The New Lean Startup

Can a company grow its revenue and product portfolio without continually expanding its team? AI tooling makes that question increasingly practical: smaller companies can defer fundraising and pursue profitability earlier. Oleve, which builds consumer software intended to help people live more productive and fulfilling lives, provides Sid Bendre’s concrete case.

Bendre reports that Oleve reached $6 million in annual recurring revenue, operated profitably, and generated more than half a billion social media views with four people. These are the founder’s historical results for a particular portfolio, rather than a benchmark for what any small team should achieve. The company is based in New York City; its path began with a student app and an unexpectedly successful launch video.

Quizard AI launched on January 26, 2023, with a TikTok video that went viral overnight. Bendre reports that one million views became 10,000 users in less than 30 hours. Distribution created demand almost immediately, before the company had built the operating systems it would later reuse across products.

The initial inference setup was unusual. Oleve used the original Codex model API during its free beta period, and Bendre describes cycling through ten friends’ accounts to generate outputs without initial LLM charges. Although Codex was designed for coding, the team could prompt it into open-domain conversation. This is a historical account of the original model API, not guidance for accessing today’s Codex products. Bendre attributes its retirement to abuse; OpenAI’s deprecation record documents the legacy models’ March 2023 shutdown without giving that explanation.

Usage became substantial enough that, according to Bendre, OpenAI contacted several of the accounts as top Codex users. The anecdote captures how quickly a consumer launch could produce inference demand, even when the product was still relying on a temporary beta arrangement.

After graduating, the founders moved to New York in fall 2023 and ran a back-to-school campaign built around street interviews at prominent US colleges. Bendre reports that Oleve reached its first $1 million ARR and profitability within nine months of operating. A further campaign in spring 2024 took Quizard to number six in the education charts, alongside Duolingo, Photomath, and Gauth. The chart positions here are Bendre’s reports; the recording does not specify the store, country, or chart type.

The team then applied its accumulated learning to Unstuck AI, a study companion for students. Bendre reports that Unstuck reached one million users in under nine weeks and generated more than 250 million social views in one month. By a few weeks before the talk, both products had reached the education top ten, with Unstuck at number three below Gauth and Duolingo.

The next launch moved beyond education. Bendre reports that Oleve’s third, still-stealth product took three weeks to build, had more than 1,000 users, and was already profitable. He credits blueprints prepared in advance for the short build time. That progression—from a successful launch to reusable methods to another product—underpins three parts of the operating model: operating principles, organizational structure, and AI and tool augmentation.

Leanness starts with whom the company hires. Oleve prefers leaving a position unfilled to lowering its standard, and looks for generalists with several complementary strengths. A product engineer might combine full-stack development, product judgment, and strong networking fundamentals. A marketer might code; a designer might build. Bendre’s 10X terminology describes this hiring ideal, not measured individual productivity. The practical requirement is that a person can carry work across related disciplines.

Profit supplies the shared decision criterion. Rather than treating it only as a financial result to inspect later, Oleve uses it to focus choices throughout the company. That priority becomes actionable through individual KPI ownership: everyone owns a metric, works to move it week over week, and validates decisions against it. Clear outcome ownership is intended to reduce the need for micromanagement.

For a recurring process, finishing the work also creates an opportunity to improve how it runs. Oleve asks what went wrong in the previous run and what could work better next time. Treating operational and technical problems as system failures creates a feedback loop: an issue should change the process that produced it.

Two further principles make those improvements reusable. First, consolidate workflows into existing platforms where possible, extending their use instead of introducing a separate tool for every task. Bendre calls these platforms super tools. Second, retain learning in technical playbooks and operational blueprints so the next product does not need to rediscover it. He connects Unstuck’s rapid growth to roughly a year and a half of learning from Quizard. The asset that compounds is the method, not just the product it first served.

LaunchDarkly illustrates what extending an existing tool means. Its baseline role is feature management: controlling which features are released and to whom. Oleve uses it for additional operational decisions inside its applications.

LLM traffic routing. Oleve places LaunchDarkly controls between LLM calls so operators can change where requests go. They can route or split traffic across providers in response to rate limits or strategic needs. This was especially useful when early Azure OpenAI endpoint quotas were tight and difficult to increase. The mechanism is manual, on-the-fly routing control; it is not a claim of automatic provider failover.

