_limit%2Cf_auto%2Cfl_progressive%2Cq_auto%2Cw_800/https://dev-to-uploads.s3.amazonaws.com/uploads/articles/ec7st3pthlak4t0x6040.jpeg" class="article-body-image-wrapper"&gt;&lt;img src="https://res.cloudinary.com/practicaldev/image/fetch/s--70TNQHR9--/c_limit%2Cf_auto%2Cfl_progressive%2Cq_auto%2Cw_800/https://dev-to-uploads.s3.amazonaws.com/uploads/articles/ec7st3pthlak4t0x6040.jpeg" alt="Image description" width="800" height="544"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;In today’s digital landscape, data security is paramount. With the increasing frequency and sophistication of cyber threats, organizations are constantly seeking innovative solutions to safeguard their sensitive information. One such solution that has gained significant traction is Encryption as a Service (EaaS) coupled with secure data vaults, a powerful combination that not only ensures data privacy but also enables seamless access and management.&lt;/p&gt;

&lt;p&gt;Encryption as a Service (EaaS):&lt;/p&gt;

&lt;p&gt;Encryption is the process of converting data into an unreadable format, rendering it inaccessible to unauthorized users. EaaS takes this fundamental concept and elevates it to a cloud-based service. Unlike traditional encryption methods, EaaS operates on a scalable and subscription-based model, allowing businesses to encrypt their data without the burden of managing complex encryption keys and algorithms. This service provides a reliable and centralized approach to data security, ensuring that sensitive information remains confidential, even in transit or storage.&lt;/p&gt;

&lt;p&gt;Benefits of EaaS:&lt;/p&gt;

&lt;p&gt;Enhanced Security: EaaS employs robust encryption algorithms to secure data, making it virtually impossible for unauthorized entities to decipher sensitive information.&lt;/p&gt;

&lt;p&gt;Simplified Management: EaaS providers handle the complexity of encryption, allowing organizations to focus on their core operations without worrying about intricate cryptographic processes.&lt;/p&gt;

&lt;p&gt;Scalability: Businesses can easily scale their encryption capabilities based on their evolving needs, ensuring that data security grows alongside the organization.&lt;/p&gt;

&lt;p&gt;Compliance: EaaS solutions often comply with industry standards and regulations, helping businesses meet legal requirements regarding data protection.&lt;/p&gt;

&lt;p&gt;Secure Data Vaults:&lt;/p&gt;

&lt;p&gt;Secure data vaults serve as fortified digital containers where encrypted data is stored. These vaults are equipped with advanced access controls, ensuring that only authorized users can retrieve or manipulate the encrypted information. Additionally, secure data vaults often come with audit trails and monitoring features, allowing organizations to track who accesses the data and when, enhancing transparency and accountability.&lt;/p&gt;

&lt;p&gt;The Marriage of EaaS and Secure Data Vaults:&lt;/p&gt;

&lt;p&gt;When EaaS is integrated with secure data vaults, organizations benefit from a holistic data security solution. Data is encrypted using EaaS before being stored in these vaults, ensuring double-layered protection. Moreover, the seamless integration between the two technologies allows for efficient data management, enabling authorized personnel to access, share, and modify data securely.&lt;/p&gt;

&lt;p&gt;Tech Behind Encryption as a Service and Secure Data Vaults:&lt;/p&gt;

&lt;p&gt;Advanced Encryption Algorithms: EaaS and secure data vaults employ state-of-the-art encryption algorithms such as AES (Advanced Encryption Standard) to safeguard data from unauthorized access.&lt;/p&gt;

&lt;p&gt;Multi-Factor Authentication (MFA): Secure data vaults often utilize MFA to add an extra layer of security, requiring users to provide multiple forms of verification before accessing encrypted data.&lt;/p&gt;

&lt;p&gt;Tokenization: Tokenization techniques are used to replace sensitive data with unique tokens, ensuring that even if the encrypted data is compromised, it remains meaningless without the corresponding tokens.&lt;/p&gt;

&lt;p&gt;Blockchain Technology: Some advanced secure data vaults leverage blockchain technology to create immutable records of data access and modifications, enhancing the auditability and integrity of stored information.&lt;/p&gt;

