Project summary

Over decades, researchers in history and epigraphy have been discussing whether the 4000 year old Indus Valley writing system carried through in its legacy as pottery graffitos in Iron Age Tamil Nadu in Keeladi area and the banks of Vaigai river. Till date, most discussions were done based on people looking at two drawings side by side and comparing visually. This method seemed highly subjective. During the last 11 months, I developed Project Vatayan, a 5-step Computer Vision and Graph matching pipeline, which tests 641 Indus signs against 42 unique Vaigai base signs. 36 out of 42 are structural matches.

Preprint: https://zenodo.org/records/23859986

Code and datasets: https://www.kaggle.com/code/arkapravac366/phase-5-vatayan-grand-unified-pipeline

What are this project's goals? How will you achieve them?

It took me 11 months of continuous trial and error to even overcome pottery erosion and stroke noises while Building Phase 1. From 31 TNSDA excavation locations, I deduced and used 42 unique Vaigai and Keeladi base signs for comparing against 641 Indus signs through skeletonization, Directed Hausdorff Distance, Graph Edit Distance, Hu moments, and Euler isomerism, thus showing that a minimum of 36/42 (85.7%) structural matches prove topology’s ability to distinguish genuine continuity from coincidence.

Proper mapping of an undeciphered script is a 2.5 to 3 year grind process, similar to how Alice Kober and Michael Ventris spent years building the grids for Linear B before testing phonetic values while Linear A remains un-decipherable because of the limited corpus.

In the first year (Phase 2), I want to scale up from 42 signs to 1,500+ potsherd graffiti marks from TNSDA reports (Keeladi, Sivagalai, Adichanallur) and the complete Mahadevan Indus concordance. Manually vectorizing is too inefficient, therefore, I will use vision-language models to trace faint scratch marks directly from raw excavation photographs through the use of Unsloth. Also, permutation testing against unrelated Megalithic marks and synthetic random strokes will be done to test for false positive rates.

For Years 2 and 3 (Phase 3), the project shifts focus from isolated signs to sign sequences, i.e., constructing positional grids to test if sign order on Vaigai pottery matches that on Indus seals and to differentiate between actual signs and scribe handwriting variations. The whole data set, weights, and search tool will be made public domain.

How will you know it worked?

An open-access and version-controlled dataset of 1,500+ segmented and skeletonized Indus and Vaigai signs is released on Zenodo and Kaggle. Apart from providing the data, the similarity in structure either stands the test of scalability (p < 0.01 for comparison with random geometric controls) or fails cleanly either way, we settle a decades-long debate with statistics rather than rhetoric. In the end, I wish to publish this as a peer-reviewed benchmark and have practicing epigraphers use the search engine for new discoveries.

How will this funding be used?

This is for Year 1 in this 3-Year project plan. By achieving my minimum target of $5,000, I will be able to afford GPU computation as well as the dataset digitization process for 6 months. With $24,000 which is the maximum target, I will be able to do all the above on full time for a whole year.

- High Resolution scanning, cleanup, and Vector annotation of corpora from TNSDA excavations, Private visits, Ground Zero data acquisition, Inter-state traveling using own means: $7,000

- GPU computation for fine-tuning and permutation runs of Unsloth Vision Language model: $5,000

- Hosting and dataset storage cost, Open Access Journal fees: $2,000

- Full Year Living allowance in India: $10,000

Who is on your team? What's your track record on similar projects?

Just Me, Arkaprava Chakraborty, and I am an independent ML & Computer Vision Researcher from India. All my research in Phase 1 of the Vatayan pipeline was carried out single-handedly with no funding from labs/universities. I created the 31-site master manifest for the TNSDA (vatayan-tnsda-master-manifest that is the notebook with culmination of a 11 month research effort). I extracted 42 base signs and matched them geometrically/mathematically with 641 signs in the Indus corpus showing distinct structural evolutionary trends between the Indus and Vaigai signs (36/42 matches), thereby bringing a new perspective about where the Indus people went. I prepared the preprint on Zenodo and made all code available on Kaggle.

What are the most likely causes and outcomes if this project fails?

There could be two pitfalls. The first one is that, as I will be applying much stricter null controls while working with 1,500+ markings of potsherds, there is a chance for the structure correlation between Indus and Vaigai signs to become purely accidental. However, such a negative outcome is still useful since it will help to resolve a longstanding dispute with quantitative support. The second possible challenge is that strong erosion of photos of raw pottery will affect computer vision algorithms negatively. In this case, I will turn to drawings which have been proven in TNSDA site publications.

How much money have you raised in the last 12 months, and from where?

Nothing yet ($0). I have funded everything out of my own pocket so far.