AI Work Flow for Business¶
Adopt AI in your business. Keep your data on your network. A Forward Deployed Engineer (FDE) makes the rollout real — they audit your data, build evals, and ship the gateway. Read the full FDE chapter. A short landing page in English and বাংলা. For the long version, see the chapters below.
Read this page in your language
This page is bilingual — every section is in English and বাংলা side by side. If you prefer to read the whole site in বাংলা, start at বাংলা - হোম.
AI Work Flow for Business is a working library of patterns and reference implementations for using language models in real enterprise operations. The default is local models on your own network for anything that touches customer data, network logs, internal SOPs, or HR records. For non-sensitive workloads the same code can target a cloud endpoint through a local gateway that scrubs the prompt before it leaves your network.
This site is not "no cloud, ever." It is local-first by policy, cloud-allowed when the data permits it.
One sentence, two languages
English: Adopt AI in your business, keep your data on your network, and use a Forward Deployed Engineer (FDE) to make the rollout real. বাংলা: ব্যবসায় AI আনুন, ডেটা নিজের নেটওয়ার্কে রাখুন, আর রোলআউট বাস্তব করতে একজন Forward Deployed Engineer (FDE) নিয়োগ করুন।
The core idea¶
| English | বাংলা |
|---|---|
| Sensitive data stays on your network. | সংবেদনশীল ডেটা আপনার নেটওয়ার্কেই থাকে। |
| A small local LLM answers first. | একটি ছোট লোকাল LLM আগে উত্তর দেয়। |
| The cloud is only used for sanitized text. | ক্লাউড শুধু ঝাড়াই-করা টেক্সটের জন্য ব্যবহৃত হয়। |
| A Forward Deployed Engineer runs the rollout. | Forward Deployed Engineer (FDE) রোলআউটটি পরিচালনা করেন। |
The decision: local or cloud?¶
| If the data is... | English default | বাংলায় ডিফল্ট |
|---|---|---|
| Customer PII, network logs, internal SOPs, HR records, payment data | Local — regulated, legal will block anything else | লোকাল — নিয়ন্ত্রিত, আইনি দল অন্যটা বন্ধ করবে |
| Anonymized telemetry, public docs, third-party SaaS output, code review | Cloud acceptable — no PII, stronger model, fine cost | ক্লাউড চলবে — কোনো PII নেই, মডেল শক্তিশালী, খরচ সহনীয় |
| Mixed: some sensitive fields in a larger request | Local gateway — local scrubs, then cloud on sanitized text | লোকাল গেটওয়ে — লোকাল ঝাড়ায়, তারপর ক্লাউডে পাঠায় |
The third row is the common case. The local gateway is the pattern this site recommends for most enterprises.
flowchart LR
U[Operator question] --> L[Local LLM: Qwen 1.5B or Gemma 4B]
L --> S[Strip PII + cache lookup]
S --> X{Cloud allowed?}
X -- no --> R[Local response]
X -- yes --> A[Cloud LLM: Azure OpenAI or Bedrock]
A -- structured JSON --> R
R --> V[Re-insert real data, return to user]
style L fill:#e3f2fd
style A fill:#fff3e0
style V fill:#c8e6c9How a Forward Deployed Engineer helps your business¶
| Step | English (what the FDE does) | বাংলা (FDE কী করেন) |
|---|---|---|
| Audit | Walks your data flows, finds what is sensitive, picks the local/cloud split. | আপনার ডেটা প্রবাহ ঘুরে দেখেন, সংবেদনশীল অংশ চিহ্নিত করেন, লোকাল-ক্লাউড ভাগাভাগি ঠিক করেন। |
| Evals | Builds offline test sets, measures accuracy, picks the model size. | অফলাইন টেস্ট সেট বানান, নির্ভুলতা মাপেন, মডেলের সাইজ ঠিক করেন। |
| Deployment | Ships the gateway, monitors it, retrains, and hands it back to your team. | গেটওয়ে চালু করেন, মনিটর করেন, পুনঃপ্রশিক্ষণ দেন, শেষে আপনার দলের হাতে ফেরত দেন। |
Why an FDE is mandatory for local-LLM businesses
Local-first AI is not "install LM Studio and walk away." It is a deployment of a small, regulated model into a real workflow, with evals, monitoring, retraining, and a human in the loop. A Forward Deployed Engineer owns that loop end-to-end. Without one, the project stalls at the demo.
Read the full chapter: adoption/fde.md. The four phases in adoption/index.md are the FDE playbook: discover → pilot → scale → build.
Start here¶
-
Why AI Work Flow for Business?
What this project is, who it is for, and why a local-first design matters.
-
Concrete examples of what you can automate today with a 1.5B-parameter local model.
-
The Forward Deployed Engineer role: audit, evals, deployment, and handover.
-
A four-phase path from "what is this" to running in production across 5+ teams.
-
How the layers fit together: LM Studio, FastAPI, ChromaDB, the modules on top.
Who this is for¶
| Audience | English — what you get | বাংলা — আপনি যা পান |
|---|---|---|
| ISP / telco operations (NOC, field ops, support) | Complaint classification, ticket routing, runbook Q&A | অভিযোগ শ্রেণিবিন্যাস, টিকিট রাউটিং, রানবুক প্রশ্ন-উত্তর |
| Bank IT teams | Internal helpdesk triage, network log analysis, SOP lookup | ইন্টারনাল হেল্পডেস্ক, নেটওয়ার্ক লগ বিশ্লেষণ, SOP লুকআপ |
| Factory operations | Shift handover summarization, anomaly flagging, maintenance drafts | শিফট হ্যান্ডওভার সারাংশ, অসঙ্গতি চিহ্নিতকরণ, মেইনটেন্যান্স ড্রাফট |
University IT, government, and other sectors — see Case Studies for the full audience map.
If you handle sensitive operational data, start local. If your workload is mostly non-sensitive and you just want strong models, the local-gateway pattern still gives you the same code path with cloud endpoints on the back end. Either way, this site is for you.
Quick start¶
- Install LM Studio and load Qwen 2.5 1.5B Instruct (or Gemma 3 4B for harder reasoning).
- Start the local server on
http://localhost:1234/v1. - Open the Getting Started guide and run your first script.
That's it for the local path. If you want to add a cloud endpoint for non-sensitive requests later, flip the base URL in your .env from http://localhost:1234/v1 to your Azure OpenAI / Bedrock endpoint, set LOCAL_GATEWAY_ALLOW_CLOUD=true, and the rest of the code is unchanged.