Career Pilot

Landing Page
Unrelated Input Handling
Agent Thought Process
AI Skillset-Matched Radar
Section 1 - Synergy Overview
Section 2 - Technical Strengths and Missing Gaps
Section 3 - Cold Outreach Draft
Section 4.1 - Architectural Fit
Section 4.2 - Day-One Roadmap
Section 4.3 - Interview Gauntlet
Section 4.4 - 14-Day Study Plan
Section 5 - Ingested Resume and Preview
Agent Execution Terminal
Extracted Profile

Inspiration

Every hiring cycle, ambitious computer science students and early-career engineers build solid technical projects—ranging from concurrent key-value stores to custom retrieval agents—yet face immediate rejections or silence from elite research groups and infrastructure teams.

Generic career tools and standard LLM chatbots fail these applicants: they hallucinate hiring criteria, cannot interpret low-level systems architectures, and spit out superficial, cookie-cutter cover letters. Team ScouterAI built Career Pilot to serve as an autonomous career radar and technical auditing agent that bridges this gap with verifiable, ground-truth intelligence.

What it does

Career Pilot conducts an end-to-end, multi-stage audit of a candidate's background against high-impact engineering teams:

Layout-Aware Parsing & Casual Note Ingestion: Ingests dense, multi-column PDF resumes via pypdf layout extraction or takes informal, unstructured student notes.
Semantic Gatekeeper: Evaluates incoming documents to ensure they contain technical, engineering, or coursework context—gracefully handling unrelated inputs without wasting inference compute.
Dual Targeting Radar: Operates in two modes:
1. Custom Target Radar: Audits the candidate against a user-specified lab or engineering team.
2. Autonomous Skillset Discovery: Analyzes codebase primitives (e.g., POSIX threading, lock-free queues, memory mapping) to infer and recommend 3 high-synergy research groups or open-source projects.
Zero-Hallucination Web Discovery: Uses Tavily search to retrieve live preprints, repository discussions, and hiring criteria while locating the verified official career/lab portal.
Socratic Synergy Audit: Yields an executive verdict, percentage synergy score, and a targeted "Recommended Move."
Plain-English (ELI5) Toggle: Features a prompt-adapted mode that translates dense systems hiring verdicts into clear, encouraging layman's terms.
Grounded Cold Outreach: Drafts a hyper-personalized outreach email citing verified lab priorities and real candidate project metrics, equipped with a one-click clipboard copy tool.
Architectural Deep Dive: Synthesizes a 30-day contribution roadmap, an accelerated 14-day gap curriculum, and a technical interview gauntlet with talking points.

How we built it

Inference Backbone: Orchestrated nvidia/Nemotron-3-Ultra-550b via the Nebius Token Factory, driving multi-stage structured reasoning, semantic extraction, plain-English translations, and architectural audits.
Web Intelligence & Grounding: Integrated the Tavily API for advanced, real-time retrieval of active opportunities, preprints, and official careers portals.
Frontend & Architecture: Built with Streamlit, featuring custom CSS styling, dynamic responsive containers, iframe PDF dossier rendering, and asynchronous clipboard components.
Resilient Parsing: Structured PDF stream extraction using pypdf with persistence logic (seek(0)) and full layout preservation.

Challenges we ran into

Hallucination in Cold Outreach: Standard prompts often hallucinated lab research topics. We solved this by strictly grounding Nemotron's audit prompt on live scraped Tavily context.
Multi-Column Resume Parsing: Typical text extraction scrambles two-column technical resumes. We leveraged layout-mode extraction and robust JSON sanitizers to handle complex formatting cleanly.
UI Streamlit Constraints: Achieving a sleek, responsive dark/cyberpunk interface with right-aligned dynamic buttons, copy-to-clipboard buttons, and reactive state persistence required targeted CSS flex wrappers and custom HTML components.

Accomplishments that we're proud of

Zero hallucinations on careers portals and lab focus areas across diverse test domains.
Sub-second plain-English translation toggles that demystify high-level systems feedback for undergraduates.
A live agent thought-stream terminal that gives users transparency into the agent's real-time reasoning and scraping telemetry.

What we learned

Designing multi-turn agent pipelines where web discovery outputs directly condition LLM audit rubrics dramatically improves factual grounding.
Large models like Nemotron-3-Ultra excel at mapping architectural codebase primitives (e.g., concurrency primitives, cache eviction strategies) to target research domains when prompted with structured JSON schemas.

What's next for Career Pilot

GitHub Repository Ingestion: Direct agent integration to inspect commit histories, PR reviews, and AST code graphs for deeper verification.
Interactive Mock Interview Mode: An interactive audio/text agent allowing students to practice defending their gap projects against Nemotron.
Automated Lab Tracking: Periodic background radar pings that alert candidates when new preprints or openings match their target skills.