Skip to content

camerongable/thesys

BranchesTags

Repository files navigation

Thesys

Agentic RAG platform for evidence-backed idea validation.

Thesys helps founders and builders turn rough business ideas into structured validation decisions. Given a rough idea, the system can run an autonomous research sprint, discover sources and competitors, ingest evidence, generate a cited research memo, identify risky assumptions, create validation plans, and guide build / pivot / pause / kill decisions.

This project demonstrates production-style AI software engineering patterns: agentic RAG, retrieval-grounded generation, persistent project memory, human-in-the-loop workflows, source traceability, and decision-oriented AI UX.

Why Thesys Exists

Most AI idea tools generate advice in a single conversation.

Thesys is built around a different premise:

Idea validation is not a chat. It is a stateful workflow.

A founder does not just need “startup advice.” They need to know:

  • Is this idea worth pursuing?
  • Who or what am I really competing against?
  • What evidence supports or weakens the thesis?
  • What assumptions must be true?
  • What should I validate before building?
  • Should I proceed, pivot, pause, or kill the idea?

Thesys turns that process into a structured AI-native workflow.

Core Product Flow

Rough idea
→ autonomous research sprint
→ source and competitor discovery
→ evidence ingestion
→ cited research memo
→ competitor / substitute map
→ ranked assumptions
→ validation plan
→ experiment results
→ decision recommendation

The goal is not to replace founder judgment. The goal is to make founder judgment more evidence-backed, structured, and repeatable.

What Makes This Different From a Chatbot

Thesys does not store a loose chat transcript as the primary product object.

It stores structured strategic state:

  • project thesis
  • target customer
  • primary problem
  • evidence sources
  • cited findings
  • open questions
  • competitors and substitutes
  • assumptions
  • risks
  • validation plans
  • experiment results
  • recommendations
  • decisions
  • strategic updates

This allows the system to reason over a project over time instead of answering disconnected prompts.

Key Features

Guided Strategic Workspace

Each project has a lifecycle:

Idea → Research → Evidence → Assumptions → Validation → Decision

The UI keeps the user focused on:

  • current verdict
  • next best action
  • project stage
  • risk level
  • confidence level
  • evidence health
  • riskiest assumption

Autonomous Research Sprints

A user can start with a rough idea and run a research sprint.

The system:

  1. creates a research plan
  2. discovers relevant sources
  3. identifies competitors and substitutes
  4. ingests evidence
  5. retrieves relevant context
  6. generates a cited research memo
  7. identifies gaps and assumptions
  8. recommends what to validate next

Source and Competitor Discovery

Thesys can discover and classify:

  • direct competitors
  • indirect competitors
  • substitute behaviors
  • incumbent platforms
  • adjacent solutions
  • manual workarounds

The system emphasizes that the real competitor is often not another startup. It may be ChatGPT, Notion, spreadsheets, Reddit, YouTube, or an existing manual workflow.

Evidence-Backed Findings

Research output is grounded in ingested sources.

The Evidence page separates:

  • supported findings
  • open questions
  • source records
  • evidence gaps

The app is designed to make uncertainty visible instead of hiding it behind confident AI prose.

Assumption Ranking

Thesys turns research into operational validation priorities.

Each assumption can be ranked by:

  • risk
  • confidence
  • evidence strength
  • validation status
  • recommended validation method

This helps the user identify the assumption most likely to kill the idea if false.

Validation Planning

For the riskiest assumptions, Thesys can generate validation assets:

  • customer interview scripts
  • screener questions
  • survey questions
  • landing page copy
  • outreach messages
  • success criteria
  • failure criteria
  • results rubric

The goal is to move from “interesting research” to “what should I test next?”

