Skip to content

sahil-tgs/gisst

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

21 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Gisst

The AI research agent that lives in your chat.


I. The Name

Gisst

Derived from "gist" — the essential point of something. The spelling is deliberate: ownable, searchable, impossible to confuse with another product. It becomes a verb: "Gisst that for me." "I Gisst-ed the whole regulatory landscape." "Just Gisst it."

Domains: Gisst.ai + Gisst.io (both available, secure both)

Pronunciation: Same as "gist." The double-s is visual only — it makes the brand scannable in a URL and distinct in search results.

Also consider acquiring: gyst.io as a redirect if available. People may drop the second S instinctively — owning the redirect eliminates lost traffic.

Agent Naming

Default agent template is called a Scout. But users name and configure their own agents. A Scout is the starting point. Users might call theirs "Hal," "Friday," "PolicyBot," or whatever fits their workflow. Gisst doesn't impose naming — it enables it.

No Branded Feature Names

Scheduled updates are scheduled updates. Knowledge bases are knowledge bases. Chat is chat. Features are self-evident; they don't need internal marketing language.


II. Product Thesis

The information economy has a fragmentation problem. The tools for finding information, discussing information, and organizing information are three separate workflows. A researcher reads an article, switches to Slack to share it, then switches to Notion to file it. A founder scans Twitter for market signals, copies links into a group chat, and manually builds a competitive intel doc.

Gisst eliminates the seams.

Gisst is an AI agent that embeds directly into the communication channels people already use — WhatsApp, Telegram, Slack, Discord — and acts as a persistent research partner. It doesn't ask users to adopt a new app. It meets them where they are. It monitors topics on a schedule, answers questions with sourced depth, and accumulates everything it surfaces into a structured, evolving knowledge base that the user owns.

The core insight: The chat thread is the research session. Every question asked, every article surfaced, every follow-up clarification is a signal about what matters. Gisst captures that signal and crystallizes it into durable knowledge — exported to Notion, Obsidian, or any vault the user controls.

The long-term bet: Knowledge bases built through Gisst become training data. Users can fine-tune their own agent, spin up specialized child agents, or license their curated datasets. The product becomes a flywheel: use generates knowledge, knowledge improves the agent, a better agent generates better use.


III. Product Principles

1. Zero New Apps. Gisst never asks the user to leave their communication channel. The chat is the interface. If you want a dashboard later, fine — but the default experience is conversational.

2. Sources or Silence. Every claim an agent makes is cited. If it can't source something, it says so. Trust is the product. There is no hallucination-tolerant mode.

3. Knowledge Compounds. Every interaction should make the system smarter. A question answered today should make tomorrow's briefing more relevant. The knowledge base grows. The agent learns. Idle agents are wasted agents.

4. User Owns the Data. Knowledge bases are exportable. Always. In standard formats (Markdown, JSON, CSV). The user can leave Gisst and take every byte of accumulated knowledge with them. Lock-in comes from value, not hostage data.

5. Opinionated Defaults, Open Architecture. Out of the box, Gisst should work beautifully for a solo researcher tracking three topics. But the system should also support a newsroom running 40 agents across 12 channels with shared knowledge bases feeding a custom model.


