Talentic

Antigravity for Recruitment. Full-stack AI company that replaces a traditional recruitment agency with a fully autonomous agent platform that gets you the best candidates for an affordable price.

Comment

Inspiration

The hiring process is fundamentally broken. Recruiters spend 80% of their time on repetitive tasks screening resumes, scheduling calls, and writing the same emails while candidates wait weeks for responses. We saw an opportunity to apply multi-agent AI systems to automate the entire hiring pipeline, not just individual tasks. The vision: what if AI could handle everything from understanding a hiring manager's needs (even via voice) to generating a personalized offer letter, while maintaining consistent reasoning across weeks of candidate evaluation?

What it does

Talentic is an autonomous talent acquisition platform powered by four specialized AI agents:

  1. JD Assist—Transforms voice/text input into structured job descriptions with skills matrices and evaluation criteria
  2. Talent Screener—Analyzes CVs against requirements, producing ranked candidate lists with match explanations (0-100 scores)
  3. Talent Assessor—Generates role-specific questions and analyzes video responses using Gemini Vision for content. quality AND behavioral signals (confidence, engagement, body language)
    1. Offer Generator—Calculates compensation packages and generates professional offer letters.

Additional innovations:

  • Marathon Agent: Autonomous hiring orchestrator with "thought signatures" persistent memory that maintains hiring decisions across multi-week processes and self-corrects when new information contradicts earlier assessments
  • Chatbot-based sourcing across LinkedIn, GitHub, Indeed with pay-per-reveal monetization
  • Web-based phone interviews with real-time transcription and AI analysis

How we built it

Stack:

  • Frontend: Next.js 14 + TypeScript, Tailwind CSS, shadcn/ui, Zustand
  • Backend: FastAPI (Python), Google ADK with Gemini 2.5 Flash/Pro models
  • Database: Supabase (PostgreSQL) with real-time subscriptions
  • Integrations: Vapi (phone calls), SendGrid (email campaigns), Apify/Proxycurl (LinkedIn scraping), WebRTC (video recording)

Architecture:

  • Sequential agent pipeline where each agent's output becomes the next agent's context via output_key state sharing
  • ParallelAgent for concurrent CV screening
  • Supabase Realtime for live agent status updates
  • SSE streaming for conversational sourcing responses

Challenges we ran into

  1. State persistence across multi-week hiring: Traditional AI has no memory. We invented "thought signatures"—structured reasoning snapshots that follow candidates through the entire journey, enabling the Marathon Agent to make consistent decisions over weeks.
  2. Video behavioral analysis reliability: Gemini Vision sometimes hallucinated behavioral cues. We implemented configurable behavioral analysis that falls back to content-only scoring when confidence is low.
  3. Agent handoff coordination: Ensuring consistent candidate evaluation across four different agents required careful schema design and shared evaluation criteria that propagate through the pipeline.
  4. Real-time streaming with agent reasoning: Balancing responsive UI (SSE streaming) with thoughtful agent responses that take time to generate.

Accomplishments that we're proud of

  • True end-to-end automation: From voice-described job requirements to signed offer letter without human intervention (Marathon mode)
  • Multimodal AI integration: Combining CV parsing, voice transcription, video analysis, and behavioral signals into unified candidate evaluations
  • Self-correcting autonomous agents: The system updates previous decisions when contradicted by new evidence, something traditional workflow automation can't do.
  • Production-ready architecture: Real-time updates, async task processing, complete audit trails, and credit-based monetization all working together
  • Novel pay-per-reveal model: Anonymized candidate sourcing with PII gating creates a sustainable business model

What we learned

  • Google ADK's SequentialAgent and ParallelAgent patterns dramatically simplify multi-agent orchestration compared to custom implementations.
  • Behavioral analysis from video requires careful calibration; confidence thresholds matter more than raw signal detection
  • Real-time features (Supabase subscriptions, SSE streaming) transform the recruiting UX from "submit and wait" to "watch AI work."
  • Agent prompts need to be specific about output schemas to ensure reliable handoffs between pipeline stages

What's next for Talentic?

  1. Interview scheduling agent—Coordinate calendars and automatically book interviews.
  2. Salary negotiation agent—Guide back-and-forth negotiations within approved parameters.
  3. Reference check automation—Structured reference interviews with AI analysis
  4. ATS integrations—Connect with Greenhouse, Lever, Workday for enterprise adoption.
  5. Custom model fine-tuning—Train on historical hiring decisions for company-specific evaluation criteria.
  6. Candidate-facing AI—Let candidates ask questions about role, culture, and compensation before applying

Built With

Share this project:

Updates