Signal Thesis

Signal Thesis makes complex investment research simple with personalized AI conversations that adapt in real time, test your thesis, and uncover evidence you can trace and trust.

Comment

Signal Thesis

Expert conversations. Evidence-backed decisions.

Inspiration

An investment thesis starts with a compelling story. The difficult part is discovering whether that story holds up when you speak with people who know the business.

Strong customer retention, for example, could reflect a product people love, high switching costs, or both. Understanding the difference requires good questions, expert conversations, and evidence that a team can revisit.

We built Signal Thesis to connect that process: from the assumptions behind a thesis to the questions asked, the answers received, and the findings a researcher is prepared to defend.

What it does

Signal Thesis helps investment teams investigate assumptions through expert interviews and review the resulting evidence.

The consultant workspace connects:

Research projects, private theses, and versioned source material.

Expert invitations, participant consent, and withdrawal.

AI-generated question proposals grounded in selected sources and expert context.

Explicit application and separate approval of interview plans.

Transcript corrections, exact citations, source-linked findings, human reviews, and current reports.

When a source changes or evidence becomes unavailable, dependent plans and findings lose their validity. The workflow keeps reviewers connected to the evidence they are actually using.

Our Expo participant app connects to a hosted video-avatar interview backend. Controlled Android testing has demonstrated avatar video, audible speech, microphone transcription, mute controls, and background capture cleanup. Public voice and camera access remain disabled while the remaining device checks are completed.

How we built it

We built a React and Vite consultant workspace and an Expo React Native participant app around a shared TypeScript contract. The backend uses Node.js, Express, and Zod.

Clerk handles identity. Google Cloud Run hosts the application, Cloud Build produces deployable images, and Cloud SQL PostgreSQL stores authoritative research records, permissions, consent, and audit events. Independent scheduled retention removes expired project data.

Gemini through Vertex AI powers source-grounded question planning and finding proposals. LiveKit and Boson connect the conversation pipeline, including participant audio, speech recognition, and streamed avatar audio and video.

We also implemented an opt-in SurrealDB Agent Memory integration on an isolated development branch. It adds persistent research projection, private and shared handoff notes, and recall resolved against current authorized evidence. PostgreSQL remains authoritative. This memory implementation is locally tested and awaits integration and live-provider qualification.

Challenges we ran into

The hardest work was coordinating changes over time.

A source can be corrected while an AI request is running. A participant can withdraw while a worker is processing a response. A delayed result must respect the current state of the research.

Memory introduced another difficult case: a deletion request can succeed while an older write is still in flight. We built durable cleanup records and regression tests so the application preserves that uncertainty until it can be resolved.

On mobile, we worked through authentication refresh, microphone ownership, background transitions, and resource cleanup. Testing on an actual phone helped distinguish a successful server response from audio and video the participant could see and hear.

Accomplishments that we're proud of

Deployed a working research preview with real Clerk authentication and persistent PostgreSQL records.

Qualified source-grounded question planning with actual Vertex calls, explicit review and approval, reload persistence, and regeneration after a source change.

Demonstrated an initial live finding-generation, human-review, JSON-export, and source-invalidation workflow.

Ran a real hosted synthetic avatar conversation and separately demonstrated Android avatar playback, user-confirmed hearing, microphone transcription, and mute controls.

Implemented and independently reviewed the isolated memory branch, with 699 automated tests passing, including 152 new memory tests, plus browser workflow checks.

These milestones give us working evidence for the research workflow while making the remaining release work clear.

What we learned

Trustworthy research depends on the details surrounding each conversation: who contributed, what they agreed to, which source version supports a claim, and what happens when that evidence changes.

We also learned that automated tests, real provider calls, and physical-device checks answer different questions. Combining them helped us find problems that a successful demo or a passing test suite alone could miss.

Human review became a central design decision. AI proposes questions and interpretations; researchers decide what to apply, approve, and accept.

What's next for Signal Thesis

Our immediate priorities are reliable finding regeneration after source edits, integration and live qualification of SurrealDB Agent Memory, and completion of the mobile conversation checks.

That includes rejoining, withdrawal, network and audio-route recovery, multi-question progression, and the newly implemented camera, avatar-selection, and caption controls. iOS qualification remains ahead.

We will also expand original-document ingestion, evaluate agent behavior with Era, and publish a complete captioned product demonstration.

The goal is a connected research loop: prepare from sources, learn from experts, review the findings, and keep every conclusion traceable as the evidence evolves.

Built With

Share this project:

Updates

Submission history