Apr 23, 2025 · 42m · latent-space
Tiny Teams: $6m ARR, 5m users with 4 employees — Sid Bendre, Oleve (Quizard AI/Unstuck AI)
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Sid Bendre, co-founder and CTO of Oleve, explains how a lean four-person team built and scaled multiple viral consumer AI applications to 5 million users and $6 million in ARR using deterministic AI architecture, clever infrastructure hacks, and a CPG-inspired distribution strategy.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 4.8 Guest teaching 5.4 Guest disagreement 1.6 The hosts pushing back 2.6
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 18:04 Sid rejects the hit-driven unpredictability framing
When the host asserts that relying on TikTok is inherently unpredictable and unsustainable, Sid counters by arguing that virality can be broken down into a repeatable engineering science.
Hardest push from the hosts ▶ 17:25 Host challenges lack of traditional retention moatsThe host presses Sid with a classic VC skepticism check, questioning how the business retains users when students naturally churn and can easily switch between competing copycat apps.
Biggest teaching moment ▶ 25:00 Sid details the LaunchDarkly load-balancing hackSid surprises the host by explaining how a single engineer repurposed LaunchDarkly feature rollouts into a resilient cross-region Azure OpenAI load balancer without writing complex infrastructure code.
The host holds their own ▶ 32:00 Host corrects deterministic vs variable LLM framingThe host demonstrates precise technical rigor by pointing out that prompt-routed agent workflows should be termed variable rather than truly deterministic because generative models remain inherently non-deterministic.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| The Genesis of Quizard AI and Unstuck AI | 3 | 5 | 1 | 1 | The host asks open-ended questions about how Sid got into the studio model via Neo and Z Fellows. Sid educates the host on how Quizard launched by prompt engineering the free OpenAI Codex preview model to avoid API costs before switching to 3.5. | |
| Organizational Architecture and the CPG Software Thesis | 3 | 6 | 2 | 1 | Sid outlines his structural division between direct-revenue product engineers and a central platform shadow organization run by agents. He reframes consumer software as mirroring distribution-first CPG rather than traditional SaaS. | |
| Early AI Architecture and Prompt Routing | 5 | 5 | 2 | 3 | The host asks about fine-tuning versus model routing architectures. Sid explains why they rejected model routing in favor of upstream feature extraction and deterministic prompt routing to maintain agility as base models improve. | |
| Viral Growth Tactics and Social Media Distribution | 4 | 6 | 1 | 2 | Sid details tactical social media playbooks, including man-on-the-street interviews and viral sticky-note templates adapted from duped.com that generated hundreds of millions of views. The host asks about well-funded competitors like TikTok-owned Gauth. | |
| Retention, Moats, and Brand-Led Consumer Loyalty | 6 | 4 | 3 | 6 | The host pushes back on the sustainability of social virality, arguing that TikTok hits are unpredictable and questioning the product's defensible moat beyond churn. Sid pushes back by arguing virality can be operationalized scientifically and defended through CPG-style brand equity. | |
| Platform Engineering and Cost-Saving Infrastructure Hacks | 6 | 7 | 1 | 2 | Sid explains clever infrastructure workarounds, such as using Azure AI Search with a de-indexing cron job to purge single-use novelty queries, and abusing LaunchDarkly feature rollout percentages as a managed cross-region load balancer. | |
| LLM Orchestration and Deterministic Intent Workflows | 6 | 6 | 2 | 3 | Sid outlines their philosophy of avoiding fragile multi-tool agents in favor of intent classification and deterministic execution workflows. The host technicality checks the terminology, noting that LLM-driven flows remain variable rather than strictly deterministic. | |
| Comparing LLM Frontier Models for Conversational Nuance | 5 | 5 | 1 | 2 | The host inquires about distinct model capabilities between top labs. Sid highlights Claude's clear superiority at natural tone and Gen-Z conversational nuance compared to OpenAI's generic emoji styling, and shares how they use RevenueCat metadata for client experiments. | |
| Developing the Agentic Shadow Org and Marketing Automation | 5 | 5 | 1 | 3 | The host raises skepticism about over-hyped agentic workflows, contrasting practical automation with error-prone autonomous agents. Sid agrees and explains their pragmatic approach of building deterministic scraping and automation tooling to augment human marketing staff before attempting autonomous agents. |
Statements from this episode (16)
Oleve reports $6M ARR, 5M users, and sustained profitability
“And so now across the portfolio, we do all, we do five million users. We had six million dollars in ARR and we've been profitable since the first nine months of operating.”
Oleve raised angel funding from Slack, Tinder, and Cognition leaders
“We raised from people like Cal Henderson, the co-founder of Slack, Mary Zhang, the ex CTO of Tinder and Russell Kaplan, who's the current president of Cognition Labs.”
Oleve is building a shadow organization staffed by AI agents
“What I'm trying to do with the platform is start up an internal shadow org within the company, and that's all run by agents. Effectively staffing like each business unit with agents, you know, our growth and marketing especially being the highest priority, but…”
Consumer software startups will increasingly operate like distribution-first CPG brands
“Our thesis on consumer software now is, it's gonna be It's going to reflect less or look less like our operating consumer technology companies is going to look less like traditional tech companies and more like CPG companies that were like very much distributi…”
Prompt routing scales better than model routing as base models improve
“For us, we've actually like figured that the like unit, the like The return on it was not actually that beneficial. What makes sense for us instead is, you know, like we use base models for like the final, like response. What makes sense, what makes more sense…”
Unstuck AI's launch campaign reached 250 million views in one month
“We were able to get two hundred and fifty million views in under a month just on this concept, just ripping this concept over and over again. And this is what sort of like scaled Unstuck's like initial, like viral growth.”
Viral marketing pushed Quizard AI to number four in education charts
“We've also had like one video do so well. We're working with some events so well that it brought our app all the way to number four in the education charts.”
One in four high school students has used Photomath
“I was going to say about photo math, one in every like four high school students has used photo math or has photo math on them.”
Oleve avoids third-party AI frameworks in favor of native SDKs
“Because it's so critical to know what goes into your prompt. I think it's been very important for us to like sort of like manually, like own those prompts, i.e. Like build the integrations ourselves and build those like workflows ourselves. So we haven't reall…”
Oleve periodically de-indexes unused user files to curb vector DB costs
“What we do have is we actually have a D indexer that runs like every like few days where basically we look at the index, we look at what's like being put in there and what hasn't been used in a while and we take it off. And the moment somebody like tries trigg…”
Oleve repurposes LaunchDarkly feature flags to load balance LLM endpoints
“What we do is we use the stage roll, like we use LaunchDarky flags to route between different Azure, Azure OpenAI endpoints setting up the percentages ourselves. And so now like every new request is a random, like a new like user effectively. And so it comes i…”
Build deterministic if-else LLM workflows instead of relying on prompt tuning
“I would emphasize people to focus on, like, building the most if-else condition-esque, like, LLM flows, because that gives you something to actually improve, i.e., like, if I identify an intent that's doing really shittily, but has high volume usage, I know th…”
Claude is far better than OpenAI at slang and Gen Z tone
“Whenever we need to do stuff that's a little more conversational or like a little more like a little better at slang, like Claude is way better at slang. Like whenever you ask like Claude to generate something that's like, that sounds like human or like sounds…”
Oleve repurposes RevenueCat paywall metadata tags to run mobile A/B experiments
“The revenue cat in the paywall has this metadata, like tag thing, which is actually meant for like you to like render your paywall in a very specific way, not so much anything else, but we actually use that to run our experiments as well.”
Teams should try basic automation before deploying autonomous AI agents
“The first thing we always do is, like, try to see if you can build, like, regular automation, see that scales, and then build the agent that, like, supports everything. What I mean by that is, like We don't want to over invest in trying to get an agent stood u…”