Aakash GuptaCollege Dropout Raised $20M Building AI Tools | Cluely, Roy Lee
CHAPTERS
- 0:00 – 2:07
Cluely’s product direction: let users’ behavior set the roadmap
The conversation opens on how Cluely decides what to build next. Roy explains they launched a broad "Interview Coder for everything" and now use massive inbound usage data and customer complaints to identify the stickiest use cases and prioritize fixes.
- •Shipped a general-purpose version first, then iterated based on real usage
- •Millions of daily requests reveal what people actually do with the tool
- •Customer emails provide a direct backlog of pain points
- •Roadmap emerges from signal density rather than formal planning
- 2:07 – 3:52
Roy Lee’s origin story: provocative by nature, expelled, and refocused on startups
Roy describes a lifelong tendency to be polarizing and outspoken, which helped him gain leadership and friends but also caused repeated discipline. He recounts being rescinded from Harvard and later getting pushed out at Columbia, framing these as catalysts for betting on company-building instead of school.
- •Self-identity as intentionally provocative and polarizing
- •Detentions/suspensions as recurring outcome of being outspoken
- •Harvard rescinded after mass reporting and suspension
- •School expulsion hardened his view that he may not need school
- •Decision to channel unconventional streak into building a company
- 3:52 – 4:25
Cluely’s early traction: rapid ARR growth and “top of timeline” distribution
Aakash and Roy discuss Cluely’s meteoric growth—near $6M ARR just weeks after writing the first code. Roy attributes the ‘been around forever’ feeling to aggressive distribution and a world-class marketing team rather than product maturity.
- •ARR update: approaching $6M
- •Product is extremely new (first code ~10 weeks prior)
- •Perception gap: timeline dominance vs. company age
- •Distribution/marketing seen as the core early advantage
- 4:25 – 8:26
Virality on X/LinkedIn: controversial, digestible, and engineered for reactions
Roy breaks down why he thinks many tech founders fail at social distribution: they optimize for sounding smart instead of being shareable. He argues short-form algorithms reward controversy and clarity, and content must invite a strong quote-tweet/response to spread.
- •X/LinkedIn reward digestible, opinion-forcing posts more than essays
- •Tech audiences often write for insiders; virality requires mass comprehension
- •Two-part test: “digestible” + “reactionable”
- •Design posts so someone can quote-post a viral reaction
- 8:26 – 12:31
Stop obsessing over funnels: iterate formats faster than the algorithm shifts
Aakash pushes on conversion measurement; Roy argues it’s hard to attribute conversions when you dominate attention, and formats expire quickly. He emphasizes constant iteration over tracking precise ROI, because today’s winning format becomes stale within weeks.
- •Brand visibility is hard to price and hard to attribute
- •Viral formats change rapidly; repetition decays fast
- •Focus on making content with clear call-to-action rather than perfect attribution
- •Roy deprioritizes conversion analytics as “noise” in fast cycles
- 12:31 – 13:59
Stunt pipeline and content ops: daily brainstorms, UGC scale, and “reverse engineering” trends
They get specific about how Cluely manufactures viral moments—stunts, trend analysis, and a steady production machine. Roy describes daily idea sessions, a large creator bench, and an emerging in-house studio designed to ship high-quality launches weekly.
- •Daily unscheduled brainstorms generate ~20 high-potential ideas
- •Team skews young; “viral sense” is treated as a core competency
- •Reverse-engineer why recent trends worked, then adapt them
- •In-house UGC studio with 60+ creators on retainer
- •Expanding toward a full internal film studio (videographers/editors)
- 13:59 – 16:29
From Ghibli trend to “Cluely anime”: turning a cultural insight into a planned viral moment
Roy explains how a trend (Studio Ghibli/4o image wave) led to a larger creative bet: an anime trailer and series concept to tap nostalgia among tech audiences. He frames this as a repeatable method—identify emotional triggers, then over-invest in execution.
- •Trend dissection: not just controversy—nostalgia and emotion drive sharing
- •Hypothesis: tech audiences have deep affinity for anime/Ghibli aesthetics
- •Plan: start with a trailer to test the waters
- •Willingness to invest heavily if confident in 100M-view potential
- 16:29 – 18:08
Cluely’s UX thesis: translucent overlays as the future AI interface
The discussion shifts from marketing to product. Roy describes how Interview Coder served as the prototype and why translucency/overlay is essential for “undetectable” assistance and, more broadly, for AI that feels integrated rather than split-screened.
- •Interview Coder as Cluely’s prototype and UX proving ground
- •Goal: seamless assistance without obvious app/window switching
- •Translucent overlay enables working while seeing AI output inline
- •Belief that Apple’s “Liquid Glass” signals the same direction
- •Rejection of separate chat windows as the long-term AI GUI
- 18:08 – 19:38
Design without designers: brute-force iteration and founder-led product taste
Roy shares that they iterated through 20–30 versions before landing the current interaction model. He says product design is founder-led and engineering-heavy, relying on fast cycles and taste rather than a traditional design function.
- •20–30 iterations to reach current UX
- •Process described as brute-force experimentation
- •No dedicated designer; founders lead design decisions
- •Belief that strong engineers can also carry product sense
- 19:38 – 21:17
Early-stage execution: lean engineering team and user-driven prioritization (no formal sprints)
Aakash presses on roadmap rigor and coordination. Roy explains prioritization is straightforward with huge user volume and a tiny team—only four engineers—so decisions are made directly based on usage patterns and feedback.
- •Strategy: launch broad, go viral, observe use cases, then iterate
- •Millions of requests + hundreds of emails dictate priorities
- •Only 4 engineers—alignment and coordination are simpler
- •Emphasis on being much earlier than people assume
- 21:17 – 24:07
Technical architecture: audio capture, screen context, compression, and latency tradeoffs
Roy outlines the core engineering challenges: capturing both system and mic audio, providing screen context, and keeping response times low. He explains they snapshot the screen at query-time, compress images to reduce token load, and explore infra/model hosting to reduce variance.
- •Custom audio engine captures system + microphone audio on Mac
- •Screen understanding uses screenshot-at-query rather than continuous video context
- •Image compression reduces token/input cost from high-res PNGs
- •Latency variability with OpenAI prompts exploration of self-hosted models
- •Currently primarily powered by GPT‑4.1
- 24:07 – 26:33
Enterprise pull vs “AI for everyone”: features shipped and the product’s wedge
They discuss recent enterprise wins and features like post-call summaries and coaching moments, especially for sales. Roy argues the long-term goal is not a vertical sales tool but a default AI interface for everyone, with personalization and workflow-native UX as the differentiator.
- •Signed a large enterprise contract; enterprise-driven feature shipping
- •Post-call summaries plus “you should have used Cluely here” coaching
- •Sales enablement positioning: objection handling and training utility
- •Vision: not “Cluely for sales” only—capture broad consumer market
- •Personalization seen as a current value unlock; onboarding experiments ongoing
- 26:33 – 31:41
System prompt ‘leak’ and eval-driven iteration: prompt isn’t the moat
Aakash asks about a purported leaked system prompt. Roy downplays it, saying prompts change constantly and will likely be open-sourced; the moat is UX plus context ingestion. He also notes they run internal evals and update prompts based on shifting use cases.
- •Prompt changes frequently (every few days) as use cases shift
- •Plans/intent to open source prompts
- •Moat framed as UX + product experience, not prompt text
- •Internal evals used to guide prompt iteration
- •End-state prompt: simply provide audio/screen context and help appropriately
- 31:41 – 36:47
Big bets: fundraising preempt, revenue mix targets, and extreme end-state visions (CRM → brain chips)
Roy describes a preempted Series A that effectively required no formal raise process and discusses aiming for consumer-led scale. He then paints ambitious futures: enterprise workflows that replace CRMs and a consumer trajectory that ultimately points toward brain-computer interfaces.
- •Series A described as a preempt after arriving in SF
- •Targeting 100M ARR; aspirational mix ~70% consumer / 30% enterprise
- •Enterprise end-state: real-time AI that reads/writes CRM data (even “kill Salesforce”)
- •Consumer end-state: brain chips as ultimate AI form factor
- •Belief in rapid tech acceleration collapsing sci-fi into reality
- 36:47 – 45:23
Company culture as a growth engine: “frat house” intensity, comp, interns, and criticism
They close on what it’s like to work at Cluely: intense, all-consuming, communal living, and high pay/equity to match expectations. Roy acknowledges risks (including potential lawsuits) but argues culture is opt-in; they also discuss an expanded ‘intern’ concept via paid UGC creators and address concerns about cheating making people ‘dumber.’
- •Culture: live/work together in a SF mansion; minimal work-life separation
- •High expectations paired with high compensation and equity
- •UGC creators treated like marketing interns; large creator “intern” pool
- •Addressing Uber-like ‘fratty’ culture comparisons and potential downsides
- •Argument that AI makes old knowledge redundant; parallels to calculators
- •Public controversy (e.g., jail rumor) framed as persona vs reality
