SynthMind: Autonomous Decision Intelligence

🧠 SynthMind: How I Built an Autonomous Co-Thinking & Decision Partner

💡 What Inspired Me Whenever I have to make a tough technical decision — like picking between cloud architectures, comparing AI models, or evaluating hardware trade-offs — I always found traditional AI chatbots frustrating. They just dump huge paragraphs of text at you and wait for you to ask another question. They don't help you structure your thoughts, they don't challenge your assumptions, and their math is often unreliable.

I wanted to build an AI partner that actually thinks with you, not just talks at you.

That was the spark for SynthMind: an autonomous research and decision platform that deconstructs problems using Socratic questioning, verifies live facts with Google Search, and organizes everything into interactive decision matrices with real math.

Live Web App: https://synthmind-ai-d39ed.web.app GitHub Repository: https://github.com/Anurag-tech22/synthmind

🚀 What SynthMind Does

Socratic Deconstruction: Instead of giving generic answers right away, SynthMind asks thoughtful questions to uncover hidden constraints, risks, and goals that you might have missed. Fast Real-Time Streaming: Responses stream word-by-word with sub-second latency, showing you which agent is working behind the scenes in real-time. Real Math (No Hallucinated Scores): Decision matrices use actual Multi-Criteria Decision Analysis (MCDA) math: $$S_i = \sum_{j=1}^{m} w_j \cdot c_{ij}$$ You can move interactive weight sliders on the screen to see rankings recompute instantly. Adversarial Critic: A built-in critic agent reviews the output to catch confirmation bias and unstated assumptions, giving every insight a calibrated confidence rating. Learns Your Style: It tracks your cognitive profile across four dimensions (Analytical, Detail, Speed, Visual) and adapts its communication to match how you think. Google Bonus Studio: Includes Google Gemma 2 for fast executive distillation and Google Veo 3.1 & Lyria for multi-shot video storyboards with audio cues.

🛠️ How I Built It

I designed SynthMind using clean hexagonal architecture so the core logic stays pure and modular:

Google Agent Development Kit (ADK): Orchestrates 6 specialized agents (Orchestrator, Clarifier, Ingester, Synthesizer, Adapter, Critic) through an automated state machine. Gemini 3.5 & 3.7 Flash: Handles multi-modal reasoning and structured synthesis, with Google Search Grounding enabled for live factual verification. Google Cloud Firestore & Firebase: Persists user profiles, research sessions, and matrix state in Firestore (synthmind-ai-d39ed), with the frontend hosted globally on Firebase Hosting CDN. Backend (Python 3.13 + FastAPI): Provides asynchronous Server-Sent Events (SSE) streaming, deterministic MCDA math, and sliding-window rate limiting. Frontend (Next.js 16 + React 19 + TypeScript): A sleek dark glassmorphism dashboard with DOMPurify sanitization to ensure zero XSS vulnerabilities. 🧗 Challenges I Faced & Solved Preventing LLM Math Hallucinations: LLMs struggle to do reliable weighted math on large tables. I solved this by having Gemini identify the qualitative criteria and options, while a dedicated Python math engine calculates all the weighted scores and rankings deterministically. Seamless Streaming with Multi-Agent Handoffs: Emitting real-time SSE chunks while simultaneously tracking background agent handoffs required designing an event pipeline that sends clean JSON events to the frontend. Client Security: Making sure rich user markdown and streaming HTML couldn't introduce XSS bugs by enforcing client-side DOMPurify sanitization.

🏆 Accomplishments I'm Proud Of

100% Automated Test Coverage: Wrote a comprehensive Pytest test suite with 17/17 tests passing in 3.5 seconds. Fully Live in Production: Both the Next.js frontend and Google Cloud Firestore database rules are live and running on Google Cloud with zero setup required to test. Genuinely Useful: It's not a generic chat wrapper — it gives you real, interactive tools (sliders, SWOT tables, video storyboard prompts, executive exports) you can use immediately.

📚 What I Learned

Socratic questioning at the start of a conversation makes the final synthesis 10 times more relevant and useful than jumping straight to answers. Decoupling deterministic calculations from creative LLM reasoning is the key to building reliable AI decision tools. Google Cloud Firestore is incredibly fast and intuitive for managing complex, real-time multi-agent session state.

🔮 What's Next

Adding collaborative multi-user rooms so teams can vote on weights together during strategy meetings. Adding headless browser scrapers for autonomous web research. One-click export to Google Cloud Run and Vertex AI workflows. 1:04 AM