FindHackers.co

A Case Study In Building A Developer Hiring OS On Modern Infra

1. Overview

FindHackers is an integrated platform that connects developers who need visibility with clients and companies who need talent. It combines narrative portfolios, AI powered assistants, analytics, and job aggregation into one product that serves both sides of the hiring market. This case study walks through:

The problem FindHackers set out to solve
How the team designed the product experience
How the stack was implemented using Next.js, AI SDK, Supabase, Hetzner, and Node.js
What worked well and what still hurts

2. The Problem

For developers:

GitHub is a code dump, not a story
LinkedIn is noisy and keyword driven
Personal sites are hard to maintain and rarely updated

For clients:

Hiring signals are fragmented across GitHub, LinkedIn, blogs, talks, and random links
It is hard to quickly understand what a developer is actually good at
Inbound applications are noisy, outbound sourcing is time consuming

The team wanted a system where:

Developers can host their entire builder identity in one place
Clients can see a real story, not just a CV and buzzwords
AI can handle repetitive questions and research so humans do not have to

3. Product Solution

At a product level, FindHackers is built around two surfaces.

3.1 Developer Surface

Profile and portfolio Developers get a single canonical profile that holds:

Projects with narrative descriptions
Screenshots, demos, stack decisions, and lessons learned
Links to GitHub, personal sites, talks, and content

Luminous AI: the “digital twin” Each profile has its own AI assistant that:

Answers questions about projects, stack choices, tradeoffs, and experience
Uses the portfolio, linked content, and structured data as context
Gives visitors rich answers without the developer having to repeat themselves

Analytics The platform tracks:

Profile views and referrers
Per project engagement
Clickthroughs for contact and external links

This gives developers a clear sense of whether their profile is actually working. Job discovery A companion Jobs experience plus a Chrome extension pulls developer relevant roles from LinkedIn and X into a distraction free dashboard. Noise goes down, signal goes up.

3.2 Client Surface

Hiring dashboard Clients can:

Search for developers by stack, seniority, and region
See consolidated portfolios that actually tell a story
Filter based on the type of work they need done

DeepResearch AI For each developer, DeepResearch AI:

Crawls public signals like GitHub, LinkedIn, blogs, conference talks, and repos
Builds a structured profile including skills, patterns, activity, and interests
Produces a narrative summary that helps non technical hiring managers understand a candidate quickly

4. Architecture And Tech Stack

4.1 High level architecture

The system is built as a Next.js application deployed on Hetzner, with Supabase as the primary data and auth layer and AI features wired through the AI SDK. High level pieces:

Web application
- Next.js App Router
- Hybrid static and dynamic routes
- React Server Components where possible
API and AI layer
- Next.js route handlers as thin APIs
- AI SDK for chat, tools, and streaming responses
- Shared TypeScript types between UI and handlers
Data and auth
- Supabase Postgres for core data
- Supabase Auth for user accounts
- Row Level Security (RLS) for multi tenant safety
- Supabase Storage for images and assets
- Supabase vectors for embeddings that power Luminous AI and DeepResearch
Infrastructure
- Hetzner cloud instances
- Docker for app, worker, and proxy containers
- Caddy on Hetzner for TLS and routing
- Node.js runtime for Next.js server and background workers
Background processing
- Worker process for:
  - Crawling external sources for DeepResearch
  - Building embeddings and refreshing AI context
  - Scheduled job imports and cleanup

4.2 Next.js and AI SDK

The team chose Next.js as the primary web framework to keep everything in one codebase:

Marketing pages, product app, and docs share a single project
App Router allows streaming server components and data fetching patterns that fit AI use cases

The AI SDK is used for:

Luminous AI chat on profile pages
DeepResearch pipelines that:
- Fetch external data
- Extract structured signals
- Generate summaries and scores

Key patterns:

Each AI feature has a dedicated route handler in app/api/...
Requests are validated with Zod schemas before anything touches AI models
Streaming responses power chat UIs so users see tokens as they are generated
Long running research tasks are split into:
- A trigger endpoint that enqueues work
- A worker that persists final results in Supabase
- A polling or subscription based UI that updates when the report is ready

