TenK

Analysts read 80-page reports for 3 hours to form a single view. TenK does it in 5 minutes you talk, it answers, from the actual document. This is your always-on earnings analyst.

Comment

TenK — Your Always-On Earnings Analyst

The Problem

Every quarter, analysts spend 3–4 hours reading an 80-page 10-K to form a single view. Existing tools parse text but ignore embedded charts. None of them let you talk back.

What We Built

TenK is a voice-first AI analyst that reads a company's earnings report — including every chart and table — and lets you interrogate it in real-time conversation.

Drop in a PDF. Start talking. Interrupt mid-answer. Get spoken responses from an agent that has read everything.

How It Works

  1. Upload — Drop in any 10-K or earnings PDF
  2. Ingest — Gemini Flash 2.0 reads the full document multimodally, text and charts
  3. Talk — Gemini Live API opens a bidirectional voice stream
  4. Interrogate — Ask anything, interrupt anytime, drill into any number
  5. Export — Session ends with an auto-generated investment memo and shareable marksheet card

The Multimodal Stack

Layer Technology Role
Vision Gemini Flash 2.0 Reads PDF text + interprets charts visually
Voice Gemini Live API Bidirectional audio, native interruption
Agent Google ADK Orchestrates document context + Q&A tools
Hosting Google Cloud Run Serverless backend, WebSocket support
Storage Cloud Storage + Firestore PDF store + session state

Why Live API Changes Everything

Most voice tools are STT → LLM → TTS pipelines bolted together. Gemini Live API is natively bidirectional — the model speaks back, handles interruptions mid-sentence, and maintains conversational context throughout.

This is what makes TenK feel like talking to a human analyst, not a search bar.

Demo

"What drove iPhone revenue this quarter?" TenK: "iPhone came in at $69 billion, up 1% year over year. The real story is —" [user interrupts] "Wait, what about China?" TenK: "China was down 11% year over year. That's the key risk heading into Q2..."

No typing. No waiting. Full interruption support. Exactly like the Live Agents category describes.

What We Learned

  • Gemini's long context window is genuinely powerful for financial documents — feeding an entire 10-K in one shot works better than chunking
  • Spoken-first prompting matters enormously — the agent needs to speak like an analyst, not read a report aloud
  • Live API + ADK wiring is the hardest part — getting document context into a live voice session requires careful session state management
  • Vision on charts is underutilized — Gemini Flash caught margin trends in embedded graphs that pure text extraction would have missed entirely

Built With

The One-Liner

The analyst that never sleeps — drop in a 10-K, start talking.

Built for the Gemini Live Agent Challenge · Live Agents category · #GeminiLiveAgentChallenge

Share this project:

Updates