đź’ˇ Inspiration
Applying for technical roles today is broken in two opposite ways:
- Manual tailoring takes hours per role: Candidates spend hours reading company engineering blogs, decoding job descriptions, checking ATS keyword density, rewriting resume bullets, and drafting cover letters.
- Existing AI resume tools hallucinate and burn frontier-model tokens: Most "AI job tools" route every single prompt—even simple keyword checks or intent classification—to a single heavyweight proprietary LLM, resulting in slow latencies, high inference costs, and fabricated resume claims that aren't grounded in the candidate's actual projects.
We built AutoApply AI: Nemotron Career & ATS Agent for the Nebius x NVIDIA Global AI Hackathon to prove that a Compound Open-Source AI Architecture—dynamically routing across NVIDIA Nemotron Nano (8B), Super (49B), and Ultra (253B) on Nebius Token Factory and grounded in live web intelligence via Tavily—can deliver higher factual precision at less than half the inference cost.
🚀 What It Does
AutoApply AI is an autonomous 4-node multi-agent career orchestration and Job-Description-grounded ATS platform:
- Node 1 — Guardrail & Structured Resume Parser (
Nemotron Nano 8B+Nemotron Super 49B):Nemotron Nanoperforms sub-second resume validation, domain classification, and initial skill tag extraction.Nemotron Superparses raw PDF/Markdown/text resumes into a strict JSON schema capturing categorized technical skills, quantified work history, and project outcomes.
- Node 2 — Live Job Discovery & Company Engineering Intelligence (
Nemotron Nano 8B+Adzuna API+Tavily Search API):Nemotron Nanodecomposes the candidate profile into high-signal search queries.- Queries Adzuna for live job openings and uses Tavily Search API to pull recent engineering blog posts, infrastructure announcements, and tech-stack news about the hiring company.
- Node 3 — Semantic Match Evaluator & JD-Grounded ATS Checker (
Nemotron Super 49B):- Evaluates the candidate against the exact Job Description across Technical Alignment, Experience Depth, and Company Stack Fit.
- Runs a full Applicant Tracking System (ATS) Audit checking section hygiene, quantified impact density, action verbs, and keyword frequency (
Matched,Injected by Ultra, orMissing), plus generating a copy-ready ATS skills block.
- Node 4 — Grounded Tailoring & Cover Letter Synthesizer (
Nemotron Ultra 253B):- Drafts a publication-ready cover letter weaving together the candidate's real project metrics and live company insights from Tavily.
- Produces Before $\rightarrow$ After ATS Resume Bullet Rewrites grounded strictly in the candidate's actual projects (zero fabricated employers or degrees) and a 3-Pillar Technical Interview Strategy.
- Live Nebius Token Factory Telemetry Dashboard:
- Exposes real-time per-step model routing badges, latency ($\text{ms}$), throughput ($\text{tokens/sec}$), credit cost savings percentage, and full JSON request/response traces.
🛠️ How We Built It
1. Dynamic Three-Tier NVIDIA Nemotron Router (src/llm/nebius_client.py)
All LLM inference runs 100% through Nebius Token Factory (https://api.studio.nebius.ai/v1/) using the OpenAI Python SDK. We engineered a centralized ModelRouter / NebiusNemotronClient that dispatches tasks based on complexity:
| Tier | NVIDIA Model on Nebius | Config | Pipeline Responsibilities |
|---|---|---|---|
Tier 1 (fast) |
nvidia/Llama-3_1-Nemotron-Nano-8B-v1 |
$T = 0.1$, $\text{max_tokens} = 512$ | Intent guardrails, query planning, Tavily signal triage, ATS structural checks |
Tier 2 (structured) |
nvidia/Llama-3_3-Nemotron-Super-49B-v1 |
$T = 0.2$, $\text{max_tokens} = 2048$ | Strict JSON resume parsing, semantic match scoring, JD keyword matrix & knockout risk |
Tier 3 (reasoning) |
nvidia/Llama-3_1-Nemotron-Ultra-253B-v1 |
$T = 0.6$, $\text{max_tokens} = 4096$ | Project-grounded bullet rewrites, Tavily-informed cover letters, interview prep strategy |
2. Mathematical Model for Multi-Tier Credit Savings & ATS Scoring
To
Log in or sign up for Devpost to join the conversation.