FindMyNext.work
FindMyNext.work
How It WorksSign in
Engineering Deep Dive

How It's Built

A vertical AI agent built on modern infrastructure: Next.js 15, Supabase, LangGraph, and multi-provider LLM orchestration.

The Stack

Serverless for scale, persistent for brain. Edge-first delivery with robust orchestration.

Frontend

  • Next.js 15App Router + RSC
  • Tailwind CSSCustom design system
  • Shadcn/UIRadix primitives
  • CloudflareEdge + DNS
  • Fly.ioContainerized deploy

Backend

  • SupabasePostgres + Auth
  • pgvectorRAG embeddings
  • Row Level SecurityTenant isolation
  • Edge FunctionsDeno runtime

Orchestration

  • LangGraphState machine flows
  • Node.js WorkerFly.io long-running
  • Python SearchMulti-source aggregation
  • Queue-drivenHeartbeats + retries

The Precision Pipeline

A multi-stage filtering system that increases intelligence (and cost) only as signal improves.

1

The Gatekeeper

•Zero LLM CostDeterministic

Deterministic SQL filters and RegEx. Hard rules like salary floor, location, and avoid-company lists discard 90% of noise instantly.

2

The Vibe Check

•GPT-5-miniOpenAI

Semantic filtering: Is this 'Director' role actually cold-calling in disguise? Does it require 15 years when I have 10? High-recall pass/fail.

3

The ATS Analyst

•GPT-5.2OpenAI

Full ATS-style scoring. Reads the complete JD against my resume. Scores Role Match, Experience Match, Skills Match. Produces structured justification.

Multi-Provider LLM

Provider diversity where it matters: OpenAI is the default foundation model for scoring, and users can optionally choose Claude for personalized prep + assistant.

OpenAI

Models
GPT-5.2GPT-5-miniGPT-5-nano
Used for
ParsingLite scoringFull scoringOnboardingPrep + assistant (default)Engineering assistance (GPT-5.2)

Anthropic

Models
Claude Sonnet 4.5
Used for
Prep + assistant (user preference)Engineering assistance (Claude Opus 4.5)

Note: We also use Gemini 3 Pro for engineering assistance (development workflows).

Evidence-Based Generation

When the agent drafts materials, it uses strict grounding to significantly reduce hallucinations. Every claim must be grounded in source documents.

1
Ingest
Resume + LinkedIn parsed into semantic chunks, stored in pgvector.
2
Retrieve
RAG pulls only relevant experience for each job.
3
Synthesize
LLM instructed to use only retrieved facts.
4
Cite
Ungrounded claims flagged as Suggestions for human approval.
Grounded Output

Privacy by Design

Built for multi-tenancy from day one. Your data stays yours.

RLS Enforced
Every query scoped to auth.uid()
No Training
User data never trains public models
Edge Only
LLM calls server-side only
Zod Validated
All outputs schema-validated

See It In Action

The technical foundation powers a delightful user experience.

See how it worksBack to home