CareerOrbit — Autonomous Personal AI Career Agent

24/7 Personal AI Career OS powered by 3-tier NVIDIA Nemotron on Nebius Token Factory & Tavily: unifies Mem0 career memory, zero-hallucination RenderCV Typst resumes, HITL auto-apply & Calendar sync.

Comment

💡 Inspiration: Solving the Off-Campus Career Bottleneck

As a 3rd-year B.Tech Computer Science student at IIIT Dharwad (Graduating Class of 2028), navigating off-campus internships, research fellowships, and hackathons is a fragmented, exhausting full-time job:

  1. The Multi-Resume Dilemma: Students maintain 5 to 10 separate resume variants (SDE, AI/ML, Generative AI, Android, Systems) across local PDFs. Standard LLM resume wrappers hallucinate unverified skills just to match a job description, risking instant disqualification during technical interviews.
  2. Discovery Fragmentation: Opportunities are scattered across consumer boards (LinkedIn, Naukri, Indeed), student competition hubs (Unstop, Internshala), and isolated corporate ATS portals (Adobe, Google, Microsoft, Stripe, Vercel).
  3. Repetitive Portal Entry & Missed Assessments: Filling out identical profile fields across dozens of portals wastes hours every week, while time-sensitive 48-hour Online Assessment (OA) links from HackerRank or CodeSignal get buried in noisy email threads.

We built CareerOrbit—not as another chat wrapper, but as an always-on, 24/7 Autonomous Personal AI Career Operating System powered by a cost-optimized 3-tier NVIDIA Nemotron router on Nebius Token Factory and Tavily Live Search.

🧠 What It Does (5-Plane Autonomous Lifecycle)

CareerOrbit operates continuously in the cloud (even when the student's laptop is powered off) and answers five core lifecycle questions:

1. Plane 1: Dual-Layer Core Memory Vault (Mem0 + Qdrant + Canonical SQLite WAL)

  • Deterministic Canonical Vault: Locks non-negotiable ground-truth constraints (Full Name, IIIT Dharwad, B.Tech CSE, Graduation Batch: 2028, CGPA: 8.69, +91 E.164 Phone, GitHub, LinkedIn, India + Global Remote) in a strict Pydantic-validated relational store.
  • Self-Improving Semantic Memory: Ingests 12+ role-specific PDF resumes and natural-language chat updates using nebius/nvidia/Nemotron-3_5-Lightning and nebius/Qwen/Qwen3-Embedding-8B (4096-dimensional embeddings) into Mem0 + Qdrant, categorizing atomic facts into PROJECT, INTERNSHIP, SKILL, METRIC, and EDUCATION cards with full SHA-256 document provenance and Contradiction Detection (FR-008).

2. Plane 2: 24/7 Multi-Channel Opportunity Discovery (Tavily API + JobSpy + Direct ATS)

  • Tavily Search & Extract Engine: Executes domain-targeted agentic queries across unstop.com (hackathons & hiring challenges), internshala.com, and direct Big Tech portals (careers.adobe.com, careers.google.com, careers.microsoft.com, amazon.jobs) alongside Greenhouse/Lever JSON feeds and JobSpy.
  • Two-Stage Eligibility Gate: Filters out senior/manager roles, non-engineering listings, unpaid training scams, and 2025/2026-only cohorts, scoring valid Class of 2028 opportunities from $0$ to $100$.

3. Plane 3: Fail-Closed Requirement Mapping & Dynamic RenderCV + Typst Synthesis

  • Zero-Hallucination Requirement Map (FR-023 & FR-024): Before generating a single line of YAML, CareerOrbit maps extracted Job Description requirements against confirmed candidate evidence in Mem0. If a role asks for Kubernetes and the student's memory only proves Python, FastAPI, Docker, and Machine Learning, Kubernetes is explicitly flagged as UNCONFIRMED — EXCLUDED FROM RESUME.
  • Deterministic Typographical Compilation: nebius/nvidia/Nemotron-3-Ultra-550b-a55b selects the top matching research and engineering projects and synthesizes a strict, Pydantic-validated RenderCV YAML (engineeringresumes theme), which compiles via Rust-based Typst into a 100% ATS-parseable PDF in under 3 seconds.

4. Plane 4: Constrained Browser Auto-Applier & Cryptographic HITL Approval Gate

  • Two-Phase "Stage-Then-Approve" Protocol: Driven by nebius/nvidia/nemotron-3-super-120b-a12b and Playwright / Browser-Use, the agent navigates to the target application portal, fills candidate fields from the canonical vault, attaches the tailored RenderCV PDF, captures a full-page verification screenshot, and halts at a strict Human-in-the-Loop policy gate before clicking Submit.
  • One-Time Cryptographic Email & Dashboard Gate: Dispatches an interactive HTML approval briefing with HMAC-SHA256 signed one-time tokens (Approve & Submit Now, Improve Resume or Change Fields, Reject). Only after explicit human authorization does the worker complete final submission and issue an auditable receipt.

5. Plane 5: 24/7 Recruiter Inbox Watcher & Google Calendar Auto-Scheduler

  • Polls incoming recruiter communications for Online Assessments (HackerRank, CodeSignal, Mettl, Unstop) and interview invitations.
  • Nemotron-3_5-Lightning extracts the company, event category, action link, and deadline, normalizes timestamps into Asia/Kolkata (IST +05:30), automatically schedules a 2-hour study/focus block 24 hours prior to 48-hour OA deadlines, and syncs directly with Google Calendar.

⚡ How We Built It: 3-Tier NVIDIA Nemotron Routing on Nebius Token Factory

Running a 24/7 background daemon that evaluates hundreds of job postings and emails daily would rapidly exhaust compute credits if every call hit a 550B frontier model. Following the official hackathon architecture guidance, we engineered a credit-aware 3-tier model router over Nebius Token Factory's OpenAI-compatible API (https://api.tokenfactory.nebius.com/v1):

Tier Nebius Token Factory Model ID Active / Total Params Cost (Input / Output per 1M) Assigned Autonomous Subsystems
Embeddings nebius/Qwen/Qwen3-Embedding-8B 8B Dense (4096-dim) $0.01 / - Vectorizing multi-PDF resumes, project artifacts, and job descriptions in Mem0 + Qdrant.
Tier 1 (High-Frequency Daemon) nebius/nvidia/Nemotron-3_5-Lightning (fallback: Nemotron-3-Nano-30B-A3B) 3B Active / 30B MoE $0.06 / $0.24 24/7 B.Tech 2028 eligibility gating, scam detection, Mem0 atomic fact extraction, and Gmail OA/interview JSON triage.
Tier 2 (Browser & DOM Control) nebius/nvidia/nemotron-3-super-120b-a12b 12B Active / 120B Hybrid Mamba-Transformer $0.30 / $0.90 Driving Playwright / Browser-Use DOM inspection, form-field mapping, and staging verification.
Tier 3 (Deep Synthesis) nebius/nvidia/Nemotron-3-Ultra-550b-a55b 55B Active / 550B MoE $1.00 / $3.00 Deep JD-to-Memory gap analysis, Action-Technology-Result bullet synthesis, and self-correcting RenderCV YAML generation.

Mathematical Formulation of Relevance Scoring & Cost Optimization

Each opportunity $j$ that passes the deterministic hard-constraint filter $H(j, C_{\text{student}}) \in {\text{ELIGIBLE}, \text{UNCERTAIN}}$ is scored via a weighted multi-signal relevance function:

$$\text{Score}(j, u) = 100 \times \left( 0.30 \cdot S_{\text{skill}} + 0.25 \cdot S_{\text{role}} + 0.20 \cdot S_{\text{project}} + 0.10 \cdot S_{\text{cohort}} + 0.10 \cdot S_{\text{loc}} + 0.05 \cdot S_{\text{fresh}} \right)$$

By routing $85\%$ of high-frequency background token volume ($V_1$) to Nemotron 3.5 Lightning ($c_1 = \$0.06/\text{1M}$), $10\%$ ($V_2$) to Nemotron 3 Super ($c_2 = \$0.30/\text{1M}$), and reserving Nemotron 3 Ultra ($c_3 = \$1.00/\text{1M}$) for the final $5\%$ ($V_3$) of high-match resume compilations, total daily inference cost is reduced by over 11.3x compared to single-model frontier routing:

$$\text{Cost}{\text{tiered}} = \sum{k=1}^{3} V_k \cdot c_k \ll \left(\sum_{k=1}^{3} V_k\right) \cdot c_3$$

🔍 Best Use of Tavily Integration

Static job-board scrapers fail on dynamic student hackathon platforms and corporate career portals due to aggressive anti-bot walls and changing DOM layouts. We integrated the Tavily Search and Extract API (tavily-python) as a core discovery and intelligence pillar:

  1. Student Hackathon & Internship Discovery: Executes targeted search_depth="advanced" queries across site:unstop.com and site:internshala.com/internship to surface live 2028-batch hackathons and remote/India internships.
  2. Direct Big Tech Career Portal Scouting: Queries (site:careers.adobe.com OR site:careers.google.com OR site:careers.microsoft.com OR site:amazon.jobs) to discover newly posted university roles directly at the source.
  3. Pre-Compilation Company Context: Enriches candidate resume synthesis with live engineering domain context before invoking Nemotron-3-Ultra-550b.

