🤖 Career OS — SOTA Multi-Agent Career Assistant
A production-quality, research-grade personal career agent built with 3-stage iterative training (SFT → DPO → GRPO), multi-agent orchestration , and Karpathy-style training visualization . Designed for publication and real-world deployment.
🏗️ Architecture
┌─────────────────────────────────────────────────────────────┐
│ Career OS Orchestrator │
│ (routes requests, manages context, synthesizes output) │
└─────────────────────────────────────────────────────────────┘
│
┌─────────────┬───────┴───────┬─────────────┐
▼ ▼ ▼ ▼
┌─────────┐ ┌──────────┐ ┌────────────┐ ┌──────────────┐
│ Resume │ │ Job │ │ Career │ │ Salary │
│ Parser │ │ Matcher │ │ Advisor │ │ Negotiator │
│ Agent │ │ Agent │ │ Agent │ │ Agent │
└─────────┘ └──────────┘ └────────────┘ └──────────────┘🧬 3-Stage Training Pipeline
Stage 1: SFT — High-Rank LoRA (r=256)
Method : Supervised fine-tuning on ~16K multi-turn career conversationsData : Resume reviews, job-fit assessments, JSON parsing, coaching dialogues, reasoning chainsConfig : LoRA r=256, α=512, target_modules="all-linear", use_rslora=TrueTime : ~60 min on A100 40GBLR : 2e-4, cosine schedule, 2 epochs
Stage 2: DPO — Preference Optimization (r=64)
Method : Direct Preference Optimization on model-generated preference pairsInnovation : Self-generated pairs using the Stage 1 model — quality scores for structure, actionability, career relevanceConfig : LoRA r=64, α=128, β=0.1Time : ~30 min on A100 40GBLR : 5e-7, 1 epoch
Stage 3: GRPO — Multi-Component Reward (r=32)
Method : Group Relative Policy Optimization with custom career reward functionReward components :
Structure (25%): Headers, bullet points, sections
JSON correctness (25%): Valid JSON when requested
Actionability (25%): Action verbs and concrete steps
Career relevance (25%): Domain-specific terminology
Config : LoRA r=32, α=64, num_generations=4Time : ~40 min on A100 40GBLR : 1e-6, 1 epoch
Total Pipeline Time: ~2.5–3 hours on A100 40GB
📊 Training Dashboard
Karpathy-style visualization tracks all training stages:
Metric What It Shows Loss curve Raw + moving average, best-loss point marked Learning rate Cosine schedule with warmup Reward trajectory Mean reward + trend line across DPO/GRPO Response length Histogram + time series (detects collapse/explosion) Gradient norm Training stability monitoring Career quality JSON correctness + actionability scores
All plots auto-generated during training and saved to ./training_dashboard/.
🚀 Quick Start (Google Colab Pro A100)
Step 1: Set up Colab
Open a new notebook at Google Colab
Runtime → Change runtime type → GPU → A100 (requires Colab Pro/Pro+)Secrets (left sidebar) → Add HF_TOKEN with write permission
Step 2: Run everything
Copy the entire contents of
career_os_complete_colab.py into
one code cell and press
Shift+Enter .
That's it. The script runs all stages end-to-end and pushes the final model to the Hub.
What you get after ~3 hours:
✅ Fine-tuned model weights on huggingface.co/Builder-Neekhil/career-agent-v1
✅ Dataset cached on huggingface.co/datasets/Builder-Neekhil/career-agent-dataset-v1
✅ Training dashboard with publication-quality plots in ./training_dashboard/
✅ Ready-to-use multi-agent Career OS
🎯 Agent Capabilities
Agent What It Does Output Format Resume Parser Extracts structured data from raw resume text JSON (name, skills, experience, education) Job Matcher Scores resume against job description JSON (score 0-100, strengths, gaps, suggestions) Career Advisor Career path planning, skill gaps, interview prep Markdown with structured headers + action items Salary Negotiator Compensation strategies, market research, scripts Markdown with data-backed scripts Orchestrator Routes requests, chains agents, synthesizes output Unified career report
🧪 Running Individual Agents
1 from career_os_orchestrator import CareerOS
2
3 cos = CareerOS ( agent_model = "Builder-Neekhil/career-agent-v1" )
4
5 # Single task
6 result = cos . process ( "Review my resume" , resume_text = "..." )
7 print ( result [ "synthesized" ] )
8
9 # Full pipeline
10 results = cos . full_pipeline (
11 resume_text = "..." ,
12 job_description = "..." ,
13 target_role = "Senior Software Engineer"
14 )
📁 File Structure
File Purpose When to Use career_os_complete_colab.py⭐ ONE-COPY COLAB SCRIPT — everything in one cell Copy into Colab, run once career_os_sota.pyStandalone 3-stage pipeline script Local GPU or scripted training career_os_orchestrator.pyMulti-agent Career OS with 4 specialized agents + orchestrator After training, for inference training_dashboard.pyKarpathy-style visualization dashboard Import during training or for analysis train_colab.pyOriginal single-stage SFT script Simple SFT only train.pyOriginal standalone training script Simple SFT only inference.pyBasic inference script Quick testing demo.pyGradio web UI Interactive demo requirements.txtPinned dependencies Setup
🔬 Research Foundation
Technique Paper arXiv ID What We Used Self-Rewarding LLMs Iterative DPO for instruction following 2401.10020 Stage 2: self-generated preference pairs iGRPO Iterative GRPO with self-feedback 2602.09000 Stage 3: multi-component reward function LoRA Without Regret High-rank LoRA for SFT, low-rank for RL Blog post r=256→64→32 pipeline AgentOrchestra Multi-agent orchestration framework 2506.12508 Career OS agent architecture ResumeFlow LLM resume generation pipeline 2402.06221 Structured output patterns TALLRec Lightweight LoRA tuning for recommendation 2305.00447 SFT data efficiency
📊 Dataset Composition
Total : ~17K multi-turn conversations with system prompts.
📈 Evaluation Benchmark
Since no standard career agent benchmark exists, we propose the Career Agent Evaluation Suite (CAES):
Task Metric What We Measure Resume parsing F1 on extracted entities Name, skills, experience, education Job fit assessment Accuracy vs ground truth Match/no-match classification Career advice quality Human evaluation (1-5) Helpfulness, specificity, actionability JSON correctness Valid JSON rate % of responses with parseable JSON Response structure Section headers, lists Markdown structure score
🔧 Hyperparameters
Stage 1: SFT
Param Value Model Qwen/Qwen2.5-1.5B-Instruct LoRA rank 256 LoRA alpha 512 Target modules all-linear RSLora True Epochs 2 Learning rate 2e-4 Batch size (effective) 8 (2 × 4) Max length 2048 Precision bf16 Gradient checkpointing True Assistant-only loss True
Stage 2: DPO
Param Value LoRA rank 64 LoRA alpha 128 Epochs 1 Learning rate 5e-7 Batch size (effective) 8 (1 × 8) Beta (DPO temperature) 0.1 Max length 2048 Precision bf16
Stage 3: GRPO
Param Value LoRA rank 32 LoRA alpha 64 Epochs 1 Learning rate 1e-6 Batch size (effective) 4 (1 × 4) Completions per prompt 4 Max completion length 512 Precision bf16
📤 Publishing Checklist
To make this publication-ready:
📄 License
Apache-2.0 (inherits from Qwen/Qwen2.5-1.5B-Instruct)
Built for those who take their career seriously. 🚀