&lt;p&gt;In conclusion, Encryption as a Service, when combined with secure data vaults and cutting-edge technologies, represents a formidable defense against cyber threats. This integrated approach not only protects sensitive data but also empowers organizations to embrace digital transformation securely, fostering trust among clients and partners in an increasingly interconnected world.&lt;/p&gt;

</description>
      <category>devops</category>
      <category>devscops</category>
      <category>kubernetes</category>
      <category>cybersecurity</category>
    </item>
    <item>
      <title>Best Practices for open banking API</title>
      <dc:creator>Mohamed Ibrahim </dc:creator>
      <pubDate>Sun, 10 Dec 2023 15:48:18 +0000</pubDate>
      <link>https://dev.to/rafmo20d/best-practices-for-open-banking-api-3ohf</link>
      <guid>https://dev.to/rafmo20d/best-practices-for-open-banking-api-3ohf</guid>
      <description>&lt;p&gt;Open banking APIs (Application Programming Interfaces) are instrumental in enabling secure data sharing and financial innovation. Here are some best practices to consider when implementing open banking APIs:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Security and Authentication:
OAuth 2.0: Implement OAuth 2.0 for secure authorization and authentication. Ensure the use of industry-standard flows like authorization code or client credentials.
JWT (JSON Web Tokens): Use JWT for securely transmitting information between parties. Encrypt sensitive data within JWTs.
Strong Encryption: Utilize TLS (Transport Layer Security) to encrypt data in transit. Employ strong, up-to-date encryption algorithms.
API Keys: Use API keys along with OAuth for additional security. Rotate keys regularly.&lt;/li&gt;
&lt;li&gt;Consent Management:
Explicit User Consent: Always obtain explicit consent from users before accessing their financial data. Clearly explain what data will be accessed and how it will be used.
Granular Consent: Allow users to provide granular consent, specifying which accounts or services the third-party can access.&lt;/li&gt;
&lt;li&gt;Data Privacy and Compliance:
GDPR Compliance: Adhere to GDPR (General Data Protection Regulation) or relevant data protection laws in your jurisdiction. Ensure data subjects' rights are respected.
Data Minimization: Only collect and share data that is necessary for the intended purpose. Minimize the scope of data shared.&lt;/li&gt;
&lt;li&gt;Rate Limiting and Throttling:
Rate Limiting: Implement rate limiting to prevent abuse and ensure fair usage. Set appropriate limits for the number of requests per minute/hour/day.
Throttling: Implement throttling mechanisms to handle traffic spikes and prevent server overload.&lt;/li&gt;
&lt;li&gt;Monitoring and Analytics:
Logging: Implement comprehensive logging of API requests and responses for auditing and debugging purposes.
Monitoring: Use real-time monitoring to track API performance, errors, and suspicious activities. Implement alerts for unusual behavior.
Analytics: Analyze usage patterns and user behavior to improve API design and user experience.&lt;/li&gt;
&lt;li&gt;Versioning and Documentation:
API Versioning: Use versioning in your APIs to ensure backward compatibility. Clearly specify the API version in the request header.
Comprehensive Documentation: Provide detailed, easy-to-understand documentation including endpoints, parameters, authentication methods, and sample requests/responses.&lt;/li&gt;
&lt;li&gt;Error Handling:
Meaningful Errors: Return clear and meaningful error messages with appropriate HTTP status codes. Help developers understand what went wrong.
Error Codes: Use consistent error codes and provide a description for each code in the API documentation.&lt;/li&gt;
&lt;li&gt;Testing:
Unit Testing: Implement thorough unit testing for individual API endpoints to validate their functionality.
Integration Testing: Perform integration testing to ensure seamless interaction with third-party applications.
Security Testing: Regularly conduct security testing (e.g., penetration testing) to identify vulnerabilities.&lt;/li&gt;
&lt;li&gt;Developer Support:
Developer Portal: Provide a developer portal with resources like API documentation, SDKs, sample codes, and a sandbox environment for testing.
Developer Support: Offer responsive developer support to assist third-party developers with their queries and challenges.&lt;/li&gt;
&lt;li&gt;Collaboration and Industry Standards:
Collaboration: Collaborate with industry stakeholders, regulators, and other financial institutions to establish common standards and best practices.
Compliance: Stay updated with regulatory requirements and ensure your API complies with evolving standards.
By following these best practices, you can create a secure, reliable, and developer-friendly open banking API that fosters innovation while safeguarding user data and privacy.&lt;/li&gt;
&lt;/ol&gt;