IV. Architecture

System Overview

┌─────────────────────────────────────────────────────────────┐
│                     Gisst PLATFORM                          │
│                                                             │
│  ┌──────────────────────────────────────────────────────┐   │
│  │              CHANNEL ADAPTER LAYER                    │   │
│  │                                                       │   │
│  │  WhatsApp · Telegram · Slack · Discord · REST API     │   │
│  │                                                       │   │
│  │  Each adapter handles auth, rate limits, formatting,  │   │
│  │  and platform-specific constraints (e.g. WhatsApp's   │   │
│  │  4096-char limit, Slack Block Kit, Telegram MD)       │   │
│  │                                                       │   │
│  │          ┌─────────────────────┐                       │   │
│  │          │   MESSAGE ROUTER    │                       │   │
│  │          │   (intent + auth)   │                       │   │
│  │          └──────────┬──────────┘                       │   │
│  └─────────────────────┼─────────────────────────────────┘   │
│                        │                                     │
│  ┌─────────────────────▼─────────────────────────────────┐   │
│  │                 AGENT ENGINE                           │   │
│  │                                                       │   │
│  │  ┌──────────────┐ ┌──────────────┐ ┌──────────────┐   │   │
│  │  │   QUERY      │ │  RESEARCH    │ │  SYNTHESIS   │   │   │
│  │  │   UNDERSTAND │→│  ORCHESTRATE │→│  & CITATION  │   │   │
│  │  └──────────────┘ └──────┬───────┘ └──────────────┘   │   │
│  │                          │                             │   │
│  │              ┌───────────┼───────────┐                 │   │
│  │              ▼           ▼           ▼                 │   │
│  │         Web Search   News APIs   Academic APIs         │   │
│  │         (Tavily,     (NewsAPI,   (Semantic Scholar,    │   │
│  │          Brave,       GDELT,     arXiv, PubMed)        │   │
│  │          Serper)      Bing News)                       │   │
│  └───────────────────────────────────────────────────────┘   │
│                        │                                     │
│  ┌─────────────────────▼─────────────────────────────────┐   │
│  │              KNOWLEDGE LAYER                           │   │
│  │                                                       │   │
│  │  ┌────────────┐  ┌──────────────┐  ┌───────────────┐  │   │
│  │  │  Vector    │  │  Graph Store │  │  Export       │  │   │
│  │  │  Store     │  │  (entities,  │  │  Engine       │  │   │
│  │  │  (semantic │  │  relations,  │  │               │  │   │
│  │  │  retrieval)│  │  topics)     │  │  Notion sync  │  │   │
│  │  │            │  │              │  │  Obsidian sync│  │   │
│  │  │            │  │              │  │  JSON/MD/CSV  │  │   │
│  │  └────────────┘  └──────────────┘  └───────────────┘  │   │
│  └───────────────────────────────────────────────────────┘   │
│                        │                                     │
│  ┌─────────────────────▼─────────────────────────────────┐   │
│  │              SCHEDULER                                 │   │
│  │                                                       │   │
│  │  Cron-based schedules (daily, weekly, custom)          │   │
│  │  Event triggers ("alert me when X happens")            │   │
│  │  Digest mode (batch updates into one message)          │   │
│  └───────────────────────────────────────────────────────┘   │
└─────────────────────────────────────────────────────────────┘

Core Components

Channel Adapter Layer — Abstraction over messaging platform APIs. Each adapter handles authentication, message parsing, rate limits, and platform-specific formatting. New channels are added as adapters without touching the Agent Engine. This is the key to "zero new apps" — Gisst is protocol-native to wherever the user already lives.

Message Router — Classifies incoming messages by intent. Is this a new research query? A follow-up? A command to schedule an update? A request to export the knowledge base? Also maps platform user IDs to Gisst accounts for auth.

Agent Engine — The core intelligence loop, in three stages:

  • Query Understanding — Disambiguates the request, identifies topic boundaries, determines depth needed.
  • Research Orchestration — Fans out to web search, news APIs (NewsAPI, GDELT, Bing News), academic sources (Semantic Scholar, arXiv), and any user-configured custom sources. Runs multiple search passes, cross-references, validates.
  • Synthesis & Citation — Compresses findings into a coherent briefing. Every factual claim links back to a source. Formats output for the target channel.

Knowledge Layer — Dual-store architecture:

  • Vector Store (embeddings of every source, every conversation turn, every knowledge base entry) for semantic retrieval.
  • Graph Store (entities, relationships, topic hierarchies) for structural reasoning — "Show me everything connected to TSMC's Arizona fab."
  • Export Engine — Transforms accumulated knowledge into structured outputs: Notion databases with properties and relations, Obsidian vaults with wiki-links and tags, or raw JSON/Markdown/CSV for custom pipelines.

