Invoice INV-4471
Meghna Textiles · ৳ 18,42,500
Total matches PO-2208
VAT 15% recalculated
Supplier on master list
AI agents and data foundations · Built in BangladeshAI agents that do the work.
AI agents that do the work.
And prove every answer.
Codemen builds AI agents for operations, sales and support on rules first and models second — so answers are checked, model bills stay small, and your data stays in Bangladesh.
Where this runs
Policy · Finance workloads
| Region | Dhaka · Tier-III |
| Data residency | Bangladesh |
| External models | Off by policy |
WhatsApp · Support
Replied in 4 seconds
আমার অর্ডার #2231 কোথায়?
Codemen agentআপনার অর্ডারটি আজ বিকেল ৪টার মধ্যে পৌঁছাবে। ট্র্যাকিং লিংক পাঠালাম।
Can I change the delivery address?
Codemen agentYes — I’ve sent you a secure link. A colleague confirms the change before dispatch.
Needs a person
LC-0912 · amount differs by 2.4%
Rules couldn’t prove this one, so nothing was approved. Sent to Farhana Rahman, Trade Ops.
Inbound call · 02:14
Voice agent · Sales
Decided this week
Accounts payable agent
94%
by rules, no model
6%
model or person
Memory updated
Kept apart from the model
Credit limit rule v3 replaces v2. Approved by Risk. Every agent uses it from now on.
Screens are illustrative.
A week in operations
Same week. Different team.
Drag the line to see what changes when agents handle the routine and people handle the exceptions.
With Codemen agents
Sun
140 invoices checked against POs
134 proven · 6 to a person
Done 09:10WhatsApp answered in Bangla & English
18 handed to the team
LiveMon
Licence expiries chased automatically
Email + WhatsApp · 2 left
Done 10:02Every call answered and qualified
9 meetings booked
LiveTue
One vendor master record
Approved by Farhana R.
ApprovedLC data flows between systems
Read & written with permissions
Done 11:40Wed
Rules decide; the model only picks
94% with no model call
ProvenAI subscription billed in taka
Local invoice
PaidThu
Weekly report drafted for review
Sent at 08:00
Done 08:00Data residency: Bangladesh
Dhaka · Tier-III
VerifiedTrusted by teams in Bangladesh, Canada and the US
- Technolvelty
- Menkes
- PremiseHQ
- Fiber@Home
- Urmy Group
- Bungalow
- Pico Public Cloud
- Catalys
- Atlas Global Advisors
As engineering partner to PremiseHQ, on work for QuadReal · Westfield · Primaris · Colliers · Morguard · Whitman Family Group
Why AI pilots stall
Most AI projects stop at the same four walls.
The demo works. Then Risk asks how it got the answer, Finance sees the bill, IT asks where the data went, and nobody can switch providers. We start from those questions.
See how we design these outAnswers nobody can check. A model that is right 90% of the time is wrong on one invoice in ten — and you don’t know which one.
Bills in dollars that keep growing. Every task sent to a model is paid for, even the ones a simple rule could decide.
Data leaving the country. Customer and financial records sent abroad for processing, with no clear answer for the regulator.
Locked to one vendor. Your knowledge trapped inside one model or one cloud, and a rebuild every time you want to move.
The Codemen way
Four decisions that make AI safe to switch on.
Most AI pilots stall on wrong answers, dollar bills, data leaving the country and lock-in. We design those problems out before the first agent goes live.
Deterministic first
Rules decide what can be proven. The model does the rest.
Arithmetic, lookups and named checks run first. A model is called only where judgment is needed — often just to pick from a shortlist. Anything unproven goes to a person, never quietly through.
More on checks first7 checks · invoice, packing list, bill of lading
Invoice total = quantity × unit price
USD 84,600.00 · all 12 lines
Totals match across all documents
Invoice · packing list · B/L
Beneficiary matches the LC
Exact match on master record
Shipment date within LC validity
14 Oct ≤ 30 Oct
HS codes on the approved list
3 of 3 codes
Port of loading is Chattogram
Named rule TF-12
Page set complete
Sorted without a model
Decision
Checking…Proven · auto-approved
7 / 7
checks proven by rules
0
model calls needed
Model confidence is never used to approve. If any check fails, the file goes to a person with the evidence.
How it runs
- 01ReadDocuments arrive by email, upload or scan.
- 02CheckArithmetic and named rules run first.
- 03MatchAgainst your master data and limits.
- 04ShortlistOnly on a miss, a model picks from options.
- A personClears exceptions with the evidence attached.
- 06RecordEvery decision logged with its proof.
Sovereign by design
Your data stays in Bangladesh. You decide what leaves.
Apps, agents, databases, memory, files and backups run where you choose — live today in Dhaka on Pico Public Cloud’s Tier-III data centre. Residency is set per workload, and outside model services stay off unless your policy allows them.
More on hosting in Dhaka| Workload | Runs in | Leaves the country? | |
|---|---|---|---|
| Customer records | Dhaka · Pico Tier-III | Never | In country |
| Agent memory (Lore) | Dhaka · Pico Tier-III | Never | In country |
| Backups | Dhaka · second site | Never | In country |
| Open-weight model | Dhaka · dedicated | Never | In country |
| Outside model services | Approved list only | Masked text, if allowed | By policy |
| Audit log | Dhaka | Never | In country |
How it runs
- 01ClassifyWhich data is sensitive, and why.
- 02Set residencyPer workload, not one rule for all.
- 03DeployIn Dhaka, on-premises or your own cloud.
- A personApproves any call to an outside service.
- 05AuditEvery access and model call recorded.
- 06ReportEvidence ready for regulators and the board.
Own the memory
Own the memory. Rent the intelligence.
What your organisation knows — policies, decisions, customers, past cases — lives in Lore, kept apart from the model. Change models without losing it, and always see which decision replaced which.
More on company memoryCredit limit rule v3
Credit limit rule v2
Return policy · বাংলা & English
Key accounts · 1,240 customers
Intelligence layer · swappable
ClaudeGPTGeminiOpen model in Dhaka
Memory unchanged when the model changes
How it runs
- 01CaptureDecisions, policies and cases, with sources.
- 02LinkRecord which decision replaced which.
- A personApproves changes to rules and structure.
- 04RecallAgents answer from memory, citing sources.
- 05SwapChange the model when a better one ships.
- 06KeepYour knowledge stays yours, in-country.
Free from lock-in
Change your cloud. Keep your way of working.
Groundfloor runs on a private cloud, a hyperscaler or a regional provider. Code, databases and data can move — a migration you plan, not a rebuild you fund. And each new app skips weeks of platform plumbing.
What can moveBuilt by handSign-in, permissions, secrets, audit, deploys, databases
320–680 h
On GroundfloorConfigured, not rebuilt
< 40 h
$27k–58k
platform cost avoided per app*
Often lower
hosting cost in Dhaka than on the big global clouds†
* Illustrative, at $85/hour. † Depends on the workload; we price yours before you commit.
How it runs
- 01MapWhat you run today, and what it costs.
- 02ChooseDhaka, on-premises, or keep a hyperscaler.
- 03MoveCode, databases and data, in planned steps.
- A personSigns off each cutover.
- 05RunOld and new side by side until proven.
- 06Switch offRetire the old bill.
Agents
Agents for the teams that carry the workload.
Each agent handles the routine, shows its working, and hands exceptions to a named person — on the channels your customers and staff already use.
140 invoices matched to POs
LC documents checked
Approvals routed
Agentic Operations
Documents checked, approvals routed, records reconciled across systems.
Explore operationsTAQualified
Tanvir Ahmed · Rangpur Agro
NSReview
Nusrat Sultana · Padma Foods
KHBooked
Kamal Hossain · Sylhet Tea
Agentic Sales & Marketing
Every lead answered, researched and followed up — no enquiry left waiting.
Explore sales & marketingবিল কেন বেশি এসেছে?
Codemen agentএই মাসে দুটি অতিরিক্ত সংযোগ যোগ হয়েছে। বিস্তারিত পাঠালাম।
Please cancel one of them.
Codemen agentSent to Billing to confirm — you’ll hear back today.
Agentic Support
Bangla and English answers, day and night, with a person one step away.
Explore supportNames and screens are illustrative.
Data foundation
Good agents need good data. We build that first.
Most organisations have three versions of every vendor, customer and asset. We set the strategy, build one golden record, and keep it clean — so agents act on facts, not guesses.
Data strategy
Which data matters, who owns it, and what AI is allowed to use. A plan your board and regulator can read.
Learn moreMaster data model
One golden record for assets, people, vendors and customers. In production for a Canadian real-estate firm with 155,000 reference rows.
