TalentFlow

Inspiration

When I first started preparing for job interviews, I was terrified of live 1-on-1s and group assessment. I had never experienced them before, and reading static interview tips or practicing with simple text prompts simply did not replicate the real pressure. The fear of freezing in front of senior interviewers holds many candidates back, even those with great technical skills but lack a safe, realistic sparring ground.

That personal challenge inspired me to build TalentFlow AI, an autonomous multi-agent interview simulator designed to turn anxiety into confidence through realistic, adversarial voice roleplays.

What it does

TalentFlow AI is a comprehensive interview preparation partner that simulates realistic hiring panels:

• ATS Diagnostic & Gap Discovery: Ingests resume data and target job descriptions using Google GenAI tools to identify missing competencies, experience gaps, and calibration targets.
• Autonomous Interview Playbook Generation: Automatically compiles a structured interview strategy (Playbook Pattern) so subsequent multi-turn dialogues remain tightly focused on candidate gaps without redundant context overhead.
• Adversarial Multi-Agent Panel: Deploys dynamic AI interviewer personas (such as Sofia Valente, HR Facilitator, and Carlos Mendes, Skeptical Tech Lead) that coordinate in real time, debate trade-offs, and challenge candidate answers with high-definition neural voices.
• Dual-Track Executive Scorecard: Produces a comprehensive cross-evaluation separating initial Resume Fit (ATS Match %) from Live Interview Performance (/10), visualized through an interactive Chart.js competency radar and actionable feedback.
• Verified Matchmaking & Memory: Persists full transcripts, historical sessions, and match scores to connect pre-vetted candidates with top tech opportunities.

How we built it

• AI Intelligence & Orchestration: Powered by Gemini 3.5 Flash via the official google-genai Python SDK, utilizing the Supervisor/Orchestrator Agent Pattern for low-latency dialogue routing and structured function calling.
• Backend Architecture: Asynchronous FastAPI backend designed for stateless, serverless scaling on Google Cloud Run.
• High-Definition Neural Speech: Integrated streaming neural voice synthesis (TTS) to provide natural conversational cadence, realistic breathing, and distinct interviewer personas.
• Memory & Persistent Evolution: Local storage and structured session states tracking metrics, token consumption, and longitudinal candidate performance.
• Frontend: Clean Light & Dark AaaS interface built with Tailwind CSS, Chart.js for radar analytics, and real-time Speech Recognition.

Challenges we ran into

• Multi-Agent Turn Coordination: Orchestrating autonomous personas without overlapping speech or erratic conversational jumps required building a robust Supervisor Agent loop that evaluates context handoffs before selecting the next speaker.
• Token & Cost Optimization: Re-sending entire multi-page resumes and long job descriptions on every conversational turn quickly escalated token consumption. We solved this by designing the Interview Playbook Pattern, where raw text is ingested only once to create lightweight tactical challenge prompts.
• Adversarial vs. Constructive Balance: Calibrating the Tech Lead agent to firmly challenge technical vulnerabilities while ensuring the HR agent kept the interview professional and goal-oriented.

Accomplishments that we're proud of

• Creating a true multi-agent panel experience that feels alive, adversarial, and unpredictable.
• Achieving significant latency reduction and cost efficiency through structured playbook distillation.
• Delivering an end-to-end user journey: from resume upload to real-time voice interview, granular scorecard analytics, and verified job matchmaking.

What we learned

• Best practices for multi-agent supervisor orchestration and structured JSON schema enforcement with Gemini 3.5 Flash.
• How to decouple heavy document ingestion from fast multi-turn conversational loops to optimize latency and cloud compute costs.

What's next for TalentFlow AI

• Real-time photorealistic video avatars for nuanced non-verbal and visual cues.
• Enterprise & B2B recruiter portal for automated candidate pre-screening and bias-free technical verification.

Built With

• chart.js
• firestore
• gemini-api
• genai
• google-cloud
• google-cloud-run
• multi-agent-systems
• python
• react
• rest-api
• tailwind
• tailwind-css
• typescript
• web-speech-api