Scheduler — Manages recurring briefings. Supports cron-style schedules ("every weekday at 7am"), event triggers ("whenever a new FDA approval drops in oncology"), and digest modes ("batch everything from this week into one Sunday message").

Data Flow: A Typical Interaction

User (in Telegram): "What's happening with the EU AI Act enforcement?"
    │
    ▼
Channel Adapter (Telegram) → Message Router
    │
    ▼  Intent: research query
    │
Agent Engine:
  1. Query Understanding → topic: EU AI Act, scope: enforcement
  2. Research Orchestration → searches 4 sources, retrieves 12 articles
  3. Synthesis → 300-word briefing with 5 citations
    │
    ▼
Response sent to Telegram (formatted for Telegram markdown)
    │
    ▼  (simultaneously)
Knowledge Layer:
  - Embeds articles + response in Vector Store
  - Updates Graph: [EU AI Act] → [enforcement] → [specific entities]
  - Appends to knowledge base (if auto-sync → pushes to Notion)

Technical Decisions

Stateless agents. Each agent is a stateless function triggered by an incoming message or a scheduled job. The Knowledge Layer is the only stateful component. Agents scale horizontally.

Model-agnostic. Frontier models (Claude, GPT-4 class) for synthesis and complex reasoning. Smaller, faster models for classification, intent routing, entity extraction. Swappable backend via abstraction layer.

Privacy-first. All data encrypted at rest (AES-256) and in transit (TLS 1.3). User messages processed but not stored beyond the Knowledge Layer (which the user controls). GDPR-compliant deletion: account deletion purges all data within 72 hours.

Source ethics. Respect robots.txt and rate limits. Cache frequently-accessed sources. Attribute all sources clearly. Provide original source links prominently — Gisst is a citation tool, not a content laundering tool.


V. Design System

Philosophy: "Invisible Until Needed"

Gisst is primarily experienced inside other apps. The design system isn't about a flashy UI — it's about making agent responses feel native, trustworthy, and scannable in every channel.

Brand Identity

Logotype: "Gisst" in a geometric sans-serif (Inter or Satoshi family). Clean, lowercase-friendly. The double-s gives it a visual rhythm that stands out in URLs and headers. No icon needed initially — the word itself is the brand.

Color Palette:

Token Hex Usage
Ink #0F0F0F Primary text
Slate #64748B Secondary text, metadata
Cobalt #2563EB Primary action, links, agent identity
Ember #F59E0B Alerts, breaking updates
Jade #10B981 Confirmations, source verified
Ghost #F8FAFC Backgrounds, cards
Bone #E2E8F0 Borders, dividers

Typography:

Element Spec
Headlines Satoshi Bold, 20–28px
Body Inter Regular, 14–16px
Metadata Inter Medium, 11–12px
Code/Data JetBrains Mono, 13px

Message Formatting (In-Chat)

Since Gisst lives in chat, formatting must be channel-aware but conceptually consistent.

Scheduled Update:

📡 Gisst — AI Policy
Tuesday, April 1 · Morning

▸ EU AI Act: First enforcement actions expected in Q3 as
  compliance office staffs up. (Reuters, Mar 31)

▸ US Senate introduces bipartisan AI transparency bill
  requiring model cards for frontier systems. (AP, Mar 30)

▸ China's CAC publishes draft rules for AI-generated
  content labeling. (SCMP, Mar 29)

—
3 sources · 2 new entities tracked · Knowledge base updated
Reply with a number (1–3) to deep-dive.

Research Response:

🔍 EU AI Act Enforcement — Quick Brief

The EU AI Office has begun recruiting enforcement staff,
with ~140 positions posted across Brussels and member
states [1]. First formal investigations are expected to
target high-risk AI systems in hiring and credit scoring
[2]. Companies have until August 2026 for full compliance
on general-purpose AI provisions [3].

