Parth Radia wants to sell you a face. Not just any face—a photorealistic one, animated in real time, that can join your Zoom meeting, respond to questions, and simulate enough micro-expressions that you might, for a fleeting moment, forget you're talking to software.
His two-person startup, Keyframe Labs, emerged from Y Combinator in spring 2026 with a pitch calibrated for the enterprise developer: turn your conversational AI into a lifelike video participant for as low as six cents a minute. The company launched its real-time video avatar platform in April 2026, betting that rock-bottom pricing and plug-and-play integrations will matter more than brand recognition in a market suddenly crowded with synthetic faces.
The product—dubbed persona-1.5-live—renders AI agents as photoreal humans during live video sessions. Keyframe claims typical rendering latency of roughly 500 milliseconds, with approximately 1 second end-to-end latency, quick enough that response times approach what the company describes as "a regular video call." It's a developer-first API designed to slot into existing conversational AI frameworks like OpenAI Realtime and ElevenLabs Agents, then handle everything visual: facial expressions, lip sync, the tiny involuntary movements that telegraph humanity. Resolution hovers somewhere between 720p and 1080p, depending on connection quality.
But the integration story may be what matters most here. Keyframe's avatars don't just appear in custom web apps—they can drop into Zoom, Microsoft Teams, and Google Meet as full-fledged participants. Using the Recall.ai Meeting Bot API, a single POST request to Keyframe's /v1/meet-bot endpoint dispatches the avatar with a persona, system prompt, and greeting. The bot enters camera-on, processes audio from other participants, and responds through whichever AI backend you've wired up. A synthetic colleague, in other words, indistinguishable from the grid of thumbnail faces until it opens its mouth.
Telepresence, Redux
The Radia brothers—Parth as CEO, Kaahan as CTO—bring résumés steeped in video synthesis and autonomous systems. Parth spent five years at Ericsson leading a frontier AI team of more than 30 people, building foundation models for voice and video generation. Before that, he worked at AT&T Labs on what the company billed as the "world's first 3D call over production networks," early-stage research into photorealistic 3D telepresence that never quite escaped the lab. Kaahan, meanwhile, put in five years at Zipline as a perception engineer, developing camera-based detection and avoidance systems for delivery drones navigating real-world obstacles.
Both hold computer science degrees from UC schools—Parth from UCLA, Kaahan from Berkeley. They founded Keyframe Labs last year and set up shop in San Francisco, though California incorporation records list Los Angeles addresses for both. A detail, perhaps, that hints at the startup's still-fluid geography.
Undercutting the Field—If the Numbers Hold
Keyframe's headline rate—as low as $0.06 per minute—positions the company well below older avatar platforms now scrambling to add real-time capabilities. The company offers a free tier with 30 minutes monthly for testing, then tiered plans scaling from 300 minutes (Starter) to 1,000 (Growth) to 3,000 (Scale), each with per-conversation and concurrency caps. Enterprise customers can request custom personas with, according to the FAQ, a 24-hour turnaround.
For context: competitors like HeyGen and D-ID have rolled out real-time streaming avatar APIs this year but don't publish simple per-minute pricing publicly. D-ID launched its V4 Expressive Visual Agents in March with enterprise positioning; HeyGen's streaming costs remain gated behind business tiers, discussed more often in developer forums than in official docs. The lack of transparent pricing from incumbents may be Keyframe's opening—or a signal that usage-based models don't pencil out at scale.
CB Insights lists a $500,000 convertible note with YC as an investor, timestamped vaguely to early spring, but no press release or SEC filing confirms the amount or terms. Dealroom and other trackers list YC but offer no financial specifics. For a company pitching enterprise sales, the opacity around capitalization is notable.
Target Verticals—and the Customers You Can't See
Keyframe's launch materials outlined target applications across corporate training, recruiting interview practice, telehealth consultations, high-touch support, sales reps, and "digital twins" for influencers. The website features anonymized testimonials from users in enterprise training, B2B SaaS support, recruiting tech, digital health, and corporate training—though no customer logos appear publicly. Not one.
For a product aimed at regulated verticals like telehealth, the absence of named customers or third-party validation is a friction point. Enterprise buyers, particularly in healthcare and finance, tend to want proof of concept from someone else first.
The technical architecture, at least, supports breadth. Developers can embed avatars via a web component (@keyframelabs/elements) that exposes connect() and disconnect() methods. A hosted deployment workflow involves creating a deployment, obtaining a publishable key, and embedding the PersonaEmbed component with allowed origins and provider credentials. Alternatively, a self-managed integration option lets customers keep LLM and text-to-speech credentials in their own infrastructure while using Keyframe's browser rendering primitives—a concession to enterprises wary of routing API keys through third-party platforms.
Emotion on Demand
The persona-1.5-live model supports anger, happiness, and sadness in addition to neutral expressions. When integrated with ElevenLabs Agents, developers can add a set_emotion client tool to their JSON config, triggering emotion shifts per response. The company maintains a library of stock personas with slugs like atlas_persona-1.5-live and caspian_persona-1.5-live. Custom personas are available for enterprise accounts, though the company doesn't specify how "custom" they get—or what guardrails exist to prevent deepfake abuse.
Keyframe claims the system supports "up to 100s of simultaneous connections," language that raises more questions than it answers. Hundreds? Low hundreds? High hundreds? Production load testing details aren't published. The documentation emphasizes "best-in-class lip-sync" and "natural micro-movements," vendor boilerplate common enough in the AI avatar space.
Timing and the Competitive Squeeze
In May, LiveKit—a popular open-source real-time communication framework—posted a LinkedIn demo featuring Keyframe avatars running on its Agents Framework. Neither company framed it as a formal partnership, but the demo suggests some ecosystem momentum. Whether that momentum translates to revenue is another matter.
The timing, though, is revealing. As of March 2026, Zoom announced AI avatar features for meetings were expected soon, alongside deepfake detection capabilities. Kaltura launched Agentic Avatars and an Avatar Video Production Studio in May, targeting enterprise knowledge management. Simli has been iterating on a real-time streaming avatar SDK with LiveKit integrations. Pika Labs introduced PikaStream 1.0 in April, claiming a "real-time visual engine for live video calls with AI agents" at 24 frames per second.
The pattern: platforms that once focused on asynchronous video generation—record a script, render a talking head—are now racing to add interactive, low-latency capabilities as conversational AI agents proliferate. The question is whether the market can sustain a dozen startups all offering roughly the same thing, or whether native features from Zoom and Microsoft will swallow the category whole.
Keyframe's documentation shows active iteration on integration guides and API endpoints through late spring. The persona-1.5-live announcement appeared on LinkedIn several months back, suggesting a late winter or early spring unveiling. Earlier blog posts from last November detailed the original Persona-1 and Persona-1-Live architectures with different latency characteristics and pricing—$0.025 per minute at the time—but those posts are now dated and the company explicitly positions persona-1.5-live as the current production model. The price more than doubled; the latency presumably improved.
The Open Question
Whether Keyframe's developer-first approach and sub-seven-cent pricing can carve out defensible market share against better-funded competitors remains unclear. The company is betting that photoreal latency under 500 milliseconds and straightforward API integrations will be enough to get enterprise developers building. Maybe it will be. Or maybe the avatar-in-meetings category consolidates around native platform features before Keyframe closes its first Fortune 500 deal.
For now, the Radia brothers are selling faces—and hoping six cents a minute is the price point that makes synthetic colleagues ubiquitous. Whether anyone will pay it at scale is the gambit still playing out.
