Inflect-AI

Market intelligence that thinks faster than the market moves. Research any Fortune 500 company in seconds - speak it, type it, trade it.

Comment

Inflect — Project Story

Inspiration

Bloomberg Terminal costs $24,000 a year. TradingView assumes you already know what RSI means. ThinkorSwim was built for professionals, not students learning to trade.

At the same time, retail investors are flooded with noise—tabs of prices, filings, dashboards, and opinions—while real answers like “What was gross margin in Q4?” or “Why is this stock moving?” remain scattered across multiple tools.

We built Inflect because financial research shouldn't require a finance degree or a six-figure budget. Every retail investor and finance student deserves the same depth of insight that institutional traders take for granted — delivered instantly, cited accurately, and explained clearly.

The name comes from the concept of an inflection point — the moment a trend changes direction. That’s exactly what we aim to help users identify before everyone else does.

What it does

Inflect is an AI-powered financial research and portfolio platform that synthesizes SEC filings, real-time market data, technical indicators, and news sentiment into instant, cited answers.

Users can ask anything about a public company — fundamentals, technicals, earnings, or price action — and Inflect returns a grounded response with citations pulled directly from SEC EDGAR filings and structured datasets.

Each response includes a structured Trade Thesis Card:

  • Fundamental signal — derived from 10-K / 10-Q / 8-K via RAG
  • Technical signal — RSI, MACD, Bollinger Bands (TA-Lib)
  • Sentiment signal — FinBERT scoring across recent headlines
  • Verdict — HOLD / WATCH / AVOID (never BUY/SELL — we are educators, not advisors)

Beyond research, Inflect provides:

  • Portfolio dashboard with live market strip
  • Paper trading with simulated capital ($100K)
  • Voice-first interaction (ask → answer → spoken response)
  • Chat-based research (Perplexity-style threads)
  • Chart screenshot upload with AI-powered vision analysis
  • Seamless switching between voice and chat with full session context

Every answer is citation-grounded. A Validator Agent blocks any response that cannot be traced back to a retrieved source document — enforcing zero hallucination tolerance.

About the project

Inflect is designed as a full-stack financial copilot, not just a chatbot.

At its core, it is a system that:

  • Understands user intent (price check vs deep research vs trade)
  • Routes queries to the correct data pipeline
  • Retrieves structured + unstructured financial data
  • Generates grounded answers
  • Validates them against source documents
  • Returns results with optional charts, thesis summaries, and voice output

The experience is built around a unified research surface:

  • A live ticker strip powered by low-latency market data
  • A portfolio home for tracking positions
  • A research panel with citations, metrics, and chart actions
  • A dual-mode interface (voice + chat) with shared context

This turns fragmented financial workflows into a single, intelligent interface.

How we built it

Frontend

  • React 18 + TypeScript + Vite
  • Tailwind CSS + shadcn-style UI
  • Plotly.js for interactive charts
  • Zustand + TanStack Query for state/data management
  • Supabase Auth for authentication
  • Deployed via Lovable / custom domain

We built a dual-mode UI:

  • Voice Mode (mic-first interface)
  • Chat Mode (threaded research)

Both modes share full session memory (ticker, timeframe, context).

Backend

  • Python + FastAPI + Uvicorn
  • Deployed on Google Cloud Run / Railway

Core query pipeline:

User Query
   ↓
Intent Classification (Groq)
   ↓
Routing Layer:
   → Price check (Finnhub / Snowflake)
   → Research (RAG + LLM)
   → Portfolio / trade actions
   ↓
Validator Agent (citation enforcement)
   ↓
Response (text + charts + voice)

The system uses a modular AI provider abstraction layer, allowing model swapping without code changes.

Data & AI Pipeline

RAG System

  • 540K+ SEC filing chunks (10-K, 10-Q, 8-K)
  • Embedded using BGE-small
  • Stored in Pinecone across multiple namespaces
  • Section-aware retrieval:
    • Item 7 (MD&A)
    • Item 8 (Financial Statements)

Data Systems

  • Snowflake → prices, fundamentals, SEC tables
  • Finnhub → real-time market data
  • Cloudflare R2 → storage

AI Stack

  • Groq LLaMA 3.3 70B → reasoning, answer generation, thesis synthesis
  • Groq Whisper large-v3 → speech-to-text (~300ms)
  • ElevenLabs / Kokoro → text-to-speech
  • TA-Lib → technical indicators
  • FinBERT → sentiment scoring
  • moondream2 → chart image analysis
  • Wolfram Alpha → technical analysis

Chart Agent

A dedicated Chart Agent dynamically generates Plotly Python code:

import plotly.graph_objects as go
fig = go.Figure(data=[go.Candlestick(...)])
fig.show()
  • Executed inside a RestrictedPython sandbox
  • Validates output before rendering
  • Supports 9 chart types (candlestick, RSI, MACD, etc.)

Infrastructure

  • Backend → Google Cloud Run / Railway
  • Auth & DB → Supabase (Postgres)
  • Cache → Redis (Upstash)
  • Vector DB → Pinecone
  • Warehouse → Snowflake
  • Storage → Cloudflare R2

Entire stack runs at approximately $1–3/month at MVP scale.

Challenges we ran into

  • Hallucination prevention at scale: balancing strict citation enforcement with useful responses
  • Market data reliability: yfinance failed in cloud → replaced with Finnhub + Snowflake batching
  • Chart code sandboxing: required secure execution environment (RestrictedPython)
  • CORS & preview environments: handling dynamic Lovable preview domains with regex allowlists
  • Intent + ticker resolution: mapping natural language queries like “Tesla price” → TSLA
  • Latency constraints: achieving <3s voice pipeline required parallelization and fast inference
Speech → STT → Intent → RAG → LLMs → TTS

Accomplishments that we're proud of

  • Fully cited, zero-hallucination RAG system over 540K+ SEC chunks
  • Chart Agent generating 9 types of interactive visualizations
  • Trade Thesis Engine combining 3 independent signals
  • Seamless voice + chat context switching
  • End-to-end voice pipeline under 3 seconds
  • Entire AI stack running at ~$1–3/month

What we learned

  • Speed matters more than model size for user experience
  • RAG faithfulness is harder than RAG accuracy
  • Data pipelines are more important than prompts in production systems
  • Evaluation tools (RAGAS, LangSmith) must be integrated early
  • Batching Snowflake reads + parallel Finnhub calls reduced latency from 30s → ~1s

What's next for Inflect

  • Earnings transcript ingestion (8-K pipeline)
  • Vision-based chart analysis (moondream2 integration)
  • Historical backtesting
  • Mobile app with offline voice support
  • Real trading integration (pending regulatory review)

Built With

  • elevenlabs
  • fastapi
  • finnhub
  • gemini-(optional)
  • googlecloudrun
  • groq
  • groq-whisper-market-data:-finnhub-cloud:-google-cloud-run-testing:-playwright
  • languages:-python
  • pinecone
  • playwright
  • pydantic
  • pydantic-auth-&-db:-supabase-data-warehouse:-snowflake-vector-db:-pinecone-llm-/-ai:-groq
  • python
  • query
  • react
  • snowflake
  • supabase
  • tailwind-css
  • tanstack
  • tanstack-query-backend:-fastapi
  • typescript
  • typescript-frontend:-react
  • uvicorn
  • vite
  • vitest
  • wolfram-alpha-(optional)-voice:-elevenlabs
  • wolfram-technologies
  • zustand
Share this project:

Updates