4.3 Supabase schema design

Supabase is the “brains” of the system. Major tables include:

profiles for user level metadata
projects with rich fields for description, stack, links, and media
research_jobs to track DeepResearch runs for a given identity
research_sources to store crawled artifacts (repo, post, talk, etc)
research_events as a log of steps in the pipeline
embeddings tables that hold vector representations of text for semantic search

RLS rules enforce that:

Developers can only edit their own profiles and projects
Clients can only see data that the developer has made public
Background jobs have service role access for write heavy tasks but queries from the app are always filtered through policies

Supabase functions and triggers are used for:

Updating derived stats such as view counts
Maintaining search indexes
Keeping AI context snapshots in sync with the latest portfolio data

4.4 Hetzner and deployment story

Instead of relying only on a managed platform, the team decided to run the production stack on Hetzner:

A main app server runs the Next.js app in production mode using Node.js
A worker container runs background queues
Caddy routes incoming traffic to the right service and handles certificates

Reasons this has worked well:

More control over resource allocation and per region deployments
Easier to colocate additional services as the platform grows
Predictable costs compared to scaling vertically on a pure serverless model

Tradeoffs:

The team owns OS level security and patching
Zero downtime deploys and migrations need more careful planning
Monitoring and alerting had to be wired manually with off the shelf tools

5. Implementation Highlights

5.1 Luminous AI on profile pages

Implementation pattern:

User visits /[username] profile
Page loads profile, projects, and precomputed embeddings from Supabase
When the user opens the AI chat, the client sends:
- The question
- A profile identifier
Server side handler:
- Retrieves relevant chunks from Supabase using vector search
- Builds a prompt with system instructions plus retrieved context
- Uses AI SDK to stream a response back to the UI

This keeps the AI cheap and targeted. It only searches data for that specific developer and limits tokens to what is actually needed.

5.2 DeepResearch pipeline

For DeepResearch, the goal was to take a minimal input such as a GitHub username and produce a rich profile. High level steps:

Client hits POST /api/research/start with a handle
A research job row is created in Supabase
Worker consumes the job and:
- Calls external APIs or scrapers for GitHub, LinkedIn, blogs, and talks
- Stores raw artifacts in research_sources
- Runs extraction prompts through AI SDK to produce normalized facts
- Embeds text into vectors for later semantic queries
Final narrative is generated and stored in research_jobs as report_json
Frontend polls or subscribes to job status and renders the final report in the client dashboard

This architecture keeps slow work in the background while the main app stays responsive.

6. Results And Impact

Some concrete effects from this architecture and product design:

Developers get a single link that acts as their “home on the internet” instead of juggling multiple half updated profiles
Luminous AI absorbs a large chunk of repetitive questions from recruiters and collaborators
Clients get a faster path from “who is this person” to “should we talk to them” because DeepResearch collapses scattered signals into one narrative

From an engineering perspective:

Using Next.js plus AI SDK meant less glue code and more focus on prompt and UX design
Supabase provided authentication, storage, SQL, and vectors in one place which reduced the need for additional services
Hetzner hosting allowed the team to scale predictably while still keeping full control of runtime and networking

7. What Worked Well

Single repo for marketing, app, and AI routes simplified development and deployment
Strong typing between Next.js API handlers and front end components reduced integration bugs
Using embeddings for everything from search to AI context made features like “ask my portfolio” and DeepResearch feel coherent
Running on Hetzner with Docker and Caddy provided a nice middle ground between raw VPS work and expensive fully managed platforms

8. Challenges And Lessons Learned

Background research and scraping can be brittle due to upstream rate limits and markup changes
Keeping AI features fast while using heavier context often required multiple rounds of prompt and retrieval tuning
With more data, Supabase indexes and RLS rules had to be revisited for performance
Observability matters: logs, traces, and query insights are critical once AI features hit production traffic

9. Next Steps

Upcoming work on FindHackers builds on the same stack:

Richer developer onboarding flows that auto import and summarize GitHub and content
Deeper hiring workflows on the client side, such as shortlists, project briefs, and interview notes linked to profiles
More automation in DeepResearch, including scheduled refreshes of developer personas

Under the hood, the plan is to keep the core stack stable:

Next.js for the app
AI SDK for model access and tooling
Supabase for identity, data, and vectors
Hetzner for controlled, cost effective hosting
Node.js for server and worker processes

The main focus now is not changing tools, but compounding product value on top of a stack that is already battle tested.

FindHackers