Sources:
[1] reuters.com/tech/eu-ai-office-hiring...
[2] politico.eu/ai-act-enforcement-priority...
[3] artificialintelligenceact.eu/timeline...

💾 Added to knowledge base: "EU AI Act" → Enforcement

Event-Triggered Alert:

⚡ Gisst ALERT — Breaking

OpenAI has announced GPT-5 release date: June 2026.
Your agent "AI Landscape" flagged this based on your
tracking rules.

→ 4 related articles found
→ Reply "brief" for full analysis

Dashboard UI (Phase 2+)

For users who want a visual command center beyond chat.

Layout: Three-column on desktop. Left: agent list + status. Center: feed/conversation view. Right: knowledge base browser + graph visualization.

Card System: Every piece of intelligence is a card with a source badge, confidence indicator, timestamp, and action buttons (save, share, deep-dive, dismiss).

Dark mode first. Knowledge workers and researchers skew toward dark mode.


VI. Product Roadmap

Phase 0 — Foundation (Months 1–3)

Goal: One agent, one channel, one knowledge base. Prove the core loop.

  • Single-channel support (Telegram — fastest API, richest formatting)
  • Core agent engine with web search + 2 news APIs
  • Basic scheduling (daily/weekly, fixed times)
  • Manual knowledge base export (Markdown files via chat command)
  • Conversation memory within a session
  • Source citation on every response
  • Waitlist + 100 alpha users

Success metric: 60% of alpha users send 3+ queries/week after week 2.

Phase 1 — Multi-Channel + Knowledge Base Sync (Months 4–6)

Goal: Be everywhere the user is. Make the knowledge base a living document.

  • Add WhatsApp, Slack, and Discord adapters
  • Cross-channel agent identity (same agent, reachable from any channel)
  • Notion integration (auto-sync knowledge base to Notion databases)
  • Obsidian integration (auto-sync to Obsidian vault with wiki-links)
  • Persistent memory across conversations
  • Topic clustering and entity extraction in the Knowledge Layer
  • User-configurable source preferences (trust tiers, blocked domains)
  • Event-triggered alerts ("alert me when X happens")
  • Agent configuration and naming (users customize their agent's name, personality, depth preferences)

Success metric: 40% of users have auto-sync enabled to Notion or Obsidian.

Phase 2 — Intelligence & Depth (Months 7–10)

Goal: Make agents genuinely smarter than a person doing the same research.

  • Academic source integration (Semantic Scholar, arXiv, PubMed)
  • Multi-step research chains (agent follows leads across multiple searches)
  • Comparative analysis ("How does Company A's approach differ from B's?")
  • Graph visualization of knowledge base (viewable in web dashboard)
  • Collaborative knowledge bases (shared across team members)
  • Web dashboard MVP (read-only view of agents, schedules, and knowledge bases)

Success metric: Research depth score (source diversity, cross-referencing, analytical quality) exceeds baseline by 3x.

Phase 3 — The Flywheel (Months 11–15)

Goal: Knowledge bases become productive assets. Agents improve themselves.

  • Knowledge base as training data export (JSONL format for fine-tuning)
  • Agent fine-tuning from knowledge base (specialize an agent using accumulated knowledge)
  • Child agents (spin up a new agent pre-loaded with a knowledge base)
  • Knowledge base marketplace (opt-in: users publish or sell curated knowledge bases)
  • API access (build custom applications on top of Gisst agents and knowledge bases)
  • Multi-agent orchestration (agents that coordinate with each other)
  • Custom source connectors (user adds their own APIs, RSS feeds, databases)

Success metric: 20% of active users have created a child agent or exported training data.

Phase 4 — Platform (Months 16–24)

Goal: Gisst becomes infrastructure. Other products are built on it.