Learn moreMDM Manager
Data stewards review, merge and approve changes in one screen. Nothing changes the master record without sign-off.
Learn moreDocument AI
PDFs, scans, spreadsheets and audio turned into clean, checked data — read, then verified, not just trusted.
Learn moreConnect existing systems
Your ERP, databases and line-of-business apps, read and written with each user’s own permissions.
Learn moreWe build on Microsoft and Google
Microsoft 365, Teams, SharePoint, Azure and Fabric. Google Workspace, BigQuery and Gemini. Years of delivery experience on both.
Learn moreWorks with what you already run
- Microsoft 365
- Teams
- SharePoint
- Google Workspace
- SAP
- Oracle
- Tally
- Yardi
- Salesforce
- HubSpot
- PostgreSQL
- SQL Server
- MySQL
- REST APIs
- SMS
Delivered on Groundfloor
Numbers from real Groundfloor work.
2.9x
ROI for a national REIT
7x
compliance capacity, same team
99%+
processing accuracy on Westfield’s tenant sales data
21
app types delivered
155k
reference rows in a live master data model
1,333
tokens for one memory query, vs 46,146 reading every file
Groundfloor-reported results; not guarantees.
The stack
Built on Groundfloor. Delivered by Codemen.
Codemen is Groundfloor’s implementation partner. The same platform runs client work in Bangladesh, Canada and the US.
Groundfloor
The production platform. Sign-in, permissions, audit, secrets, databases and deploys — configured, not rebuilt for every app.
Early access
Nirman
The AI workbench. Chat with your data, build an app or agent, and publish it to Groundfloor.
Source-available
Lore
Knowledge and memory for agents, kept apart from the model — with a record of which decision replaced which.
Loom
The runtime for digital employees: each with its own role, memory, permissions and schedule.
Atlas
Code and knowledge intelligence. Structured answers instead of reading every file — far fewer tokens.
Dataplane
One governed layer over your databases, so agents read and write with each user’s own permissions.
How it starts
Small first. Proven before it grows.
No year-long programme. One process, live in weeks, measured against how your team does it today.
- Week 1
Pick one process
We sit with the team that owns it, map the rules that already exist and agree what “done” means. You see a working prototype on your own documents.
- Weeks 2–3
Run it alongside your team
The agent works in parallel with people. Every decision is shown with its evidence, and anything it can’t prove goes to a named person.
- From then on
Hand over, then add the next one
Your team runs it on Groundfloor, in Bangladesh or wherever you choose. The memory it built stays yours when the next process joins.
AI Subscriptions
Claude, GPT and more. In one workbench, billed in Bangladesh.
Use leading AI models inside Codemen’s AI workbench, connected to each provider’s API and billed locally. Your company memory stays with you — own the memory, rent the intelligence.
What does “deterministic first” actually mean?
Anything that can be checked with arithmetic, a lookup or a named rule is decided that way. A language model is used only where judgment is needed, and often only to choose from a shortlist. If nothing can prove an answer, it goes to a person. That is why model bills stay small and wrong answers don’t slip through.
Can our data stay in Bangladesh?
Yes. Groundfloor runs live in Dhaka on Pico Public Cloud’s Tier-III data centre. Apps, agents, databases, memory, files and backups can all stay in the country. Outside model services stay off unless your policy allows them.
Which AI models do you use?
Whichever fits the job and your policy — Claude, GPT, Gemini or an open model running in Dhaka. Because your organisation’s memory is kept apart from the model, you can change models without losing what the system knows. See AI Subscriptions.
Do we have to replace our current systems?
No. Agents work with what you already run — Microsoft 365, Google Workspace, your ERP, databases, email and messaging. They read and write with each user’s own permissions.
We already use AWS or Azure. Do we have to move?
No. Groundfloor also runs on a hyperscaler. If you want to move to save dollars or keep data local, we plan the migration in steps and run old and new side by side until it is proven.
How long until something is live?
Usually a working prototype on your own documents in the first week, and a supervised pilot running alongside your team within three weeks. It depends on the process and the systems involved.
Who is Codemen?
A software company in Dhaka and the implementation partner for Groundfloor. We design, build and run AI agents and data foundations for enterprises in Bangladesh, Canada and the US. About us.
Bring us one process.
We’ll show you it running.
A 30-minute call with a Codemen architect. Bring one business problem. We map the workflow, show where rules decide and where a person approves, and what it costs to run in Bangladesh.
Free. No slides unless you ask for them.