aKriti

Grounded document intelligence with verifiable extraction and coordinated review

Overview

aKriti is an open-standard VLM repository for document intelligence.

Core focus:

• Grounded block extraction via aKritiDoc
• Confidence-aware verification workflows
• Multi-file document parsing pipeline
• Coordinated read UI with click-to-highlight block/image grounding
• Export-ready outputs (JSON, HTML, Markdown) and API-first design

This repo is intentionally kept as the VLM foundation layer; domain products such as Vinti can consume it as a downstream system.

Repository layout

aKriti/
├── core/
│   ├── akriti_doc/          # aKritiDoc schema, validators, migrations
│   └── ...
├── verification/            # 5-agent harness, voting, confidence, RL/RLVR signals
├── workbench/               # UI (split-pane + blocks list + glow + chat panel)
├── data/                    # synthetic generators, public data loaders, tokenization assets
├── evaluation/              # benchmarks, unit tests, regression artifacts
├── specs/                   # architecture specs, API contracts, roadmap, commit strategy
├── docs/                    # user/project documentation and design notes
├── .github/workflows/       # CI/quality gates (CI, lint, tests, smoke checks)
├── images/                  # project branding and screenshots
├── requirements.txt
├── .gitignore
└── README.md

LOCAL_DOCS/ is ignored by default (private/local reference materials).

Standards for v1

1. aKritiDoc as single source of truth
2. Canonical grounded structure
   • normalized bounding boxes
   • page metadata + scale transforms
   • confidence + provenance
   • reading order + history
3. Async parse flow
   • submit job
   • poll status
   • return parse artifacts + quality metadata
4. Grounded UI first
   • left: document + overlays
   • right: synchronized blocks list
   • low-confidence glow + verify queue
   • follow-up: translate / modify / ask

Long-term intent of aKriti

• Build a robust document understanding stack for mixed scripts and degraded scans.
• Keep extraction grounded with block-level coordinates, reading order, and provenance.
• Make uncertainty explicit through low-confidence signals and review workflows.
• Keep APIs and schema stable so aKriti can power multiple downstream products.
• Add verification and learning loops from human corrections.
• Scale from general documents toward specialized deployments in later phases.

aKriti Ownership Policy

• Repo-facing docs intentionally avoid naming external systems. External research is inspiration-only; the only external artifact allowed into model lineage is open weights with manifest provenance. Detailed named research notes stay outside the project repo.

Research Reference Index

• aKriti research reference index: numbered paper references [1], [2], etc. mapped to the relevant akriti-*.md docs. References are for aKriti-owned implementation only; they are not product dependencies.