OtaskuScout
An applied agentic AI job intelligence platform that turns job alerts and manual postings into structured, deduplicated, human-reviewed recommendations. The project explores how agentic workflows can support career decision-making without removing human judgment.
Most job-search tools either dump listings into a queue or overpromise automation. OtaskuScout is designed around a more practical question: how can AI agents reduce noise, surface better opportunities, and explain their reasoning while keeping the human in control?
- How to structure agent runs as durable, inspectable system activity
- How to separate recommendations from actions using approval workflows
- How to evaluate job fit using both heuristic scoring and LLM-based Deep Fit analysis
- How to track token usage and estimated cost for AI-assisted workflows
- How to learn from approval/dismissal patterns without pretending the system is fully autonomous
- Gmail and manual posting ingestion
- Job normalization and deduplication
- Fit scoring and bucketed recommendations
- Persistent agent runs and run history
- Task proposal workflows
- Human-in-the-loop approvals
- Deep Fit LLM analysis
- Daily prioritized plans
- Token and cost tracking
- Decision-learning signals
- Next.js application structure
- PostgreSQL/Neon-backed persistence
- Gmail ingestion and manual ingest paths
- Redis/caching layer for repeated lookups and workflow performance
- LLM review layer for Deep Fit analysis
- Agent run/task proposal tables for durable workflow state
- Human approval layer before recommendations become actions
- Cost-aware token tracking for AI operations
Architecture posture
OtaskuScout is built as a full-stack web application with a practical emphasis on traceability, cost awareness, and controlled automation. Agent activity is modeled as persistent runs rather than invisible background magic, making outputs easier to inspect, audit, and improve.
OtaskuScout is intentionally designed to avoid fully automated career decisions. The system can propose, prioritize, and explain, but human approval remains central. This keeps the tool useful without turning it into an opaque decision engine.
- Recommendations require human review
- Fit analysis includes reasons and gaps
- Token/cost tracking makes AI usage visible
- Decision-learning signals improve future suggestions without hiding uncertainty
- The system avoids presenting LLM output as final authority
AI-assisted development
The project also serves as a practical testbed for modern AI-assisted software delivery. I used tools such as Codex, Continue.dev, Ollama/local models, and OpenAI APIs to accelerate feature planning, debugging, refactoring, and agentic system design while maintaining human review over implementation decisions.
Current status
Active / Building. Public-facing material should describe the system and architecture without exposing private job-search data, Gmail content, or internal application screens containing sensitive information.