Mentora
Built Mentora — an agentic AI career coach platform with six specialized agents, a three-tier memory system, and a RAG pipeline keeping per-session cost at ~$0.02.
Mentora is an agentic AI career coach platform powered by a six-agent system — diagnostic, planner, accountability, mock interview, escalation, and memory — running on Groq LLMs. I designed a three-tier memory layer with resume/document chunking and embeddings in PostgreSQL + pgvector, Redis episodic caching, and request-scoped working memory, plus a RAG pipeline that powers natural-language admin queries and voice mock interviews.
Project At A Glance
Timeline
Apr 2026
Industry
AI · Agentic Systems · EdTech
Contribution
Multi-agent orchestration, three-tier memory layer, RAG pipeline, voice mock interviews, and full-stack delivery
Collaboration
Project Visual
Mentora
Summary
What this project demanded.
Mentora is an agentic AI career coach platform powered by a six-agent system — diagnostic, planner, accountability, mock interview, escalation, and memory — running on Groq LLMs. I designed a three-tier memory layer with resume/document chunking and embeddings in PostgreSQL + pgvector, Redis episodic caching, and request-scoped working memory, plus a RAG pipeline that powers natural-language admin queries and voice mock interviews.
Role
Full Stack Developer
Year
2026
Problem Space
Multi-agent systems break when memory, retrieval, and routing aren't designed together. Mentora needed agents that stayed coherent across sessions, a retrieval layer that grounded their responses, and a cost profile that made running it at scale realistic.
I built a three-tier memory (vector store for long-term context, Redis for episodic state, request-scoped working memory) and a semantic RAG pipeline over resume and document chunks. Voice mock interviews used Groq Llama 3.3 with Whisper STT, and the whole system was tuned to ~$0.02 per session.
Capability 01
Next.js
Capability 02
TypeScript
Capability 03
Groq
Capability 04
PostgreSQL
Capability 05
pgvector
Capability 06
Supabase
Capability 07
Redis
Capability 08
RAG
Agentic Build
Mentora had to think across sessions, not just within them.
A multi-agent coach is only useful if the agents remember what happened last time, retrieve grounded evidence in this turn, and don't burn through budget. Memory, retrieval, and cost were the actual product surface.
01
Agents & Memory
- Six-agent system: diagnostic, planner, accountability, mock interview, escalation, and memory
- Three-tier memory: pgvector long-term, Redis episodic, request-scoped working memory
- Resume and document chunking with semantic embeddings for grounded retrieval
02
RAG & Voice
- RAG pipeline powering natural-language admin queries over project data
- Voice mock interviews with Groq Llama 3.3 and Whisper STT
- Per-session AI cost kept at ~$0.02 through retrieval and routing tuning
Signal
6 Agents
Diagnostic, planner, accountability, mock interview, escalation, and memory agents on Groq LLMs
Signal
3-Tier Memory
pgvector long-term, Redis episodic, and request-scoped working memory
Signal
~$0.02
Per-session AI cost via the RAG pipeline and Groq Llama 3.3 + Whisper STT
Insight
Agentic products live or die on memory and retrieval. Get those right and the surface above can feel like a real coach.
Core Interaction Shifts
The product decisions that changed how the experience felt.
Shift 01
Six-Agent Orchestration
Designed a six-agent AI system (diagnostic, planner, accountability, mock interview, escalation, memory) on Groq LLMs for end-to-end career coaching.
Shift 02
Three-Tier Memory Layer
Built long-term memory in PostgreSQL + pgvector, episodic caching in Redis, and request-scoped working memory so agents stayed coherent across sessions.
Shift 03
RAG + Voice Mock Interviews
Engineered a RAG pipeline over resume and document embeddings powering natural-language admin queries and voice mock interviews (Groq Llama 3.3, Whisper STT) at ~$0.02 per session.
Influence & Validation
What changed because of the work.
Mentora shipped as a working agentic platform with the memory, retrieval, and cost discipline needed to run multi-agent coaching at scale.
Six specialized agents coordinated through a shared memory and routing layer
Long-term retention via pgvector embeddings, Redis episodic memory, and per-request working memory
Per-session AI cost held at ~$0.02 through retrieval tuning and Groq Llama 3.3 + Whisper STT
Next Step
Explore Mentora