Free tools Windows power users keep installed
One-click scans. No signup required.
TechCrunch’s March 13, 2025 roundup highlighted 10 companies from Y Combinator’s Winter 2025 (W25) Demo Day. The list was an editorial selection from a batch of approximately 160 startups—not YC’s official ranking and not proof that any company will become a venture winner. Its significance is the pattern: AI is moving from chat interfaces into labor, software execution, marketplaces, and physical machines.
For founders and investors, the useful question is not which pitch sounded most impressive. It is whether each company can turn an interesting demo into repeatable revenue, safe deployment, defensible technology, and workable unit economics.
What W25 Demo Day was—and was not
Y Combinator describes Demo Day as an invitation-only presentation for investors and media. It is a showcase of a new batch, while the broader YC company directory contains companies from many batches and stages. A pitch demonstrates a thesis; it does not independently validate product-market fit, retention, revenue, safety, or regulatory readiness.
TechCrunch’s article, published March 13, 2025, selected the following 10 startups from W25. The companies below should be treated as businesses worth monitoring, not as “YC’s top 10” or guaranteed breakout companies.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →#1 Best Overall
YC’s Demo Day FAQ explains the investor- and media-oriented format. That context matters when interpreting founder-reported downloads, users, letters of intent, and range claims.
The 10 companies at a glance
| Startup | What it does | Primary user or buyer | Reported signal | Key unresolved risk |
|---|---|---|---|---|
| Abundant | API for human teleoperation of AI agents | Businesses deploying agents | Human fallback for agent failures | Cost and latency of intervention |
| Browser Use | Open-source browser control for agents | Developers and AI platforms | 28,000 daily downloads reported during a surge | Reliability, security, and monetization |
| GradeWiz | AI-assisted grading | Universities, instructors, teaching assistants | Founder experience as Cornell teaching assistants | Accuracy, bias, and privacy |
| Misprint | Bid/ask marketplace for Pokémon cards and collectibles | Collectors and traders | Company-estimated $3.5 billion annual secondhand Pokémon-card market | Liquidity, authenticity, and speculation |
| NextByte | Assessment for AI-assisted (“vibe”) coding | Employers hiring engineers | Targets a newly important hiring skill | Predictive validity and test integrity |
| Pickle | AI body double for video calls | Consumers and professionals | Company reported more than 1,500 paying users | Consent, disclosure, and platform policy |
| Rebolt | Agents for restaurant inventory and procurement | Restaurants and franchise groups | Pricing discussions with a Burger King parent company | Integration and operational errors |
| Red Barn Robotics | Autonomous agricultural weeding robot | Farm operators | Company claimed $5 million in letters of intent | Field reliability and capital intensity |
| Retrofit | AI-curated vintage-fashion marketplace | Vintage shoppers and sellers | Uses curation to solve discovery | Two-sided liquidity and fulfillment |
| Splash | Autonomous patrol boats | Maritime security and defense customers | Company claimed 200 autonomous miles and 800-mile range | Procurement, regulation, and mission proof |
AI agents need execution and recovery
Abundant: a human safety net for autonomous software
Abundant offers an API for teleoperating AI agents. An agent can work autonomously until it detects an error or unfamiliar situation, then hand the task to a human operator. TechCrunch described that failure-detection and takeover model in its W25 coverage.
The opportunity is practical: companies may adopt agents sooner if there is a defined recovery path instead of an all-or-nothing promise of autonomy. The business risk is that the “fallback” becomes the real product. Every intervention adds labor, latency, training, quality-control, and liability costs. Investors should ask who supplies operators, how intervention time is priced, and who is accountable for an agent’s action while a person is taking over. The sources do not establish that teleoperation economics scale; that remains a central question.
Browser Use: the web as an agent execution layer
Browser Use provides open-source tooling that lets agents navigate websites, click controls, fill forms, and work across services that do not expose convenient APIs. That makes it an important practical layer: businesses could automate existing web workflows instead of waiting for every application to become agent-native.
Recommended Free Tools
TechCrunch reported that daily downloads reached 28,000 during a surge associated with attention around the Chinese AI agent Manus. Downloads are an adoption signal, not evidence of active production installations, paid conversion, or recurring revenue.
- Web layouts change, creating continuing maintenance work.
- CAPTCHAs, multifactor authentication, bot detection, and terms of service can block automation.
- Credentials and customer data require strong isolation and audit trails.
- Open-source popularity may not create a defensible commercial business if model, browser, or cloud vendors bundle similar capabilities.
NextByte: measuring engineers who work with AI
NextByte evaluates candidates on effective use of AI coding tools rather than traditional coding ability alone. Its premise is that modern engineering includes prompting, code review, debugging, architecture, and judgment about generated code.
The hard problem is measurement. A useful assessment must show that its score predicts production performance, not merely tool familiarity. It must also prevent candidates from outsourcing the test, accommodate different toolchains, and avoid penalizing developers whose workflows do not match the test designer’s assumptions. Because AI-assisted development is changing quickly, any benchmark can age quickly too.
Vertical AI targets repetitive operations
GradeWiz: grading assistance with human accountability
GradeWiz applies AI to repetitive grading work. TechCrunch reported that its founders were Cornell teaching assistants who disliked the workload, giving the product a clear user-origin story and a narrow workflow to test.
Speed is not enough. Open-ended answers, multilingual submissions, handwriting, accommodations, and discipline-specific rubrics can expose inconsistency or bias. Faculty and institutions remain responsible for grades, appeals, accessibility, and student-data governance. Buyers should examine whether GradeWiz assists a human reviewer or is being positioned as a substitute, how retention is controlled, and how disagreements are audited.
Rebolt: restaurant operations as an agent workflow
Rebolt is building agents for inventory tracking, supplier communication, and procurement. Restaurants operate with thin margins and fragmented systems, so eliminating repetitive coordination could have a direct economic benefit.
TechCrunch said Rebolt was in pricing discussions with the parent company of Burger King. A pricing discussion is not a signed customer, rollout, or disclosed revenue. The commercial test is whether Rebolt can integrate with point-of-sale, inventory, accounting, and supplier systems and earn permission to act. Managers may want approval controls for substitutions, changing prices, late deliveries, and food-safety requirements; one procurement mistake can erase the savings from automation.
AI leaves the screen
Red Barn Robotics: autonomous weeding in difficult conditions
Red Barn Robotics’ “The Field Hand” is designed to remove weeds in agricultural fields. TechCrunch reported company claims that it was 15 times faster than a human, cost about one-quarter as much as human labor, and had approximately $5 million in letters of intent for the upcoming growing season.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsThose are company claims, and letters of intent are preliminary commercial interest—not revenue or necessarily non-cancellable contracts. Farm economics depend on crop, soil, weather, acreage, supervision, maintenance, financing, and the cost of missed or damaged plants. Prospective customers need field-trial results, acres covered per day, uptime, service arrangements, purchase or lease pricing, and payback periods. Dust, mud, heat, slopes, and changing light are not edge cases; they are the operating environment.
Splash: autonomous patrol boats
Splash is developing small autonomous boats for maritime border or security patrol. TechCrunch reported that Splash said its boats had autonomously traveled 200 miles in the San Francisco Bay Area and claimed an 800-mile range.
These figures are not interchangeable: a demonstrated autonomous trip is different from a range specification, and neither by itself proves a deployable mission profile. Speed, payload, weather, communications, battery or fuel configuration, navigation in GPS-denied areas, and behavior around civilian vessels all matter. Buyers also face procurement, insurance, surveillance, and cross-border rules. Investors should determine whether Splash sells boats, autonomy software, or a managed service and how much working capital a defense sales cycle requires.
Rank #4
- Author: Guillebeau, Chris.
- Publisher: Currency
- Pages: 304
- Publication Date: 2012-05-08
- Edition: NO-VALUE
Marketplaces and synthetic presence
Misprint: exchange mechanics for collectibles
Misprint is building a bid/ask marketplace for Pokémon cards and other collectibles. Its exchange-style structure could make fragmented prices more visible and trading more continuous.
Misprint estimated that about $3.5 billion of secondhand Pokémon cards are sold annually; that is the company’s estimate, not an independently established market total. TechCrunch also reported that co-founder Eva Herget left Goldman Sachs to sell Pokémon cards and reached approximately $40,000 per month in sales. That is a founder-origin and traction anecdote, not audited company revenue.
