Poneglyph
Poneglyph is institutional memory for development consulting projects.
Core trick
Claude vision extracts pixel-coordinate bounding boxes from handwritten Hindi forms, then an auditor agent re-reads the source images to verify each claim.
Description
Poneglyph is institutional memory for development consulting projects. Six specialized agents on Claude Opus 4.7 read scattered field evidence — scanned Hindi forms, meeting transcripts, CSV exports — catch silent commitment drift across stakeholder meetings, and draft pre-meeting briefings with every claim independently audited.
Three Opus 4.7 capabilities exercised at production quality. Scout extracts evidence from handwritten Hindi attendance forms with pixel-coordinate bounding boxes — no OCR pipeline, no fine-tuning. Archivist uses on-demand tool reads to navigate project memory the way a human consultant uses a binder, detecting silent walk-backs the project team did not acknowledge. Auditor takes the Drafter's claims and re-reads source images independently with vision, refusing to confirm anything it cannot ground — same model, two different roles, with vision, refusing to lie.
12/12 evidence extractions, 3/3 drift detection across variance runs, 89% aggregate eval pass rate across 28 test cases. Seven documented failure modes in the repo.
$200B+ flows annually from World Bank, GIZ, and UN agencies through this kind of consulting work. Built in five days. Live demo: https://poneglyph-chi.vercel.app
README excerpt
# Poneglyph **Institutional memory for development consulting projects.** A multi-agent system on Claude Opus 4.7 that reads scattered field evidence, catches silent commitment drift across stakeholder meetings, and drafts pre-meeting briefings — every claim cited, every fact independently audited. Built in five days for the [Built with Opus 4.7 Hackathon](https://cerebralvalley.ai) · April 2026. --- **Live demo:** https://poneglyph-chi.vercel.app **Video:** [3-min demo on YouTube](VIDEO_URL_HERE) **Stack:** Python 3.12 · FastAPI · Next.js 14 · Anthropic Claude Opus 4.7 --- ## Table of contents - [What it does](#what-it-does) - [The six agents](#the-six-agents) - [Three Opus 4.7 capabilities](#three-opus-47-capabilities) - [Architecture](#architecture) - [Evaluation](#evaluation) - [Failure modes](#failure-modes) - [How to run locally](#how-to-run-locally) - [Deployment](#deployment) - [Why this exists](#why-this-exists) - [What's next](#whats-next) --- ## What it does A typical donor-funded development project — say a five-year, $40M World Bank farmer producer company program in central India — generates evidence faster than any team can structure it. Field officers submit paper registration forms in Hindi. Block coordinators send WhatsApp updates with phone-camera photos. State managers log meetings in Word documents. Donors receive quarterly reports compiled by hand. Nothing links to the logframe. Nothing cross-checks itself. When the World Bank Independent