CEHPOINT Portal
An AI-assisted operations portal for an IT services company, covering client project intake, delivery tracking, hiring and payments for admins, developers and clients.
git clone git@github.com:the-client/cehpoint-portal.git
Cloning into 'cehpoint-portal'...
fatal: could not read from remote repository.
The company owns the code. I just have the commit history.
hint: the code is private, the product isn't. It's running live at projects.cehpoint.co.in ↗
- Role
- Lead Full-Stack Engineer and Architect
- Tools
- Next.js · React · TypeScript
- Type
- AI Operations Portal
- Year
- 2025 to 2026

01 · Overview
One portal for the whole company
The portal is the system CEHPOINT uses to run its work: clients submit and pay for projects, admins approve and staff them, developers deliver against milestones and the hiring team screens candidates. Each role gets its own portal inside one Next.js app.
I owned it end to end. I took over a React codebase that didn't run, migrated it to Next.js, designed the UI, and built the architecture, backend, AI features and payments, working one to one with the CEO on requirements. Interns added a few small features along the way.

02 · Problem statement
Every step needed a person
The company did everything by hand. Taking in client projects, getting them approved, handing them to developers, tracking progress, screening applicants and collecting payments all needed several people at each step, and the React codebase meant to fix it didn't run.
The goal was to automate the flow with AI. A client submits a project without help from the team; once it's agreed and approved it moves to a developer, and client and developer see the same milestones. The admin controls every step.
- Manual project intake
- Manual candidate screening
- Heavy human process
03 · Architecture
One Next.js app, four portals
Everything runs in one Next.js 15 App Router app on Vercel. Users sign in with Firebase Auth, the ID token lives in an httpOnly cookie, and a Node middleware verifies it with the Firebase Admin SDK and sends each role to its own portal. Server actions and API routes read and write Firestore, and files go to Cloudinary.
All AI work goes through one dispatcher: it reads the active provider, model and keys from Firestore, calls Gemini or OpenRouter, and switches to the other if the first fails. Payments use PayU hosted checkout: the server works out the amount, signs the request and checks the signed response before it activates a project.
04 · Hiring
AI screening for every applicant
Candidates apply with a resume. The server extracts its text with LangChain's PDF loader and the AI analyses it against the role, then scores the candidate's written interview answers the same way.
Admins see an overall score, authenticity and technical accuracy, and a recommendation before they accept or reject. The AI recommends; a person still makes the call.
- Resume analysis
- Answer scoring
- Human decision

05 · Build
AI that finishes inside 60 seconds
The hardest part was making the AI dependable. Project documentation, task generation and candidate analysis are long generations inside a serverless function with a 60 second limit, and with one Gemini key a quota error or timeout left the client or admin stuck partway through.
I moved the API keys out of environment variables into Firestore, where admins manage them. For each request the server plans retries across every active key, giving each round 20% more time while the total stays inside the 60 second budget. Each failure is classed as quota, auth or timeout, which sets the wait before the next key, and Gemini and OpenRouter back each other up. An admin can change provider, model or key at runtime and every AI feature picks it up without a redeploy.
- Key rotation
- 60s time budget
- Provider fallback

06 · Project results
From manual process to one portal
Project intake, delivery tracking, hiring and payments used to be done by hand, and now all of it runs through one portal. It has screened 200+ candidate applications and tracks 24 client projects, with 20+ admin and developer accounts working in it.
- Candidate applications screened
- 200+
- Client projects tracked
- 24
- Admin and developer accounts
- 20+
What shipped
- Migrated a non-working React codebase to Next.js and built all four role portals from scratch.
- Built a project intake wizard that turns a client brief into AI-written project documentation and a quotation PDF.
- Shipped AI candidate screening that analyses resumes and scores written interview answers against the role.
- Built an AI layer where admins switch between Gemini and OpenRouter, change models and rotate API keys without a redeploy.
- Integrated PayU checkout with server-side pricing, coupons, SHA-512 hash verification and one-time project activation.
