# Fynex Copilot:

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Louis Langanay](https://cerebralvalley.ai/u/LouisLanganay)
- **GitHub:** https://github.com/LouisLanganay/fynex-copilot
- **Demo video:** https://youtu.be/Iql8snmdIvM?si=Be9MO78MZNO5v3L7
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1611

Fynex Copilot tells an individual investor what changed in each stock of a portfolio, with the source. The user drops a CSV. For each holding, the app fetches the latest official documents (SEC EDGAR, the French AMF, ESEF annual reports for other European issuers), and Muse Spark explains them in plain language as a structured card. Every claim links to its exact passage, highlighted in the document text, with the page number and a link to the original. The results open on the largest lines with a material change, ordered by portfolio weight only, by code, with no model call.

A triage agent with three tools (list_candidate_documents, read_document_start, select_documents) picks the documents to analyze: 2 to 4 tool calls per holding in the recorded run. The analysis returns through a strict JSON schema. The server then checks every quote word for word against the document, and every number against the quotes or the weights computed by code. A claim that fails is dropped. If nothing survives, the model gets one retry with the list of problems. If that fails too, the AI summary is withheld.

Earnings calls are read too. Microsoft's release of Oct 28, 2026 says its outlook comes on the earnings call, so the guidance is only in the audio. Muse Voice Transcribe (muse-voice-transcribe-1.0) turned the official 57 minute recording into a transcript with timestamps: the recording is cut at its pauses into 173 clips, each keeping its offset. Muse Spark reads that transcript like a filing, and the same verifier checks every quote. A chip such as [10] 27:14 opens an audio player on Microsoft's own recording at that second. The app links to Microsoft's file, as published on its investor event page; it does not host the audio, and only the quoted passages are stored in the repository, not the full transcript. Quote timestamps were checked against Microsoft's own captions: median error 1.0 s on 988 phrases.

The app explains documents and never recommends buying, selling or holding a security. A question asking for advice is refused before any model call, and the model's own text is checked too.

Recorded live run of Oct 29, 2026, on a fictional sample portfolio, with the Microsoft card and its earnings call recorded on Nov 2, 2026: 12 holdings, 51 claims, 118 quotes, all 118 verified, $0.84 of Meta models, of which $0.17 of speech to text. One card needed the retry, because its first answer was cut off and failed the JSON check. The repository has 215 unit tests that run with no network and no model.

What existed before: Fynex, the company (fynexapp.com), a portfolio tracker for European retail investors. What was built during the hackathon: this repository and all its code. The first commit is dated Oct 27, 2026, and the application code was committed on Oct 29, Oct 30 and Nov 2. No Fynex code is used, and the video does not show the Fynex product. The only link is that the CSV importer also accepts the column layout Fynex uses for portfolio tables. The code was written with an AI coding agent (Claude), listed as co-author on the commits.

The demo video replays the recorded run with no model call, and its narration is text-to-speech. The voice heard for about ten seconds in the middle is Microsoft's official earnings call recording, played from the quoted moment and labelled on screen. A live run needs a Meta Model API key.

## More from Meta Global AI Developer Hackathon

- [Anubhav](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1608)
- [FEN meta synthesis](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1609)
- [Noticer](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1610)
- [Devcules](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1612)
- [Sahil Singh](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1613)
- [KisanAi](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1614)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1611. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt