Talha Tanveer

Case study · Web Applications

Beostoc

A connected B2B platform for Irish agrifood — sales, live-animal supply and production, now with supervised AI agents.

Screenshot of Beostoc
Category
Web Applications
Role
Full-stack developer (freelance)
Team
Joined an existing team on an established codebase
Duration
Oct 2026 – present (ongoing)
Technologies
  • Laravel
  • Vue 3
  • TypeScript
  • Python
  • FastAPI
  • MySQL
  • Redis
  • Docker

01

Overview

Beostoc Agrifood runs B2B sales, live-animal supply and production from one role-based workspace, where each person’s role decides what they can see and do. I joined an established platform and, working in milestones, added a supervised AI operations layer, a Data Centre for imports and exports, a Tools module for farm workspaces, and the back end for a new public website — with tests and deployment tooling throughout.

02

Key features

  • Role-based workspace for staff, farmers and B2B clients
  • Sales, quotations, orders and settlements
  • Supervised AI agents with human approval gates
  • Data Centre: previewed imports, exports and audit log
  • Tools module for farm workspaces and teams
  • Responsive admin panel

Stack

How it’s built

Back end

  • Laravel (PHP)
  • MySQL 8
  • Redis
  • Sanctum / JWT auth
  • Swagger API docs

Front end

  • Vue 3
  • TypeScript
  • Pinia
  • Vite
  • Bootstrap

AI service

  • Python
  • FastAPI
  • arq queue
  • Anthropic, OpenAI & Gemini APIs
  • OCR document pipeline

Quality & delivery

  • PHPUnit
  • Pytest
  • Playwright E2E
  • GitHub Actions CI
  • Docker

Client

In the client’s words

Verified on Fiverr
“I initially left a three-star review, but I decided to continue working with Talha and put my trust in him. I’m glad I did. My experience since then has been very positive, and I’m happy to revise my opinion. He has shown himself to be skilled, professional and knowledgeable in his field.”
premiumlm · United Kingdom · 2 weeks project

03

My contribution

  1. 01

    AI Operations Workforce

    Designed and built a separate Python (FastAPI) service where a supervisor agent plans the work and specialist agents read data through permission-checked tools. Runs go through a state machine with budgets, an emergency stop and routing across Anthropic, OpenAI and Gemini — and anything that changes business records waits for a person to approve it.

  2. 02

    AI Operations Centre in the panel

    Added the screens staff use to work with the agents: a command centre, task runs whose findings link back to source records, an approvals inbox, schedules, provider health and per-task cost. “Ask AI” and “Explain status” now sit on customer, order and product pages.

  3. 03

    Data Centre

    Brought imports, exports, templates, scheduled reports and the audit log into one place. Every import runs as a preview first, validates each row, guards against duplicates and records rejected rows for download.

  4. 04

    Tools module

    Built farm workspaces with their own owners, team invitations and access levels, keeping each client’s prices isolated from other clients.

  5. 05

    Public website back end

    Built the Laravel back end for a new public website: content approval, governed facts and quality claims, automation rules, chatbot hand-off and image processing — plus release scripts with health checks, automatic rollback and backups.

  6. 06

    Security, testing and delivery

    Closed an account-enumeration gap in password reset, signed the internal API between services, and added role-based end-to-end tests, isolated nightly test runs and CI pipelines.

04

Screens

Select an image to view it larger. Arrow keys switch images; Esc closes.

05

Challenges

Challenge

Letting AI agents help with real operations without letting them make business decisions.

How I addressed it

Every tool call is permission-checked as the requesting user; business changes need human approval; budgets per role and facility, kill switches and a “never autonomous” guard keep agents in bounds, and duplicate calls are replayed instead of run twice.

Challenge

Joining a platform that had been in development for nearly three years.

How I addressed it

I mapped the existing modules before changing them, followed the codebase’s conventions, and added tests around each area I touched so new work couldn’t quietly break old behaviour.

Challenge

Tests that passed locally but failed in CI.

How I addressed it

MySQL 8 in CI reorders keys inside JSON columns, so I made the comparisons key-order independent and moved end-to-end runs into isolated environments that clean up after themselves.

06

Outcome

Beostoc is live at mybeostoc.ie.

  • Delivered in agreed milestones — AI workforce milestones 1–4 and a final completion round — with tests, documentation and handover notes.
  • 479 commits across five repositories (Laravel API, Vue panel, website back end and front end, AI service) between October and November 2026.
  • The client moved from a three-star to a five-star review and continued the collaboration (see the review above).

07

What I learned

“Clear contracts between services — signed APIs, documented endpoints and explicit approval rules — are what make it safe to add AI to a system people already depend on.”

Next project

Reel Loyalty

A loyalty-card platform extended with video reels, ads, paid plans, an influencer portal and single sign-on to partner apps.

Have an idea? Let’s build it.

If a project like this is on your roadmap, I’d be glad to hear about it.