Services offered by tiptop division

Understanding "dam koto vai?" is not the hard part any more — most large models manage it. The hard part is doing it without paying a foreign API for every message, and without the model inventing a price in front of your customer. That is what we build.

The Bangla voice input and form-free posting on this site are live right now. Try them before you ask us anything.

Laravel · PHP
Next.js · React
Self-hosted LLM — Ollama, vLLM
Bangla & Banglish NLP
Ubuntu · Nginx · queue workers

If your customers message you

Stop losing sales to unanswered messages

1WhatsApp & Messenger auto-reply

আপনার কাস্টমার রাত ১২টায় মেসেজ দেয় — এখন উত্তরও যায় সাথে সাথে

The same five questions all day: দাম কত? ডেলিভারি চার্জ কত? স্টকে আছে? An AI answers instantly in Bangla, Banglish or English, using your own prices and stock. When it isn't confident, it doesn't answer — it hands the customer to you instead of guessing.

How it works: WhatsApp Cloud API → intent and confidence gate → our own self-hosted model → a cloud model only for what's left. Prices are read from your data at answer time, so nothing is improvised.

For: Facebook & Instagram shops, any business selling by message

2Let people describe it instead of filling a form

একটা লাইন বলুন বা লিখুন — বাকিটা AI সাজিয়ে দেবে

Long forms are where users quit, and typing Bangla on a phone makes it worse. Let them speak or write one ordinary sentence — "রংপুরে ৩ রুমের ফ্ল্যাট ভাড়া, ১৫ হাজার" — and your system builds the complete listing: title, category, location, price, description.

How it works: speech runs in the browser, so no audio is uploaded and there's no transcription bill. A local model then filters and classifies before a cloud model returns strict JSON — most requests never leave our server. Location resolves against a real district → upazila → union table, so two places with the same name are never confused.

For: marketplaces, classifieds, job boards, property sites
Try the live demo →

If your data cannot leave your building

Private AI, with no per-message cost

3AI on your own server

Most AI services send every customer message to a foreign provider and charge for each one, for as long as you run it. We install the model on your server instead: your data never leaves the building, and there is no cost per message however heavily you use it.

How it works: open-weight models (Qwen, Llama) on vLLM or Ollama. Honest sizing — 14B fits a single 24 GB GPU; 32B needs about 48 GB. We size to your hardware before quoting, and say so when a cloud model is the better answer.

For: banks, financial services, hospitals, insurers, government
Starts with a paid proof of concept on your own data

4Read every call, not a sample of them

You record thousands of calls and a supervisor listens to a handful. We turn Bangla recordings into what each customer complained about, which problems repeat, how the agent handled it, and which calls need review — across every call, not a 2% sample.

How it works: Whisper for Bangla speech-to-text, then our text classification pipeline over the transcript, all on your hardware. An honest limit: word-perfect transcription of compressed phone audio isn't solved, so we sell the classification layer, which stays reliable even where individual words are wrong.

Send us 50 real calls — we'll analyse them free

No commitment. You'll see what your customers actually complained about and how your agents handled it.

Ask for the free analysis →

If you already have a system

Build it properly, or rescue what's broken

5Code review for bugs that only appear in production

AI writes a lot of code now, and it writes certain bugs very confidently. We find the failures that pass every test and break under real use: errors swallowed in silence, fallback code that can never run, timeouts set above the limit that kills them first, data quietly saved wrong.

How we read code: for failure modes, not style — empty catch blocks, unreachable fallbacks, timeout ordering between client, server and proxy, silent type coercion on write paths, and retries that multiply load during an outage. You get a prioritised list with reproduction steps, and the fixes if you want them.

For: teams shipping AI-assisted code, or inheriting a codebase

6Servers set up right, and rescued when they aren't

Security, SSL, automatic backups, and a deploy your team can run without fear. We also fix servers that already exist — slow pages, unexplained errors, and requests that die halfway leaving nothing in the logs.

Typical setup: Ubuntu with Nginx, PHP-FPM and Node under pm2, supervised queue workers, off-server backups. On rescue work we start from the evidence — slow query logs, FPM limits, worker memory — not guesswork.

For: anyone whose server is a source of anxiety

If you are starting something new

No system yet? We build the whole thing

7Web applications, bilingual from the first line

Bangla and English built in from the start rather than translated in afterwards. Mobile-first, fast on slow connections, built to be found on Google, with an admin panel your own team can operate.

We build with AI, at AI speed — then review it the way we review everyone else's code. Delivery runs in weeks rather than months, and you pay for the hours a build actually takes. Service 5 above is exactly why this is safe: we know which mistakes these tools make confidently, so every generated line goes through that same review before it reaches you.

Stack: Next.js App Router with server components, Laravel with token auth on the API. Locale routing, canonical and hreflang from the first commit, and tag-based cache invalidation so an edit clears only the pages it touched. Delivered deployed and running.

Weeks, not months
For: entrepreneurs and businesses starting something new
No existing system needed — we start from your idea

Describe what is going wrong in your own words — Bangla or English. We'll tell you honestly whether we can help, and whether it's worth what it costs.

[email protected]
8801961349181
Zend Certified Engineer · 15+ years in production