InboxPilot

What the company appears to be: InboxPilot is a self-reported project that claims to offer human-in-the-loop AI support for SMS, email, and web chat. It positions itself as a system that normalizes inbound messages across channels into a shared inbox, uses grounded knowledge to assist AI replies, enforces safety rules before LLM calls, and allows agents to control or override AI decisions in real time. The author states it is built with Next.js, InsForge, Codex, PostgreSQL, and other technologies.

What changed: The project description does not indicate any prior version or evolution — it is presented as a new submission to a hackathon.

The single most important open question: Is there evidence of actual use, traction, or revenue beyond the author’s own account? The description contains no data on customers, adoption, or monetization.

The description states that InboxPilot is a system for managing support conversations across SMS, email, and web chat in one shared inbox. It claims to:

• Normalize and deduplicate messages from multiple channels.
• Use organization knowledge to inform AI-generated replies.
• Apply deterministic escalation rules before any LLM interaction.
• Allow human agents to approve, edit, or escalate responses.
• Provide real-time synchronization between agent inbox and web-chat visitor.
• Maintain audit trails of decisions, knowledge used, delivery outcomes, and handoffs.

The system is described as a Next.js app with backend functions in Deno, using InsForge for authentication and Postgres for persistence. It includes features like tenant isolation via PostgreSQL RLS, durable jobs, and traceable AI decisions.

Evidence: Self-reported by the author. No independent verification or demonstration of functionality beyond the project write-up.

The author claims that InboxPilot is not just a chatbot but a system that makes AI-assisted support trustworthy through:

• Safety gates.
• Grounded knowledge.
• Confidence-based behavior.
• Durable retries.
• Tenant boundaries.
• Auditable handoffs.

It positions itself as an alternative to traditional automation that "optimizes for sending the next reply" by instead making reasoning and handoffs observable.

Evidence: Self-reported. No indication of prior positioning or evolution in claims.

The author states that InboxPilot is aimed at support teams working across multiple channels (SMS, email, web chat) who need AI assistance but also require safety, control, and auditability.

It targets organizations with:

• Multi-channel support needs.
• Concerns about AI-generated responses missing context or violating policies.
• Need for human oversight in automated workflows.

Evidence: Self-reported. No evidence of actual customers or personas beyond the author’s description.

No pricing, monetization, or business model information is provided in the description.

Evidence: Not evidenced.

The system is built with:

• Next.js (App Router)
• Deno functions
• InsForge (for auth, storage, serverless functions)
• PostgreSQL with row-level security (RLS)
• pgvector for vector search
• React and TypeScript
• Codex and ChatGPT used during development

It includes features like:

• Deterministic escalation before LLM calls.
• Durable job handling with retries and dead-letter queues.
• Tenant isolation via RLS.
• Real-time delivery using authenticated events.
• Atomic RPCs to prevent race conditions.
• Knowledge traceability (pgvector chunks used in AI decisions).

The public source is available on GitHub, including setup instructions, tests, and integration suites.

Evidence: Self-reported. No evidence of production deployment or delivery performance.

There is no evidence of traction, customers, revenue, or adoption beyond the author’s own account.

The project was submitted to a hackathon (OpenAI 2026), and the author states it was built during a "Build Week iteration."

Evidence: Not evidenced.

No mention of competitors or market context is provided in the description.

Evidence: Not evidenced.

• Unverified claims: All features, functionality, and positioning are self-reported.
• No traction or revenue: No evidence of customers, usage, or monetization.
• Single-person team: The project was built by one person (Xuefeng Zhu).
• Hackathon submission: Not a product in production or market-ready form.
• No external validation: No third-party reviews, partnerships, or user feedback.

Inference: Given the lack of evidence for any commercial activity, the project may be an experimental or prototype effort rather than a viable business.

1. What is the actual use case or problem you are solving in production?
2. Have you tested this system with real users or support teams?
3. Are there any existing customers or pilot programs?
4. How do you plan to scale beyond a single developer’s effort?
5. What is your path to monetization, if any?
6. Has the system been deployed in a live environment?