Ingestion-service ordering. Unstuck accepts many files, and some formats pass through a waterfall of third-party ingestion services. If one service becomes unavailable, operators can change the priority order through LaunchDarkly so another service can take precedence. The application already has multiple ingestion paths; changing their order helps keep file processing available when a dependency fails.

UI and paywall experiments. Oleve also built its own experimentation layer around LaunchDarkly. That layer lets the team start experiments without a code push. The distinction matters: configuration selects behavior supported by the application and its experimentation layer; it does not deploy arbitrary new code. Together, these uses put several common operational adjustments behind a platform the team already uses.

The engineering organization gives different people responsibility for product outcomes and shared capabilities. Bendre draws an analogy to Palantir’s Dev and Delta roles, calling Oleve’s corresponding roles harvesters and cultivators.

Harvesters are product engineers, analogous to forward-deployed engineers or Palantir Deltas. They live in product metrics, run A/B experiments, build features end to end, and work with marketing. Their responsibility extends to the product’s commercial existence: build something people want and will pay for.

Cultivators are AI software engineers building shared automation and infrastructure—the foundation Bendre calls the company’s agentic operating system. Their work spans marketing, design, and product, with the intention that a capability built once helps the company in multiple markets.

The division preserves direct product accountability while giving someone explicit responsibility for systems that make every subsequent product easier to ship and scale.

Tool spending follows the same hiring philosophy. Bendre describes the ambition as taking a 10Xer to a 100Xer: an aspirational multiplier, not a measured result. Tools should amplify a capable person’s strengths rather than compensate for hiring below the company’s standard.

In daily work, paid services support scriptwriting, campaign analysis, operations, code generation, and communications. The intended effect is to give each person support resembling a chief of staff: help carrying out the surrounding work that would otherwise consume their attention.

Reusable technical blueprints make that support cumulative:

• Code-complete templates encode lessons from two and a half years of building products for viral growth.
• Internal libraries and modules package core services, including access to LLM providers.
• Shared infrastructure supplies capabilities such as notifications and experimentation across applications.

Experimentation is an initial layer in every new product because testing consumer behavior is central to how Oleve develops it. A new launch starts with those capabilities already available.

These blueprints provide deterministic reuse: known components and processes can be applied again. The next question is whether systems that learn, adapt, and coordinate can execute the blueprints on demand. That is the transition from reusable infrastructure to the company’s broader automation ambition.

Oleve describes three stages for that transition, beginning with tooling embedded in work people already perform.

1. Human-led tooling. Build dashboards, scraping tools, and prompt-chained agents inside marketing, product, and design workflows. People continue to run the work, with better assistance for individual tasks.

2. Workflow automation. Consolidate those individual tools so the system takes over an entire process. The intended gain is bandwidth: people can move from carrying out the process to pursuing more valuable opportunities.

3. Autonomous coordination. Bring the automated workflows under a decision-making system in which several agents work together in an employee-like role. This is the larger goal: hire people for strategic insight, talent, and taste while agents carry out the company’s operations. It goes beyond simply giving each employee more tools.

The first applications target decisions about where to build and what to acquire. Bendre says Oleve has started building market-research platforms to identify promising markets for new products and support acquisitions. Agents scan daily for opportunities, successful categories, and market trends, while acquisition targets are scored for strategic fit. These are efforts already underway, rather than evidence that the entire autonomous-company model has been completed.

Growth is another target. Oleve is working to automate its virality engine through content monitoring, feedback loops triggered by particular events, and work involving creator and influencer relationships. The accompanying examples show outreach and social content—the business activity being supported, rather than a demonstration of an autonomous system running it.

The envisioned next step is one strategic person directing specialized agents: the person sets objectives, the agents execute, and the system improves over time. From there, the scope expands to a product organization, with clusters of agents serving different business units. Eventually, shared systems could let one person operate an entire portfolio of apps. This progression remains a future ambition, not a demonstrated staffing outcome.

Bendre describes the destination as a portfolio of one-person billion-dollar companies, an ambition rather than an achieved valuation. He closes by introducing TRELLIS: Targeted Refinement of Emergent LLM Intelligence Through Structured Segmentation. Bendre describes TRELLIS as a framework for scaling reliable AI user experiences to five million users, designed for products whose primary go-to-market strategy is virality. That figure describes audience scale, not a measured reliability score. The recording previews its purpose without explaining its internals, leaving reliability under sudden consumer demand as the next technical problem behind the portfolio ambition.