The first thing you notice is the spacing. You ask a coding agent for a login form, and what comes back is technically correct,a username field, a password field, a submit button. But the whitespace between the elements feels arbitrary. The button’s corner radius is a generic 8 pixels. The typography is a default system stack. It works, but it doesn’t look like your product. It looks like every other piece of software an AI has ever spat out. This uncanny valley of generated UI, the gap between functional code and branded interface, is the space Softlight is trying to own.
Based in New York, Softlight is building what it calls "infrastructure that makes coding agents design like your company" [softlight.com]. Its product aims to sit between an AI coding tool and a company’s output, teaching the agent to adhere to specific design systems, component libraries, and product conventions. The ambition is to move AI from generating generic software to producing interfaces that feel native to a brand from the first commit.
The pivot from discovery to design
Softlight’s current positioning represents a subtle but significant sharpening of its wedge. Earlier materials described the company as building "product and design agents that help teams discover what to build," a system that would learn a business and generate product ideas for managers, engineers, and designers [LinkedIn, October 2025]. That framing suggested a tool for the fuzzy front end of product development. The new focus on "infrastructure for coding agents" is more technical, more infrastructural, and arguably more urgent. It answers a concrete pain point happening right now in engineering teams experimenting with AI: the output doesn’t match the design system, creating more rework than automation.
The company’s founders,Rohan Kapoor, Aditya Raju, and Ashwin Madavan,bring product-building pedigrees from Meta, Uber, Citadel, and Faire [LinkedIn, October 2025][Aditya Raju - Founder @ Softlight, 2026]. This background in shipping at scale informs their bet. They’ve seen how costly and slow it is to maintain design consistency across large engineering organizations. Their thesis is that as AI assumes more coding work, the bottleneck will shift from writing logic to enforcing aesthetic and experiential cohesion.
The team and the $3.2 million bet
In October 2025, Softlight announced it had raised $3.2 million [LinkedIn, October 2025]. The company described its backers as product executives, CEOs, and founders associated with Stripe, Airbnb, and Plaid, though a lead investor was not disclosed. The round suggests a conviction that the problem of AI-generated design inconsistency is both real and valuable to solve.
The capital appears to be fueling a build-out. The company’s careers page lists five open engineering roles, spanning AI, backend, frontend, and product engineering, all framed as "founding" positions [Ashby]. This hiring push indicates a focus on developing the core technical infrastructure to parse design systems, understand their rules, and inject those constraints into an AI agent’s workflow.
| Founder | Role | Key Prior Experience |
|---|---|---|
| Rohan Kapoor | Co-Founder, CEO | Runway, Meta [RocketReach, 2026] |
| Aditya Raju | Co-Founder | Faire [Aditya Raju - Founder @ Softlight, 2026] |
| Ashwin Madavan | Co-Founder, CTO | Citadel, Two Sigma [RocketReach, 2026] |
An infrastructural wedge in a noisy market
The market for AI coding assistants is crowded, but Softlight isn’t building another one. Its bet is orthogonal. It’s not competing with GitHub Copilot or Cursor on raw code completion. Instead, it aims to be the layer that makes those tools,and any future coding agent,output work that aligns with a customer’s unique product language. This is a classic infrastructure play: become the plumbing that ensures quality and consistency, and let other companies fight over the user-facing application.
- The proprietary dataset. The core asset isn’t a better foundational model, but a system trained on a company’s own design tokens, component library, and existing UI code. This repository of taste is what the AI learns to mimic [softlight.com].
- The closed-group launch. Softlight is currently "available to a closed group of high-growth companies" [softlight.com]. This early-access strategy allows it to refine the product with design-mature customers who feel the pain of inconsistency most acutely.
- The expanded mandate. While the website focuses on coding agents, founder Ashwin Madavan has described building "AI agents that think like world-class product designers" [Ashwin Madavan - CTO @ Softlight, 2026]. This hints at a longer-term vision where the system could assist in the design process itself, not just the implementation.
Where the concept meets the code
The most immediate question for Softlight is one of technical depth. Can its system truly understand the nuance of a design system? It’s one thing to enforce a hex color or a font size. It’s another to comprehend and apply higher-order principles like visual hierarchy, spacing scales, or interaction states consistently across a complex application. The difference between a simple style linter and an AI that "thinks like a designer" is vast.
The company’s answer likely lies in its team’s combined experience. They’ve built products where these systems were paramount. Their early design partners reportedly range from growth-stage startups to Fortune 500 companies [LinkedIn, October 2025]. Success with these initial users will be the proof point. If they can demonstrate that their infrastructure measurably reduces the time designers and engineers spend correcting AI output, they will have a compelling wedge into the modern software development lifecycle.
The next twelve months
The coming year will be about moving from a closed group to a more open product, scaling the underlying compute as their website notes, and proving that their infrastructure can handle the diversity of design systems across their target market. The five open engineering roles suggest this build phase is fully underway. The key milestone to watch will be the transition from a private beta to a public product, and with it, the first named customer case studies that quantify the time saved or consistency gained.
Ultimately, Softlight is answering a cultural question that the rise of AI coding has made unavoidable. As software creation becomes more automated, what happens to a product’s soul? Its look, its feel, the subtle cues that make it distinct? The company is betting that in a world of infinite generative capacity, the highest value will accrue not to the thing that makes the most software, but to the thing that makes software that looks like it was made by you.
Sources
- [softlight.com] Softlight · Making coding agents design like your company | https://softlight.com/
- [LinkedIn, October 2025] Product design is one of the most misunderstood fields in tech | https://www.linkedin.com/posts/rohanneilkapoor_product-design-is-one-of-the-most-misunderstood-activity-7388972842429845504-ZnoH
- [LinkedIn, October 2025] Founding AI Engineer (NYC) at Softlight | https://www.linkedin.com/jobs/view/founding-ai-engineer-nyc-at-softlight-4319487550
- [Ashby] Softlight jobs | https://jobs.ashbyhq.com/softlight
- [Aditya Raju - Founder @ Softlight, 2026] Aditya Raju - Founder @ Softlight | https://www.linkedin.com/in/aditya-raju-1a3132104/
- [Ashwin Madavan - CTO @ Softlight, 2026] Ashwin Madavan - CTO @ Softlight | https://www.linkedin.com/in/ashwin-madavan/
- [RocketReach, 2026] Rohan Kapoor profile | https://rocketreach.com
- [RocketReach, 2026] Ashwin Madavan profile | https://rocketreach.com