
AI & RAG Search Developer
- or -
Post a project like this- Posted:
- Proposals: 40
- Remote
- #4531220
- OPPORTUNITY
- Open for Proposals

















Description
Budget: 1,500–3,000 fixed price. Quote your own the that gives the best quote I will use but it needs to be in the 1500 - 3000 range
I've built the front-end of an AI recruitment platform. I need one developer to build the database, backend, payments and AI behind it. An employer chats with an AI recruiter, which asks follow-up questions, builds a structured brief, searches candidates, and returns the top 10 with explanations. Read the steps carefully — the method matters as much as the result.
STEP 1 DATABASE
Postgres storing employers, candidates, briefs, saved candidates, connections, and an append-only event log. Every state change writes a row. Nothing updated in place, nothing deleted.
STEP 2 LOGIN
Supabase Auth or Clerk. Don't build auth from scratch. Each account has a subscription status.
STEP 3 API
Endpoints: create brief, fetch shortlist, save/ignore/connect candidate, list saved, list connections. Connection status moves through a fixed sequence enforced in code.
STEP 4 STRIPE
One annual subscription. Server-side Checkout Session, webhook with signature verification, idempotent handler so duplicates don't double-grant. On lapse, account goes read-only — data stays visible, new briefs blocked. Never hide data behind the paywall.
STEP 5 SCHEDULED EMAILS
Cron jobs reading the event log. Interview reminders at 24h and 2h before. Stale-save reminder at 3 days with no contact. Overdue feedback reminder. All triggered by code rules. AI never decides timing.
STEP 6 CHAT TO BRIEF (DeepSeek Flash)
Reads the conversation and returns strict JSON: role, qualifications, min years, flexibility, visa sponsorship, location, start date, plus missing fields. Once per employer message, capped at 6.
STEP 7 FOLLOW-UP QUESTIONS (DeepSeek Flash)
Code decides which field to ask about. AI only writes the wording of that one question. AI must not choose the order or number of questions.
STEP 8 SEARCH — TWO STAGES, AI NEVER READS THE DATABASE
Stage 1: SQL filter on years, qualifications, location, work authorisation. Free and instant.
Stage 2: vector search on survivors only. Embed candidate profiles with Google text-embedding-005 (~$0.005 per 1M tokens), store in pgvector, embed the brief the same way, retrieve the closest 30–50 using an approximate nearest neighbour index (HNSW) so it doesn't scan every vector.
Output: 30–50 candidates. Nothing else in the system sees more than this.
STEP 9 RANKING (DeepSeek V3)
Send the brief plus those 30–50 summaries. Return strict JSON: top 10, each with fit score, 2–3 sentence explanation, and gaps. This is the main AI call and the only one using the stronger model. Valid JSON only, no prose.
STEP 10 REMINDER COPY (DeepSeek Flash)
When a cron job triggers, send a small fact object (e.g. "saved 3 days ago, 3 candidates"). Return one short message. Never receives the candidate database.
STEP 11 COST RULES — NOT OPTIONAL
DeepSeek Flash for steps 6, 7, 10. DeepSeek V3 for step 9. No GPT-4, no Claude, no frontier models anywhere. Implement prompt caching with a stable prefix — saves ~98% on the fixed part of each call. Batch all embedding work; never embed one document at a time (10× more expensive). Use pgvector or free Qdrant only. No paid vector database. No AI call where a rule will do. No agent framework that lets the model choose its own actions.
STEP 12 DEPLOY AND HAND OVER
Deploy to Railway or Render. Settings file showing which model is used at each of the four AI call sites. Cost log per brief: cached vs uncached input tokens, output tokens, computed dollar cost. Test set of 10–20 example briefs. One-page handover stating model versions and expected monthly cost at 100 briefs. 30-minute handover call.
WHAT THE AI MUST NEVER DO
Never filter candidates (step 8). Never decide who or when to email (step 5). Never change candidate, connection, or subscription status. Never answer a candidate's question about visas, salary, or contracts. Never message a candidate unless a code rule triggered it.
WHAT I'M LOOKING FOR
You've built both a backend and an AI search or chat system — both, not one. Send a link. You've done Stripe subscriptions and webhooks — show me one. You can state what your last AI project cost per search in real numbers, not a range. Clear written English.
Fos M.
100% (3)New Proposal
Login to your account and send a proposal now to get this project.
Log inClarification Board Ask a Question
-

Hi, quick question before I send my quote: where will the candidate profiles come from, and roughly how many are there at launch? For example, is there an existing dataset or CSV, or do candidates sign up and fill in a form (or upload CVs that need parsing)? This affects the embedding batch work and how I scope Steps 1 and 8.
-

A few questions:
1. Is your existing frontend built in React/Next.js, and is it already connected to any backend?
2. Do you have candidate profiles and sample recruitment briefs ready for the initial search and ranking tests?
3. Should the event log be the primary source for reconstructing application state, or do you also expect separate current-state tables maintained transactionally?
4. Do you already have DeepSeek and Google AI API accounts? -

-/ Are candidate profiles already available in a database or CSV, or should we also build the candidate import and profile management functionality?
-/ For the employer chatbot, should previous conversations and incomplete hiring briefs be saved so employers can resume them later?
-/ For employer-candidate connections, should the system send emails to candidates, or will communication initially remain within the platform? -

Hi Fos,
For Step 8, roughly how many candidate profiles will be loaded at launch, and are they in a structured format (CSV or database), or do they still need to be imported and cleaned?
Additionally do you have any specific timeline for this project?
Thanks,
Asif -

Is your candidate database and frontend already prepared for integration, or would you like us to build the complete backend pipeline, including candidate ingestion, embeddings, and AI-powered matching from scratch