Decision Workflow

Thesys supports structured decisions:

  • proceed
  • pivot
  • pause
  • kill
  • continue research

Decisions can include:

  • rationale
  • supporting evidence
  • unresolved risks
  • revisit triggers
  • experiment results

AI Engineering Concepts Demonstrated

This project demonstrates several modern AI engineering patterns:

  • Agentic RAG workflows
  • Retrieval-grounded generation
  • Source discovery and evidence ingestion
  • Competitor and substitute discovery
  • Persistent memory beyond chat history
  • Human-in-the-loop approval
  • Cited findings and open-question tracking
  • Assumption and risk scoring
  • Validation experiment generation
  • Decision logging and recommendation history
  • Stage-aware AI product UX
  • Multi-step workflow orchestration
  • Evidence traceability

Architecture Overview

Frontend
  ↓
API Layer
  ↓
Project Memory / Domain Model
  ↓
Research Sprint Orchestrator
  ↓
Source Discovery / Competitor Discovery
  ↓
Evidence Ingestion Pipeline
  ↓
Retrieval Layer
  ↓
LLM Gateway / Model Abstraction
  ↓
Cited Synthesis / Critique / Memory Update

Core System Layers

1. Web UI

The frontend provides a guided project workspace with tabs for:

  • Overview
  • Research
  • Evidence
  • Competitors
  • Assumptions
  • Validation
  • Decisions

The UX is verdict-first: users see the current recommendation and next best action before digging into evidence or process details.

2. API Layer

The backend exposes project and workflow APIs for:

  • project lifecycle state
  • research sprint execution
  • source discovery
  • competitor discovery
  • evidence ingestion
  • artifact generation
  • assumption scoring
  • validation planning
  • decision recording

3. Agentic Research Workflow

The research workflow is designed as a multi-step agentic RAG process.

Research Planner
→ Source Discovery
→ Competitor Discovery
→ Evidence Ingestion
→ Retrieval
→ Gap Detection
→ Synthesis
→ Critique
→ Human Approval
→ Project Memory Update

The system does not blindly update strategic state. Major updates can be reviewed before being committed to project memory.

4. Retrieval Layer

The retrieval system supports evidence-grounded generation.

Core responsibilities:

  • ingest source content
  • extract text
  • chunk documents
  • generate embeddings
  • store source metadata
  • retrieve relevant evidence
  • link claims to sources
  • surface open questions when evidence is weak

5. Persistent Strategic Memory

The app models strategy as durable state, not chat history.

Key entities include:

Project
Thesis
EvidenceSource
EvidenceChunk
Finding
OpenQuestion
Competitor
Assumption
ValidationPlan
ExperimentResult
Decision
StrategicUpdate
ResearchSprint
ResearchMemo

Example Agentic Research Sprint

Given this idea:

An AI assistant for independent fitness coaches that helps them manage client programming, check-ins, and workout adjustments.

Thesys can produce:

Verdict:
Do not build a generic fitness AI assistant yet. Validate whether independent coaches will pay for workflow automation around check-ins and client adherence.

Best wedge:
Client check-in summarization and program adjustment support for solo online coaches.

Top competitors/substitutes:
Trainerize, TrueCoach, Google Sheets, manual WhatsApp/Instagram check-ins, ChatGPT.

Riskiest assumption:
Independent coaches will pay for a dedicated AI workflow instead of continuing manual client management.

First validation test:
Interview 5–10 independent coaches and test willingness to pay for automated check-in summarization.

Tech Stack

Frontend

  • Next.js
  • TypeScript
  • Tailwind CSS or equivalent styling system

Backend

  • FastAPI
  • Python
  • PostgreSQL
  • pgvector

AI / Orchestration

  • LangGraph-style workflow orchestration
  • LiteLLM-style model abstraction
  • LLM provider APIs
  • Embeddings
  • Vector retrieval

Infrastructure

  • Docker
  • Docker Compose
  • Environment-based configuration

Local Development

Prerequisites

  • Node.js
  • Python 3.11+
  • Docker
  • Docker Compose
  • PostgreSQL with pgvector support
  • LLM provider API key

1. Clone the Repository

git clone <github-url>
cd thesys

2. Configure Environment Variables

cp .env.example .env

Example environment variables:

# Application
APP_ENV=development
FRONTEND_URL=http://localhost:3000
BACKEND_URL=http://localhost:8000

# Database
DATABASE_URL=postgresql://postgres:postgres@localhost:5432/thesys

# LLM / Model Gateway
OPENAI_API_KEY=your_key_here
ANTHROPIC_API_KEY=your_key_here
LITELLM_API_KEY=your_key_here

# Embeddings
EMBEDDING_MODEL=text-embedding-3-small

# Optional
LANGFUSE_PUBLIC_KEY=
LANGFUSE_SECRET_KEY=
LANGFUSE_HOST=

3. Start Infrastructure

docker compose up -d

4. Install Backend Dependencies

cd backend
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

5. Run Backend

uvicorn app.main:app --reload --port 8000

6. Install Frontend Dependencies

cd frontend
npm install

7. Run Frontend

npm run dev

Open:

http://localhost:3000

Suggested Demo Flow

Use this flow when reviewing the project:

  1. Create a project from a rough idea.
  2. Run an autonomous research sprint.
  3. Review the research result.
  4. Inspect discovered sources and competitors.
  5. Review supported findings and open questions.
  6. Open the assumption matrix.
  7. Create a validation plan for the riskiest assumption.
  8. Log experiment results.
  9. Record a proceed / pivot / pause / kill decision.

Product Design Principles

Verdict First

The app should answer:

What should I do next?

before showing the user the machinery behind the recommendation.

Evidence With Receipts

Every important factual finding should link back to evidence or be marked as an open question.

Strategic State Over Chat

The system should model durable strategic objects rather than relying on a chat transcript.

Human-in-the-Loop

The app should assist with research and synthesis, but users remain responsible for approving major strategic updates and decisions.

Validate Before Building

The core workflow is designed to prevent premature building by identifying the most important assumptions to test first.

Current Status

This is a V1 proof-of-concept.

Implemented or demonstrated:

  • project lifecycle workflow
  • autonomous research sprint
  • source discovery
  • competitor/substitute discovery
  • evidence ingestion
  • cited research memo generation
  • assumption ranking
  • validation planning
  • decision recommendation
  • guided UI around next best action

Planned future work:

  • recurring market monitoring
  • team collaboration
  • multi-project portfolio dashboard
  • integrations
  • advanced evaluation dashboard
  • consultant / product discovery / investor workflow packs

Roadmap

V1: Autonomous Research and Validation Copilot (current)

Core promise:

Give me a rough idea, and I’ll investigate the market, identify competitors, gather evidence, and tell me what to validate next.

Key capabilities:

  • autonomous research sprint
  • source discovery
  • competitor discovery
  • evidence ingestion
  • agentic RAG synthesis
  • assumption extraction
  • validation plan generation
  • decision recommendations

V2: Living Strategic Intelligence Platform

Core promise:

Continuously track strategic evidence, monitor changes, help teams make decisions, and manage portfolios of opportunities over time.

Potential capabilities:

  • recurring watchlists
  • market monitoring
  • strategic alerts
  • team workspaces
  • collaboration
  • portfolio comparison
  • integrations
  • advanced evals
  • workflow packs for consultants, PMs, investors, and innovation teams

Lessons Learned

This project explores several product and engineering lessons:

  1. AI product value comes from workflow, not just generation. A generic AI answer is easy to produce. A persistent, stateful workflow is harder and more valuable.

  2. RAG needs UX. It is not enough to retrieve sources. Users need to understand what the evidence supports, what remains uncertain, and what action follows.

  3. Agentic systems need constraints. Human approval, inspectable steps, and source traceability are essential for trust.

  4. Strategic products need opinionated outputs. Users do not only need summaries. They need judgment, tradeoffs, and next actions.

  5. Memory should be domain-specific. Long-lived state should model the real workflow: thesis, evidence, assumptions, experiments, and decisions.

Portfolio Notes

This project was built to demonstrate AI systems and product engineering skills relevant to modern AI software roles:

  • full-stack AI application architecture
  • agentic RAG workflow design
  • retrieval-grounded generation
  • persistent memory modeling
  • human-in-the-loop AI workflows
  • evidence-backed UX
  • product-oriented AI system design
  • workflow orchestration
  • stateful decision support

License

This project is currently provided for portfolio and educational purposes.

Author

Built by Cameron Gable.

About

No description, website, or topics provided.

Resources

Readme

Security policy

Security policy
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages