StartNet AI
Small vision and tabular models trained on real cluster data, the language layer rented from whoever is best that quarter, the data access always owned.
What StartNet builds and sells with industry.
Small vision and tabular models trained on real cluster data, the language layer rented from whoever is best that quarter, and the data access owned. The nooru.ai cohort sits here (a hundred founders, one flywheel cycle), alongside GCC partnerships and the domain clusters, Namakkal first.
How we build
BloombergGPT (50 billion parameters, a 363-billion-token proprietary corpus) was beaten by GPT-4 with zero financial pretraining at all. Seven medical LLMs tested against their own base models won 12.1% of the time and were significantly worse 38.2% of the time. Krutrim spent ₹10,000 crore and left model-building inside thirty months.
And the counter-evidence is what makes the corrected thesis work: the same Bloomberg paper shows small fine-tuned models beating GPT-4 by 26 points on domain-specific entity recognition, and a 2026 industrial study found retrieval-augmented generation the most cost-efficient adaptation method, with 3-billion-parameter open models matching frontier ones.
Small models win where the data is proprietary and the task is narrow. Frontier models win at language. We build the first and rent the second: training small vision and tabular models, renting the language layer from whoever is best that quarter, and owning the data access.
Domain AI
No hardware, no new model, no IT literacy required of the farmer: the deliberately unglamorous first product.
Cluster 1: poultry
The Competition Commission of India’s 2022 cease-and-desist order against the NECC removed the cluster’s price-coordination mechanism, and nobody replaced it with a forecast, an unserved need created by a regulator, the most durable kind. Cluster losses were reported at ₹5 crore a day during the March 2026 price crash.
Cluster 2: knitwear
A mature product category with real vendors and real prices, a harder sale than Namakkal, taken second on purpose.
Cluster 3: home textiles
Third in sequence, each cluster a harder sale than the last.
An open-source computer-vision system has already run in 150+ Namakkal farms for seventeen years. Establishing whether that is a partner or a competitor comes before the first meeting, not after.
nooru.ai
A hundred founders, one flywheel cycle, real outcomes: a curated, diagnostic-gated cohort run through a structured programme, with direct access to StartNet’s investor network. It is the one name in the house that could stand alone if the cohort works, and it is treated with the lightest endorsement in the portfolio for exactly that reason.
Apply to the cohortGCC partnerships
Domain talent pipelines and live projects, sourced directly from StartNet Research labs and Vibe cohorts.
A named pathway from a specific college to a specific GCC, not a generic placement drive.
Real GCC problems, run inside a lab or a cohort, not a simulated brief.
Search across the full StartNet X directory, filtered by domain and diagnostic result.
AI transformation: assessment, roadmaps, build, support. Custom development is opportunistic work, deliberately not the business: it has no leverage and it will eat the team if it becomes one.
Start an assessment