🛠️ Challenges We Ran Into & How We Solved Them

  1. 4096-Dim Embedding Alignment & SQLite Concurrency in Cloud Containers:
    • Challenge: Qwen3-Embedding-8B outputs 4096-dimensional vectors, whereas local Qdrant collections default to 1536 dimensions. Additionally, concurrent APScheduler discovery cycles and FastAPI requests caused SQLite lock contention on Render.
    • Solution: We enforced a process-wide @lru_cache singleton for Mem0 + Qdrant with explicit 4096 dimension alignment, enabled SQLite Write-Ahead Logging (PRAGMA journal_mode=WAL; PRAGMA busy_timeout=30000;), and decoupled network I/O from database write locks.
  2. LaTeX PDF Kerning Artifacts & Table-of-Contents Noise:
    • Challenge: Extracting raw text from LaTeX-compiled PDFs produced broken ligatures (B.T ech, F ull-Stack, UA V-LLM) and polluted memory with . . . . 13 Table of Contents lines.
    • Solution: We built a deterministic de-hyphenation, regex kerning normalizer, and SHA-256 deduplication pipeline before vector indexing.
  3. RenderCV E.164 Phone Validation & Ephemeral Container Disks:
    • Challenge: RenderCV's Pydantic validator strictly requires international E.164 phone formatting (+91 ...), and cloud container restarts wipe local /tmp PDF files.
    • Solution: We added automatic +91 E.164 phone normalization, a 2-attempt self-correcting YAML compiler backed by our canonical portfolio bank, and on-the-fly PDF regeneration from persisted database YAML so split-pane previews never 404.
  4. Cloud SMTP Port 587 Blocking & Email Scanner Prefetching (FR-030):
    • Challenge: Cloud free tiers block outbound port 587 SMTP, and automated email security scanners prefetch GET links—which would accidentally trigger application submissions.
    • Solution: We built a multi-channel HTTPS (port 443) email dispatcher + an in-app HTML Email Previewer, and separated GET review rendering from POST one-time token consumption (409 Conflict on replay).

📋 Hackathon Engineering Changelog (Built During Oct 17–19, 2026)

  • Engineered the complete CareerOrbit monorepo from the ground up with an Apple Human Interface Guidelines (Cupertino Light Mode) control plane (Next.js 16 + React 19 + Tailwind CSS) and FastAPI backend.
  • Implemented the 3-Tier Nebius Token Factory Router (backend/llm_router.py) integrating Nemotron-3_5-Lightning, Nemotron-3-Nano-30B-A3B, nemotron-3-super-120b-a12b, Nemotron-3-Ultra-550b-a55b, and Qwen3-Embedding-8B with live latency/token telemetry (GET /api/v1/operations/metrics).
  • Built the Dual-Layer Memory Vault (backend/memory/mem0_manager.py & backend/api/memory.py) with multi-PDF upload, provenance tracking, and contradiction resolution (FR-008).
  • Built the 24/7 Opportunity Discovery Engine (backend/discovery/) combining Tavily Search API, JobSpy, and Greenhouse ATS JSON readers with strict B.Tech 2028 & engineering-role eligibility gating.
  • Built the Fail-Closed Requirement Mapper & RenderCV Typst Compiler (backend/resume_engine/keyword_matcher.py & rendercv_compiler.py).
  • Built the Playwright / Browser-Use Auto-Applier, Mock Corporate ATS Sandbox Portal (/mock-portals/adobe-aiml-intern-2028), Cryptographic HITL Approval Center, Gmail OA Watcher, Google Calendar Auto-Scheduler, and a 7-Phase Automated Pytest & GitHub Actions CI/CD Suite.

Built With

  • browser-use
  • docker
  • fastapi
  • github-actions
  • gmail-api
  • google-calendar-api
  • mem0
  • nebius
  • next.js
  • nvidia
  • nvidia-nemotron
  • playwright
  • pydantic
  • python
  • qdrant
  • react
  • render
  • rendercv
  • sqlite
  • tailwindcss
  • tavily
  • typescript
  • typst
  • vercel
Share this project:

Updates

Submission history