</description>
      <category>openbanking</category>
      <category>fintech</category>
      <category>api</category>
      <category>cybersecurity</category>
    </item>
    <item>
      <title>Best Practices for DevSecOps Implementation:</title>
      <dc:creator>Mohamed Ibrahim </dc:creator>
      <pubDate>Sun, 10 Dec 2023 15:46:03 +0000</pubDate>
      <link>https://dev.to/rafmo20d/best-practices-for-devsecops-implementation-h6g</link>
      <guid>https://dev.to/rafmo20d/best-practices-for-devsecops-implementation-h6g</guid>
      <description>&lt;p&gt;1️⃣ Shift Left, Think Secure: Start security considerations from the project's inception. By integrating security at the early stages of development, we identify and rectify vulnerabilities when they are less complex and costly to fix.&lt;/p&gt;

&lt;p&gt;2️⃣ Automate Security Checks: Embrace automation for security testing. Automated security testing tools seamlessly integrated into the Continuous Integration/Continuous Deployment (CI/CD) pipeline allow for quick and consistent security assessments with every code commit, ensuring that security is not compromised at any stage.&lt;/p&gt;

&lt;p&gt;3️⃣ Culture of Continuous Learning: Foster a culture of continuous learning and collaboration. Security awareness programs, workshops, and knowledge-sharing sessions empower developers, security professionals, and operations teams to stay updated with the latest threats and countermeasures.&lt;/p&gt;

&lt;p&gt;4️⃣ Real-time Threat Detection: Implement continuous monitoring and real-time threat detection mechanisms. By actively monitoring applications and networks, we can promptly identify and respond to security threats, minimizing the potential impact on our systems.&lt;/p&gt;

&lt;p&gt;5️⃣ Compliance and Beyond: Ensure compliance with industry standards and regulations, but don’t stop there. Go beyond compliance and strive for a security posture that exceeds the minimum requirements. This proactive approach ensures a robust defense against emerging threats.&lt;/p&gt;

&lt;p&gt;Essential DevSecOps Tools:&lt;/p&gt;

&lt;p&gt;1️⃣ OWASP ZAP: An open-source security testing tool that helps find security vulnerabilities in web applications during development and testing. Its dynamic application security testing (DAST) capabilities are invaluable.&lt;/p&gt;

&lt;p&gt;2️⃣ Snyk: Snyk is a developer-first security solution that helps you use open source code and stay secure. It finds and fixes vulnerabilities for Node.js, Ruby, Python, Java, and more, empowering developers to write secure code.&lt;/p&gt;

&lt;p&gt;3️⃣ SonarQube: SonarQube is an open-source platform for continuous inspection of code quality. It performs automatic reviews with static analysis of code to detect bugs, code smells, and security vulnerabilities on 25+ programming languages.&lt;/p&gt;

&lt;p&gt;4️⃣ Docker Security Scanning: Docker Security Scanning automatically scans Docker images for vulnerabilities. It provides security intelligence about the software used in your application and its vulnerabilities.&lt;/p&gt;

&lt;p&gt;5️⃣ HashiCorp Vault: HashiCorp Vault secures, stores, and tightly controls access to tokens, passwords, certificates, API keys, and other secrets in modern computing.&lt;/p&gt;

&lt;p&gt;Incorporating these practices and tools into our development lifecycle isn’t just about ticking security checkboxes; it’s about fostering a security-first mindset that permeates every line of code we write. Let's make security not just a part of our process but a part of our DNA as technologists. Together, we can build a digital world that is not only innovative but also inherently secure. 💪&lt;/p&gt;

</description>
      <category>webdev</category>
      <category>devops</category>
      <category>devsecops</category>
      <category>cybersecurity</category>
    </item>
  </channel>
</rss>