The marketplace must solve grading, authenticity, condition disputes, shipping, returns, custody, and liquidity. Users appear to be trading physical collectibles rather than securities, but the company still needs to avoid presenting speculative price movements as guaranteed investment returns. A bid and an ask do not prove that enough buyers and sellers will transact.
Retrofit: making vintage inventory discoverable
Retrofit uses AI to curate vintage clothing according to trends and shopper preferences. Vintage marketplaces have abundant, inconsistently described inventory; better discovery could increase conversion without owning every item.
The business model matters. Aggregating listings, operating a seller marketplace, and owning inventory create different margins and liabilities. Retrofit must handle sizing, condition, authenticity, counterfeit goods, returns, fulfillment, seller acquisition, and buyer acquisition. Curation is a wedge, not a complete moat: larger marketplaces can add recommendation features, while weak liquidity can make even excellent search frustrating.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallBest Value
Pickle: an AI body double for video calls
Pickle replaces a user’s camera image in a video call with a polished, lip-synced digital version. YC’s profile describes it as a virtual body double that can appear presentable when the user is not camera-ready: Pickle’s YC company profile. TechCrunch reported that Pickle claimed more than 1,500 paying users in March 2025.
The product’s durable use case could be privacy, convenience, accessibility, or entertainment. Its trust problem is disclosure. Employers, schools, and meeting participants may require users to identify synthetic video; platforms may label or restrict it. Identity, face, voice, and consent controls are central product requirements. Reliability also needs testing under poor lighting, unusual speech, and rapid movement. Paying users are a useful signal, but the reported figure was company-provided and does not establish retention or policy acceptance.
How to evaluate the “watch” list
Use different standards for software, marketplaces, and robotics. The following questions are more informative than Demo Day applause:
- Problem severity: Is the product solving expensive, frequent, urgent work?
- Buyer clarity: Who controls the budget, and can that person approve deployment?
- Evidence quality: Separate downloads, pilots, letters of intent, paid users, signed contracts, renewals, and revenue.
- Deployment burden: Does adoption require hardware, integration, regulation, insurance, or field service?
- Defensibility: Look for proprietary data, workflow integration, operational know-how, hardware, distribution, or regulatory expertise.
- Human-in-the-loop economics: Calculate intervention and review costs rather than assuming autonomy is free.
- Safety and liability: Examine grading appeals, synthetic identity, procurement authority, and autonomous-vehicle incidents.
- Market structure: Marketplaces need liquidity and trust; infrastructure needs ecosystem adoption; hardware needs manufacturing and service capacity.
- Incumbent response: Browser vendors, restaurant platforms, education systems, marketplaces, and defense contractors can copy or bundle features.
- Time and capital: A robotics or defense company should not be judged by the same cash cycle as a developer tool.
What this snapshot cannot tell you
The TechCrunch roundup is useful for identifying memorable pitches, but it does not establish current operating status, retention, revenue, customer concentration, pricing, deployment scale, follow-on financing, or whether any company remained active through August 2026. The available sources also do not independently verify the reported traction figures. Those gaps are reasons to monitor the companies, not to fill in the blanks with assumptions.
For an investor, the next evidence to seek is recurring revenue and renewal behavior for software; verified field performance and service economics for robots; signed procurement and mission data for Splash; and repeat transactions, fraud rates, and contribution margins for Misprint and Retrofit. For a founder, the lesson is similar: a compelling AI demo becomes a company only when reliability, distribution, liability, and unit economics work together.
The takeaway
W25’s most interesting theme was AI as an operating layer. Abundant and Browser Use address execution and recovery; GradeWiz and Rebolt embed agents in narrow workflows; Red Barn Robotics and Splash carry autonomy into farms and waterways; Misprint and Retrofit apply AI to trust and discovery; Pickle tests whether people will pay for synthetic presence; and NextByte measures a new form of engineering work.
That makes all 10 worth watching. It does not make them proven winners. The decisive signals will be paid conversion, retention, reliable deployment, defensible data or operations, and the ability to manage safety and liability as usage expands.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
