NEXAI — Development Story | AMELA Project Portfolio

CLIENT

About the customer

A leading domestic "mid-sized IT company" driving operational efficiency.

The mid-sized IT company is an organization carrying diverse knowledge-work across multiple departments. High-volume application CV Screening in HR, multilingual document translation in Sales/Marketing, email distribution in the CRM department, and periodic reports to executives ─ in every department, pressure on working hours and decision-making speed had become the norm.

With the goal of letting every employee embed AI into their work with one click, we built an internal AI productivity platform in-house. It is designed as a unified workspace where the same AI backend can be used across three channels: Web, API, and Telegram.

Cumulative CV processed

12,847items

Translated files

3,421items

Email verification

89,234items

BUSINESS FLOW

STEP

Business request

Employees and executives request translation / CV Screening / data checks

STEP

Unified intake

From Web / API / Telegram to the same backend

STEP

AI engine processing

Automated processing with Gemini / DeepSeek + a dedicated pipeline

STEP

Instant output

Results returned in seconds to tens of seconds, compressing working hours

CHALLENGE

Design challenges to solve

The four structural requirements demanded of the platform.

None of these were technology-selection problems but business requirements tied directly to cost, processing characteristics, touchpoints, and storage. A from-scratch new build required a design that "satisfies them all at once."

① Allowing individual LLM calls causes cost management to collapse

If each business tool called Gemini / DeepSeek directly,Unpredictable monthly costsand Rate-limit contention happens.

② Consolidating AI tasks into a single processing queue causes processing to jam

CV parsing takes seconds, translation takes minutes, and the Bot responds instantly ─ pushing AI jobs with different processing characteristics through the same queue would meanBot responses that need to be instant also get delayed by translation jobs。

③ Implementing AI per channel causes features to fragment

Employees use the Web, executives use Telegram, and external systems use the API ─ the touchpoints differ. Implementing AI logic per channel would meanThe same feature gets implemented three times state.

④ A flawed storage design exposes confidential data to leakage risk

CV PDFs and translation files are Contains a lot of confidential information. Without designing access control, encryption, and usage logs from the start,Extremely hard to bolt on later becomes.

ARCHITECTURE

The architecture we built together

Turning knowledge-work into a "never-stopping foundation" across four domains.

We built an AI platform supporting the operations of a mid-sized IT company entirely in-house, from the ground up. Four structural requirements ─ the AI processing foundation, backend architecture, user touchpoints, and data management ─ were each designed together with Amela, brought together into an architecture that satisfies cost, performance, touchpoints, and security all at once.

Area ①

AI processing platform

Challenge

When each business tool calls the LLM individually, cost management and rate control become fragmented.

Solution

Centralized through a shared AI gateway

Calls Gemini / DeepSeek through a single abstraction layer, implementing cost caps, caching, and fallback

Area ②

Backend architecture

Challenge

Diverse AI tasks (CV parsing, translation, Bot) differ in both processing characteristics and completion time.

Solution

Asynchronous separation via role-based microservices

NestJS microservices + Bull Queue + Redis split AI workers by role, processing the instant-response Bot and long-form translation independently

Area ③

User touchpoints

Challenge

Employees use the Web, executives use Telegram, and external systems use the API ─ the touchpoints split into three lines.

Solution

From three channels to the same backend

Next.js Web UI + Grammy Telegram Bot + REST API make the same AI features usable from three routes

Area ④

Data management

Challenge

Many CV PDFs and translation files get uploaded, requiring storage and access control.

Solution

S3 + Prisma unify persistence and tracking

Confidential files are stored encrypted on AWS S3 and delivered securely via signed URLs. Usage logs are fully tracked with Prisma

TECH STACK

Frontend

Next.js

React T TanStack Router s shadcn/ui

Tailwind

i18n

Backend

NestJS

Prisma B Bull Queue

Python Flask

Grammy (Telegram)

Infra & AI

Gemini D DeepSeek

MySQL

Redis

AWS S3

Docker

IMPACT

The value the platform creates

The mid-sized IT company's "operational speed" accelerates.

This foundation continuously produces business outcomes ─ reduced working hours, faster decision-making, and cross-departmental use of data ─ rather than technical metrics.

① A foundation that compresses knowledge-work hours

CV Screening, translation, email verification, report aggregation ─ AI automates each department's routine tasks. Employees can focus on judgment-based work.

② A design that makes executive decisions instant

When executives ask the AI on Telegram, Revenue / pipeline / Alerts come back on the spot. Report wait time approaches zero.

③ A design that maximizes AI value across departments

HR, Sales, Marketing, and executives all use the same AI backend. Add a new feature and every employee can use it instantly.

④ An architecture that lets you focus on continuously expanding AI use

With a shared AI gateway + microservices architecture, a new AI use case can go into production simply by "adding a new worker."

👥

STAKEHOLDER

HR staff

Hiring screening

Typically

At other organizations, the only option is to have people read dozens of CVs a day by eye.

At a mid-sized IT company

AI scores every entry within 24h, so HR only needs to read the top 10% closely.

Business impact

Screening 10× faster

Fewer top candidates slipping through

📄

STAKEHOLDER

Sales / Marketing

Multilingual documents

Typically

At other organizations, multilingual document translation tends to fall back on outsourcing or hours of manual work.

At a mid-sized IT company

Submit from the browser and translation finishes minutes later, with Source ⇄ Translation viewable on the spot.

Business impact

Translation effort cut 70%

Faster overseas expansion

✉️

STAKEHOLDER

Marketing / CRM

Email delivery quality

Typically

At other organizations, bounces on large volumes of email addresses are often only detected after sending.

At a mid-sized IT company

AI verifies in bulk before sending, excluding invalid addresses in advance.

Business impact

Delivery failure rate sharply reduced

Improved sender reputation score for the delivery provider

📊

STAKEHOLDER

Executives

Decision-making speed

Typically

At other organizations, the process of report request → aggregation in each department → sharing takes hours to days.

At a mid-sized IT company

Ask the AI on Telegram and KPI / pipeline / Alerts come back instantly.

Business impact

Faster decision-making

Realizing data-driven management

Knowledge workBuilt in-house from scratch into an AI workspace.

A leading domestic "mid-sized IT company" driving operational efficiency.

The four structural requirements demanded of the platform.

Turning knowledge-work into a "never-stopping foundation" across four domains.

The mid-sized IT company's "operational speed" accelerates.

Experience the mid-sized IT company's journey with a live demo.