  • Self-hosted option for enterprise
  • Plugin system for custom agent behaviors
  • Knowledge base federation (link knowledge bases across organizations)
  • Real-time collaborative research sessions (multiple users + agent in one thread)
  • Agent-to-agent communication protocol
  • Compliance and audit trail features for regulated industries
  • White-label offering

VII. User Personas

"The Solo Researcher" — Priya

AI policy researcher at a think tank. Tracks regulatory developments across 6 jurisdictions. Currently uses 14 browser tabs, 3 newsletters, and a messy Notion database. Needs an agent watching each jurisdiction, a morning briefing, and a knowledge base that auto-builds her literature review.

"The Founder" — Marcus

Series A startup founder in climate tech. Needs competitive intel, investor news, and market signals. Lives in Slack. Wants an agent in his team's Slack channel that anyone can query, with a shared knowledge base the whole team references.

"The Content Creator" — Aisha

Writes a paid newsletter on AI. Needs to find stories before anyone else, understand them deeply, and have organized source material. Wants event-triggered alerts and a knowledge base she can use as her story research archive.

"The ML Engineer" — Jun

Building domain-specific models. Needs curated, high-quality datasets on narrow topics. Wants an agent that accumulates and structures data for months, then exports a clean JSONL file for fine-tuning.


VIII. Monetization

Free — "Observer"

  • 1 agent
  • 1 channel
  • 5 queries/day
  • Weekly scheduled updates only
  • Manual knowledge base export (Markdown)
  • 30-day knowledge retention

Pro — $15/month — "Analyst"

  • 5 agents
  • All channels
  • Unlimited queries
  • Custom schedules + event triggers
  • Notion + Obsidian auto-sync
  • Unlimited knowledge retention

Team — $12/user/month (min 3) — "Bureau"

  • 20 agents (shared pool)
  • Collaborative knowledge bases
  • Shared channel deployment
  • Admin controls and audit log
  • Priority source access

Enterprise — Custom Pricing

  • Unlimited agents
  • Self-hosted option
  • Custom source connectors
  • Fine-tuning and child agent capabilities
  • SLA and dedicated support
  • Knowledge base federation
  • SSO/SAML

Flywheel Revenue (Phase 3+)

  • Knowledge base marketplace transaction fee (15%)
  • Training data export as premium feature
  • Fine-tuning compute as usage-based billing

IX. Competitive Landscape

Gisst Feedly AI Perplexity ChatGPT Artifact
Lives in your chat
Scheduled briefings
Builds knowledge base Partial
Cited sources Partial
Exports to Notion/Obsidian
Fine-tunable from usage
Multi-channel
User-named agents

Gisst's moat is the compound loop: Chat → Research → Knowledge Base → Better Agent → Better Chat. No competitor connects all four.


X. Key Risks & Mitigations

Platform dependency on messaging APIs. WhatsApp and Telegram can change API terms. Mitigation: Multi-channel from Phase 1. No single channel exceeds 40% of usage. API-first architecture means a new adapter takes weeks, not months.

Source quality and misinformation. Agents could surface bad information. Mitigation: Source trust tiers, cross-referencing requirements for factual claims, explicit confidence indicators. "Sources or Silence" principle.

User data sensitivity. Knowledge bases may contain proprietary research. Mitigation: Encryption, user-controlled retention, self-hosted option for enterprise, clear data processing agreements.

LLM cost at scale. Heavy research queries are expensive. Mitigation: Tiered model usage (cheap models for routing, expensive for synthesis), aggressive caching, query complexity estimation.


XI. Taglines

Primary: "Just Gisst it."

Alternatives:

  • "The gist of everything. Faster."
  • "Research that lives where you work."
  • "Your chat already knows everything. Now it actually does."
  • "Send an agent. Build a knowledge base. Know more tomorrow."

Gisst.ai · Gisst.io

About

An AI research agent that lives in your chat. Connect it to WhatsApp, Telegram, Slack, or Discord. Get scheduled briefings, ask questions with cited sources, and auto-build a knowledge base in Notion or Obsidian. Your chat is the interface. The knowledge compounds.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors