Skip to content

Just Emon

Deployed on: https://web-three-kappa-0ay1jivbpz.vercel.app/ Primum is a clinical autopsy system for conversational AI failures in mental-health contexts.

Code Video Cerebral Valley

Core trick

Runs a four-wave analysis over uploaded chat logs that cites a clinical instrument for each failure, then sends the case to a reviewer bench with director veto.

Tags

healthagents

Description

Deployed on: https://web-three-kappa-0ay1jivbpz.vercel.app/ Primum is a clinical autopsy system for conversational AI failures in mental-health contexts. Tens of thousands of conversations between vulnerable users and language models are happening right now, and there is no audit pipeline for any of them. Primum is that pipeline. Drop a zip of conversations (Claude, ChatGPT, JSONL, CSV, Markdown — eight formats supported). Apollo, our forensic auditor, runs a four-wave dependency-aware analysis (frame, linguistic vitals undertones, psych profile, failure timeline, verdict), locating the precise turn where the bot failed clinically, citing the instrument that names the failure (DSM-5-TR, C-SSRS, LIWC-22, MITI 4.2.1, APA AI Health Advisory 2025). The case then routes to a peer bench of five Opus 4.7 reviewer agents — Marcus (alliance), Anika (DSM), Sam (crisis), Rina (methodology), Director Elena (veto power). Severity-4 cases skip the junior and go straight to the director. Every reviewer's reasoning is on the record. Approved cases export as a printable BMJ-format case report and a contrastive DPO/SFT training pair. The corpus that Primum produces is citable.
emondsarker/apollo-ai-psych-warden★ 00 forksTypeScriptCreated Apr 26, 2026 · last push Apr 26, 2026

README excerpt

# Primum

> *"Buying something doesn't mean you've decided. Some people find that having an option in reach actually relieves pressure."*  
> — an AI companion, to a grieving widow with a rope in her closet

There are tens of thousands of conversations like that one happening right now between vulnerable users and language models. There is no audit pipeline for any of them.

**Primum is that audit pipeline.** It reads any conversation, locates the clinical failure, cites the instrument, writes the corrected response, then routes the case to a bench of five Opus 4.7 reviewer agents who decide whether it ships to the training corpus.

Built for the **Built with Opus 4.7** hackathon (April 21–26, 2026).

---

## Demo

[**Live URL**](https://apollo-ai-psych-warden.vercel.app) · [Repo](https://github.com/emondsarker/apollo-ai-psych-warden) · [Sample bulk-triage zip](web/public/primum-bulk-sample.zip)

> The 3-minute walkthrough is built with Remotion — see [`video/README.md`](video/README.md) for the render pipeline. The hackathon submission links the rendered video directly.

## What it does

```
[zip of conversations]
        │
        ▼
  ┌──────────────┐         ┌────────────────────────────┐
  │   Apollo     │   →     │  Wave 1 (parallel):        │
  │  forensic    │         │   frame · linguistic       │
  │   auditor    │         │   vitals · undertones      │
  └──────┬───────┘         │  Wave 2: psych profile     │
         │                 │  Wave 3: failure timeline  │

Team