LEANcrush

Building the right product better and faster.

I help startups between Seed and Series A to gain more traction and reach Product-Market Fit (PMF) faster by implementing product discovery and delivery processes that allow them to quickly test and iterate business models.

Recent work:
• 🥫KraftHeinz Venture Builder: led team of 15 to test 30 ideas and deliver 3 MVPs to disrupt the $25b Food Market.
• 💰 Q.ai [backed by Forbes]: built team of 10 and delivered V1 AI-investor to 50k early access users.
• 💊 Menavitin by Pelion: launched a new healthcare product in Germany by executing discovery and design sprints.
• 🤖 Hulecos SA: launched discovery-and-design sprints with an MVP to validate a deep-tech and mobile app business.

By combining world-class methodologies such as Customer Development, Lean Startup, Business Model Generation,Value Proposition Design, Jobs-to-be-Done, you’ll be able to: uncover Big Opportunities, Test Riskiest Assumptions and Build Desirable Products.

Fractional Product Leader

Techstars

Mentored +50 startups on:
- Workshopped with +50 startups on PMF Discovery
- Designing their value proposition, identify critical assumptions to find critical risks 
- Building product and launching experiments
- Multiple industries such as education, health, mobile app, SaaS, eCommerce, and fintech.

Startup Mentor for Product-Market Fit

Toptal

I work as a Product Leader solving pressing product challenges and giant market opportunities.

- Launched an award-winning Robo-investor for Forbes as Interim VP Product
- Led a corporate Venture Builder for Kraft-Heinz

A.Team

I helped A.Team to increase the quality of their network by evaluating +5k PM Profiles and interviewing 350 Product Managers (Designers & Marketers) to find superstar builders from companies like Dropbox, Microsoft, Uber etc.

Lead Product Manager & CPO Interviewer

Founder Institute

Worked with +180 startups:
- Identifying assumptions within their business model and building experiments to validate them.
- Building customer journeys, interviewing customers, defining value propositions and customer segments, analysing existing alternatives, prioritising unmet needs and building MVPs for innovative business industries (health, mobile app, SaaS, eCommerce, finance).
- Voted #1 mentor by startups.

Startup Mentor (Customer Development, Product Discovery)

The Kraft Heinz Company

Leading a 15 team of builders to launch innovative companies to disrupt the food industry.

- Designed Discovery and Delivery Sprint Process
- Trained team and Executed multiple sprints
- Results: 300 innovative ideas mapped, 30 validated, 3 MVPs launched

Head of Venture Builder

Building the next generation of ventures to disrupt the food and beverage industry.

Digital Ventures Lead

Fintech co. (backed by Forbes)

Led the creation, since inception, of consumer fintech mobile investing app (backed by major publishing giant Forbes).

- Defined the value proposition, business model and positioning.
- Built 1st prototype and validated assumptions by surveying 1000s of prospects and interviewing 100s.
- Crafted a product vision, strategy and product feature roadmap with job and user stories. 
- Led design sprints to build several iterative MVPs and user-tested with 100s of users.
- Built marketing material ie. landing pages and videos to help gather over 40k early access users
- Led agile development with scrum to deliver an Alpha 1.0, 2.0, and Beta 1.0 (fully working and compliant).
- Managed design, development and marketing projects involving +15 people

The app has garnered multiple Global Fintech Awards.

Interim VP Product

Proudsugar

Built a team of domain domain experts and contracted to high growth startups to help them grow:

- Openme (acq. by Rowl):
Led a team of seven, added analytics, launched surveys, optimized customer journey, built a testing roadmap, redesigned UX, and launched continuous A/B tests for eCommerce Openme.com.

- Plug.dj:
Led a team of nine, added analytics, surveyed and interviewed, redesigned UX, road mapped, and launched  continuous A/B tests for Plug.dj to increase activation, retention and revenue (a web/mobile music startup with 7 million users).

- Pluto.tv (acq. by Viacom):
Worked with a team of seven, implemented web analytics and created dashboard reports for Pluto.tv (an internet TV startup that sold to Viacom for $300 million). This led to identifying product conversion problems, UX redesigns and better KPIs.

- MROSupply: 
Led a team of six, implemented web analytics, interviewed/surveyed customers, identified opportunities, redesigned UX and launched continuous A/B tests for MROSupply (eCommerce with 1 million SKUs for mechanical parts) resulting in an +$5 million/year sales.

- Discuto
Led a team of five for Discuto.io (a B2B SaaS ideation and collaboration platform with +20,000 users), implemented analytics, performed user surveys/interviews, launched a UX redesign, resulting in higher onboarding of 35%, as well as 2x retention.

Co-founder | CEO | CPO

Javier Rincon

Freelance CRO and Web Analytics practitioner. Helped companies establish a web analytics infrastructure to gain insights on major KPIs. Then worked with CRO methodologies to increase main business and product metrics.

- Moviepilot (acq. by Webedia):
Worked with a team of 15, added analytics, did user research, built an opportunity roadmap, and launched continuous A/B tests for Moviepilot (a movie content platform that sold to Webedia), resulting in higher engagement metrics eg. +520,000 active onboarded users/year.

Product Analytics, Growth & CRO specialist

6Wunderkinder

Chief Marketing Officer for globally recognized Berlin startup, 6Wunderkinder, creators of Wunderlist with more than 2Million Downloads and new productivity app Wunderkit.

Building PMF Engines @ startups | Product Leader | Lean

Javier has +15 years experience building products that millions have used. 

• Startups: @Wunderlist (acq. by Microsoft), @Plug.dj, @Open.me (acq. by Rowl), @Pluto (acq. by Viacom), @Moviepilot (acq. Webedia).

• Venture Building: @LaSalle, @Forbes , @KraftHeinz

• Startup PMF mentoring & financing: @Keiretsu Forum, @Founder Institute and @Techstars.

• Industry: Business Intelligence @Oracle; System Analyst @Transmitton Siemens

He's also founder and builder; he ran a product agency in Berlin, co-founded a mobile social offers company in Madrid and launched an e-pharmacy out of University.

He has a MEng in Systems Engineering from Loughborough University in the UK. He's an avid learner, passionate about playing guitar and piano, storytelling in film and comics, war strategy, oil and gold drilling, railroads, and the expansion of empires.

We ran 200k Tokens of finance docs Into LLMs… and accuracy COLLAPSED!

Everyone is celebrating the new long-context models.

500k tokens! 1M tokens! Just throw in the docs, the model will figure it out.

At 42Automation (Adrian Tineo, Ph.D. and me), we tested to see where accuracy would break. (= Chroma's Context Rot!)

The models: Gemini 2.5 Flash, GPT-5 Mini and Qwen 3-235B.

At 200k tokens, all models failed.

Let's break it down.

We picked a very practical use case:
Verifying whether the numbers in an IPO prospectus actually match the company’s annual report.

In the Link Mobility IPO prospectus, we manually found an error — some numbers didn’t line up with the annual report.

Normally, an analyst would do this check. In this case, the mistake slipped through. So the question was: would AI catch it?

This is a high-stakes test. If AI misses errors here, investors could be making million-dollar decisions on bad data.

We conducted 3 experiments:

1/ Short context (~2k tokens)
We fed the model just the relevant sections.

✅ All models solved it. Easy.

2/ Medium context (~286k tokens)
One long document + one short extract.

✅ Gemini and GPT-5 Mini still spotted the discrepancies.
⚠️ Qwen got confused — it knew there was a mismatch but couldn’t give the details.

3/ Full long context (>200k tokens)
We stuffed both long docs into the context window.

❌ The models collapsed.
– GPT-5 Mini hallucinated a revenue value that never existed.
– Qwen downgraded from:
“There are discrepancies” →“There are discrepancies, not sure which” → “There are no discrepancies.”

The takeaway?
> Short docs? → they catch the errors.
> Full long documents? → they collapse and make mistakes.

More context can actually create new mistakes that weren’t there before.

That’s a big problem if you’re relying on AI for: IPO filings, Financial audits, Compliance checks, Due diligence, Contract verification, Regulatory compliance

What’s next?
We’re testing mitigation on how to increase reliability under long-context loads. Stay tuned!

🔗 Full report in comments

#AI #Automation #AIAgents

Fractional Head of Product & AI | The bottleneck guy

What’s the right way to feed large documents into LLMs?

For example, with financial filings you don’t just have text — you’ve got tables, graphs, and charts.

So at 42Automation (Adrian Tineo, Ph.D. and me) we asked...
...what format should we use as input?
- Raw text?
- Markdown?
- PDF as-is?

To make this work at scale, we’d need to homogenize all docs.

But here’s the twist: when you pass the same document in different formats…
the token counts explode or shrink.

We tested the same Tesla filing 3 ways:
- PDF → 53 tokens 
- TXT → 787 tokens 
- MD → 872 tokens

Same correct answer. Completely different token counts.

What is this black magic?

At first we thought: “Okay, PDFs must be internally compressed.”

But then things got weirder:
- As text, Claude hits the 200k ceiling.
- As PDFs, the same docs slide through (we even fit 3 together).

Yet the API cost across formats was almost identical.
So what’s going on?

- How does the LLM actually consume tokens for each format?
- Are PDFs stored in a special compact representation?
- Is the token counter misleading us?
- Or is Claude on POE capped at a smaller context window, and we’re counting wrong?

Right now, we don’t trust the results. 

Either tokens aren’t being reported correctly… or the model is internally treating PDFs as images + OCR, extracting only the relevant text while hiding the real token cost.

The Takeaway?
Choosing between PDF, text, or markdown isn’t just cosmetic.

It changes token usage, costs, and even whether your workflow runs at all.

The way tokens are counted for PDFs still feels off, so we’ll keep investigating.

And for now, we’re leaning on texts — although we won’t be able to answer 
about charts, the costs are dependable.

How do you feed large documents into your LLM?

👇 Full report below:

#AIAgent #AI #GenAI

The hidden cost of document format in LLMs

Taplio

Javier Rincon 💡's Linkedin Analytics

's Best Posts (last 30 days)

Want to drive more opportunities from LinkedIn?

Taplio

.css-1jgreyo{background-image:linear-gradient(284deg, rgba(39,43,72,1) 0%, rgba(33,47,183,1) 100%);color:transparent;-webkit-background-clip:text;background-clip:text;}Javier Rincon 💡's Linkedin Analytics

's Best Posts .css-qqfgvy{font-size:var(--chakra-fontSizes-md);}(last 30 days)

Want to drive more opportunities from LinkedIn?

Javier Rincon 💡's Linkedin Analytics

's Best Posts (last 30 days)