# Ashr: custom models and evals for production AI

Canonical: https://mudpie.ai/companies/ashr/
Breadcrumb: [Home](https://mudpie.ai/) / [Companies](https://mudpie.ai/companies/) / [Ashr: custom models and evals for production AI](https://mudpie.ai/companies/ashr/)
Author: Ali Abouelatta (https://mudpie.ai/authors/ali-abouelatta/)
Published: 2026-09-19
Updated: 2026-09-19
Research type: Company profile
Method: Company and accelerator sources checked 2026-09-19. Product claims are attributed to their sources; this is research, not a hands-on product trial.

## What it does

Ashr builds Manifold, a platform for training, hosting, evaluating and continuously improving purpose-built open-weight models. Its [current homepage](https://ashr.io/) describes a loop that captures an organization’s documents, transcripts, logs and corrections, trains against a written rubric, benchmarks against a current frontier API and keeps retraining in the customer’s cloud. The [YC profile](https://www.ycombinator.com/companies/ashr) also describes the earlier Ashr evals product for mimicking real user behavior and testing agents.

The fit is a company with enough production data and a repeatable AI workload to justify a specialist model, but not enough research infrastructure to build the training/evaluation loop alone. Ashr’s key promise is ownership: encode the organization’s judgment in weights and run them in an environment the customer controls.

## Why I’d look closer

The public workflow is unusually disciplined. Ashr says teams define correctness, freeze a dataset, run a head-to-head benchmark and only promote a model when it clears a written bar. The docs expose a Python SDK with offline evals, server-side grading and optional production observability. That gives an evaluator a concrete first step before fine-tuning a model.

The homepage displays customer references and one company-reported example of a custom model tripling browser-agent performance on EHR workflows at more than 20 clinics. Those are company/customer claims, not an independent benchmark. The [YC founder biographies](https://www.ycombinator.com/companies/ashr) describe Rohan Kulkarni as a Berkeley-trained co-founder of Ask Geri and Shreyas Kaps as building Manifold for purpose-built models.

## What I’d ask

Which workloads are repeatable enough for a specialist model, how are labels and graders validated, and what leaves the customer cloud? I’d run a frozen holdout against the current frontier model, inspect latency/cost and failure slices, and compare a specialist model to a better prompt before committing to retraining.

## My editorial take

Ashr is a strong fit for teams that have real production traces and a measurable model-cost or quality problem. The written benchmark gate is the best part of the pitch. The buyer should demand evidence on its own workload, not generalize from the EHR claim.

## Quick facts

| Field | Sourced detail |
|---|---|
| Buyer fit | Teams with repeatable AI workloads and proprietary production data |
| Product | Evals, post-training, hosting, observability and continual learning |
| Deployment claim | Customer-cloud runs and open-weight model control; company-stated |
| Public pricing | Not exposed in the sources checked |

## Sources checked

Checked 2026-09-19.

| Source | Used for |
|---|---|
| [YC company profile](https://www.ycombinator.com/companies/ashr) | Product history, founders and launch context |
| [Ashr homepage](https://ashr.io/) | Manifold workflow, customer claims and deployment posture |
| [Ashr Python SDK docs](https://ashr.io/evals/docs/python-sdk/overview/) | Eval/observability implementation surface |
| [Ashr blog](https://ashr.io/blog/) | Public product and research context |

## Cohort context

Ashr is listed in Winter 2026. In our 2026-09-18 directory snapshot, 126 of 199 listed companies in that cohort have YC’s primary industry label B2B (63.3%). This is a current-directory comparison, not an original intake count or a performance ranking. [Nine-cohort dataset](https://mudpie.ai/research/yc-cohorts-2026-09-19.json).

## Public website snapshot

Observed 2026-09-19T16:20:00.978Z in raw homepage HTML. This records visible metadata and advertised links, not agent execution or product quality.

| Signal | Homepage observation |
| --- | --- |
| Product description metadata | Observed |
| Canonical link | Observed |
| H1 or H2 heading | Observed |
| Typed structured data | Observed |
| Docs/developer link | Observed |
| Pricing link | Observed |
| llms.txt link | Not observed in this response |
| Markdown alternate | Not observed in this response |

[Public observations](https://mudpie.ai/research/yc-homepage-links-2026-09-19.json) · [Collection method](https://mudpie.ai/research/yc-homepage-methods/README.md). Missing links here do not establish that a capability or file is absent elsewhere.


## Author disclosure

I cofound Lazyweb and publish Mudpie. This is an owner-written publication, not an independent testing organization. Research notes distinguish observations, sourced reporting and editorial judgment.