Frontexh

Inspiration

In fast-moving fields like AI, crypto, and global geopolitics, hundreds of papers, new technologies, and discussions appear every single day. Professionals and students don't struggle with finding information — they struggle with content overload. The most dangerous knowledge gap isn't what you don't know; it's what you knew last month that's no longer true.

Current tools force a false choice: news aggregators give you a firehose of links with zero understanding; online courses are outdated the day they publish; AI chatbots only answer questions you already know to ask. Nobody proactively teaches you what you don't know you're missing. There's no intelligent curation — no system that discovers, filters, and explains what actually matters to you.

We built Frontexh to solve this: an AI-powered EdTech platform that revolutionises learning in fast-moving domains by combining autonomous content discovery, adaptive multimodal explanations, and a persistent knowledge base that grows with you — saving hours of daily scanning and turning information overload into structured, personalised understanding.

What it does

Frontexh is an AI learning companion that operates in three stages:

DISCOVER — An autonomous discovery agent powered by Exa semantic search performs intelligent curation across papers, news, GitHub repos, and blogs. It doesn't keyword-match — it runs multi-angle semantic queries from a personalised monitoring profile, then applies aggressive editorial filtering to separate signal from noise. Only developments that materially matter make it through. What used to take 45 minutes of daily source-scanning is reduced to a curated feed delivered in minutes — a real timesaving multiplier.

LEARN — Every discovered topic is synthesized into an adaptive briefing personalised to the user's expertise level and goals. The AI chat companion — powered by MiniMax M2.5 — can search the web in real-time, fetch and read full articles, and generate explanatory images using MiniMax image-01. A real-time voice mode streams conversational responses through MiniMax Speech-2.6 TTS with sentence-level parallelization for natural, low-latency spoken explanations. Three MiniMax modalities — text, speech, and image — are creatively orchestrated by a single agent that decides which modality to deploy based on what the user is learning.

RETAIN — This is Frontexh's core innovation. The AI agent has full file system access — it can read, write, edit, list, and search files in a persistent workspace. Every topic, synthesis, source, and note is stored as real files (YAML metadata, Markdown syntheses, raw sources) organized in date-stamped folders. Your knowledge base isn't a chat history that disappears — it's a permanent, growing file system that the agent can reference, update, and build upon across sessions. This architecture makes learning compound over time and has clear potential to scale beyond the hackathon into a full knowledge management platform for teams and organisations.

How we built it

Backend (FastAPI + Python):

• CortexAgent: A custom tool-calling loop using MiniMax M2.5 via the OpenAI-compatible API. The agent has 9 tools: exa_search (Exa semantic web search), web_fetch (full page content extraction via trafilatura), generate_image (MiniMax image-01), read_file, write_file, edit_file, list_files, search_files (regex across workspace), and execute_command. The agent runs up to 10 tool rounds per conversation turn, enabling complex multi-step research workflows.
• Exa Discovery Pipeline: A 3-stage autonomous pipeline — (1) load personalised user monitoring profile, (2) run a discovery agent that executes 3-5 targeted Exa searches with date filtering, fetches and verifies sources, then outputs ranked candidates as structured JSON, (3) synthesize each candidate into a topic folder with meta.yaml, synthesis.md, and raw_sources.md. The discovery agent uses up to 20 tool rounds for thorough coverage.
• Voice Pipeline: WebSocket-based streaming — browser speech recognition → MiniMax M2.5 streaming text generation → sentence splitting → parallel MiniMax Speech-2.6 TTS calls → base64-encoded MP3 chunks streamed back to the client. Sentences are synthesized in parallel for minimal latency.
• File-based Knowledge Storage: All topic data is stored as real files on disk. Topics live in data/topics/{date}_{slug}/ folders. The agent's file tools enforce workspace sandboxing (path traversal protection) to keep operations safe.

Frontend (React 19 + Vite + TypeScript):

• Clean, responsive UI with a curated topic feed, topic detail pages with tabbed content (Summary/Audio/Diagrams), and an integrated chat sidebar.
• Voice chat interface with real-time speech-to-text and audio playback.
• Tool call visualization — the UI shows which tools the agent used (search, fetch, file ops) as interactive chips during conversations.

