• 

  Step-by-step plan to bridge resume gaps: HIPAA, EHR tools, insurance workflows, and resume tweaks.

• 

  Generated interview questions tailored to resume gaps and target role in US healthcare admin.

• 

  Skill matrix showing clinical strengths, missing US billing tools, and emerging digital health skills.

• 

  Resume analysis with strengths, gaps, and tailored advice to align with US healthcare roles.

• 

  Radar chart showing resume fit: 25% match with JD, highlighting missing skills and risk areas.

• 

  CareerPilot login/register screen with action selector, secure password input, and clean dark-themed UI.

• 

  CareerPilot dashboard lets users upload or paste resumes for analysis, matching, and interview prep.

Inspiration

The job search process is broken. Candidates spend hours rewriting resumes, tailoring applications, and preparing for interviews — yet still get filtered out by ATS systems or mismatched expectations.
We wanted to build an Action‑Era AI agent that doesn’t just “assist” but actually does the work: analyzing resumes, understanding job descriptions, extracting insights from video, and generating a complete, personalized application package automatically.

CareerPilot was born from a simple question:
What if applying for a job felt like having a personal career strategist working beside you?

What it does

CareerPilot is an Autonomous Multimodal Job Application Agent powered by Gemini 3 that:

• Analyzes resumes, job descriptions, and even video inputs
  
• Extracts skills, gaps, strengths, and role alignment
  
• Generates a personalized FitGraph and insights
  
• Crafts a tailored resume rewrite
  
• Produces a preparation plan and mock interview questions
  
• Evaluates user answers in real time
  
• Stores history and learns from past interactions
  

It transforms raw candidate data into a high‑impact, job‑ready application package — instantly.

How we built it

CareerPilot is engineered as a modular, production‑grade AI pipeline:

• Streamlit UI for a clean, interactive user experience
  
• FastAPI backend orchestrating authentication, routing, and streaming
  
• LangGraph agent coordinating the multimodal workflow
  
• Gemini 3 for embeddings, knowledge generation, and final analysis
  
• MongoDB Atlas Vector Search for RAG augmentation
  
• Redis for caching and performance optimization
  
• Caddy + FreeDNS for secure HTTPS deployment
  
• k3s cluster running containerized microservices
  

We instrumented every step with millisecond‑level timestamps to understand real‑world latency and optimize the pipeline end‑to‑end.

Challenges we ran into

• Gemini LLM latency: 50–60 seconds per analysis, dominating 95% of total runtime
  
• DNS + HTTPS deployment: DuckDNS and Dynu failed for ACME challenges; solved via FreeDNS + HTTP‑01
  
• Containerd vs Docker: Local images weren’t visible to k3s until we rebuilt the pipeline
  
• Multimodal extraction: Video‑to‑text required careful frame handling and fallback logic
  
• Prompt engineering: Ensuring structured, deterministic outputs from Gemini
  
• State management: Coordinating multiple async steps inside LangGraph
  

Every challenge forced us to rethink architecture, improve instrumentation, and build a more resilient system.

Accomplishments that we're proud of

• Built a fully autonomous multimodal agent in under 48 hours
  
• Achieved sub‑2‑second performance for everything except LLM calls
  
• Designed a production‑grade HTTPS deployment on a homelab cluster
  
• Created a transparent latency dashboard exposing real bottlenecks
  
• Delivered a complete job application package from raw inputs
  
• Built a system that feels like a career co‑pilot, not just a chatbot
  

And the best part — it’s live and testable by anyone.

What we learned

• Instrumentation beats intuition — measure first, optimize second
  
• RAG pipelines are only as fast as their LLM bottleneck
  
• Multimodal workflows require careful state design
  
• Caching transforms user experience
  
• Deployment is half the battle
  
• Gemini is powerful, but long‑form reasoning latency needs serious improvement
  

This project taught us how to build real AI products, not just demos.

What's next for CareerPilot — Autonomous Multimodal Job Application Agent

CareerPilot is just getting started. Next steps include:

• Autonomous job search: Scan job boards and auto‑match roles
  
• Auto‑apply workflows: Fill forms, rewrite resumes, and submit applications
  
• Video interview agent: Real‑time feedback during mock interviews
  
• Portfolio builder: Auto‑generate GitHub projects based on JD gaps
  
• Skill gap learning paths: Personalized upskilling recommendations
  
• Faster LLM pipeline: Model distillation, streaming, and hybrid reasoning
  
• Mobile app for on‑the‑go career coaching
  

Built With

• caddy
• docker
• fastapi
• freedns
• google-gemini-3-api
• kubernetes-(k3s)
• langgraph
• mongodb-atlas-vector-search
• python
• redis
• streamlit