Creative MiniMax API Integration (3 modalities working together):

• MiniMax M2.5 (Text): Powers both the conversational agent and the discovery pipeline's reasoning/synthesis. Used via OpenAI-compatible function calling.
• MiniMax Speech-2.6 (Voice): Real-time TTS for the voice chat pipeline. Parallel sentence-level synthesis for natural conversational flow.
• MiniMax image-01 (Vision): On-demand image generation through the agent's generate_image tool — the agent autonomously decides when visuals would help explain a concept.

Deployment: Docker containers on AWS ECS Fargate with Nginx reverse proxy.

Challenges we ran into

• Agent reliability: Getting MiniMax M2.5 to consistently output valid JSON for the discovery pipeline required careful prompt engineering, <think> tag stripping, and robust JSON extraction with multiple fallback patterns.
• Voice latency: Streaming TTS one sentence at a time was too slow. We solved this by splitting the LLM's streaming output at sentence boundaries and firing off parallel MiniMax Speech-2.6 requests, so audio for earlier sentences plays while later ones are still being generated.
• File system safety: Giving an AI agent write_file and execute_command access required careful sandboxing — path traversal protection, command blocklists, output size caps, and workspace-relative resolution to prevent escapes.
• Discovery quality: Early discovery runs returned SEO spam and shallow listicles. We iterated the discovery agent's system prompt to act as a "skeptical editorial desk" with explicit kill/keep criteria, forced multi-angle search strategies, and source verification via web_fetch before inclusion — achieving genuinely useful real-world learner impact.

Accomplishments that we're proud of

• A real coding agent, not a chatbot wrapper: Our agent autonomously searches the web, reads full articles, generates images, and manages a persistent file-based knowledge workspace — up to 10 tool rounds per turn for complex multi-step research.
• Persistent memory through the file system: Unlike every other AI chat tool, Frontexh's knowledge doesn't vanish when you close the tab. The agent writes structured files that persist across sessions, creating a growing personal knowledge base — a genuinely novel approach to AI-assisted learning.
• Three MiniMax modalities creatively orchestrated: Text (M2.5) for reasoning, Speech-2.6 for real-time voice, and image-01 for on-demand visuals — all controlled by a single agent that picks the right modality for the moment. Not three separate features, but one creative, unified multimodal experience.
• Autonomous intelligent curation: The Exa-powered discovery agent independently formulates search queries, verifies sources, filters noise, and produces structured topic folders — turning content overload into personalised, accelerated learning.

What we learned

• Agent architecture matters more than model capability: A simple tool-calling loop with well-designed tools and clear system prompts outperforms complex framework-based approaches. We deliberately chose a no-framework agent design and it was the right call for reliability and debuggability.
• File-based storage is underrated: Using the file system as the knowledge layer made the agent's capabilities trivially composable — the same read_file/write_file tools that power the chat also power the discovery pipeline's output. Files are inspectable, versioned by git, and easy to debug.
• Voice UX requires architectural thinking: Good voice interaction isn't just "add TTS" — it requires streaming architecture, sentence-level chunking, parallel synthesis, and a completely different response style.

What's next for Frontexh

• Knowledge graph & collections: Connect topics into a browsable knowledge graph — group related discoveries into collections (e.g. "Attention Mechanisms", "Regulation"), visualise how concepts relate, and let the agent reason across your entire knowledge base to answer cross-cutting questions.
• Gap detection: Analyse what the user has and hasn't explored to proactively surface blind spots — "You've been tracking training methods but haven't seen the 3 new inference optimisation papers that affect your deployment pipeline."
• Template editor & onboarding presets: AI-assisted curation profile builder — pick a field preset (AI/ML, Crypto, Climate Tech), then refine sources, depth, focus areas, and update intervals through a conversational editor.
• Video explainers & mnemonic jingles: Use MiniMax Hailuo-02 for short video concept explainers and MiniMax music-2.5 for mnemonic jingles — completing the multimodal learning experience across text, voice, image, video, and music.

Built With

• aws-ecs-fargate
• docker
• exa-api
• fastapi
• minimax-image-01
• minimax-m2.5
• minimax-speech-2.6
• python
• react-19
• tailwind-css-v4
• typescript
• vite
• zustand