Broke to $60K in 3 Months: Raj's Enterprise RAG for Pharma and Banks
With only a few thousand dollars left and his AI project about to fold, serial entrepreneur Raj didn't quit. He targeted the most unsexy corner of enterprise AI — document gold mines in pharma and banking. Three months later, $60K in the bank.
Process
Six months ago, Raj's company bank account was down to a few thousand dollars.
He was a Sri Lanka-born serial entrepreneur who had been building RAG (Retrieval-Augmented Generation) systems for legal documents in the US. The project was burning cash faster than expected, and his AI venture looked finished.
Instead of quitting, he stopped and observed the market carefully for what others had missed.
He Targeted the Most Unsexy Corner
Almost every established enterprise sits on a massive "document gold mine" with no good tools to extract value from it. Pharma companies have decades of clinical trial papers piled up. Banks have mountains of regulatory compliance filings. Law firms have endless case histories. Employees who need to use this knowledge can only search manually, page by page.
Nobody wanted to touch this work because it was "dirty" — chaotic document formats, inconsistent data quality, nightmarish implementation complexity where demos worked but real deployments crashed. That was exactly Raj's opportunity.
He focused on pharma and banking: the two verticals with the deepest document accumulation and the highest ability to pay.
The Question That Opens Every Door
Raj never asked "do you need AI?" — that question is too easy to deflect with "we're evaluating options."
Instead, he asked: "How many hours does your team spend reading documents every day?"
That question almost always starts a real conversation, because the pain is immediate and concrete. His first three clients all came from his personal network — someone in his circle was always suffering from document chaos.
The First Trap: Selling Too Cheap
His first MVP was priced at $5,000–$10,000. The client agreed instantly. Raj was pleased.
He later realized: an instant yes means you priced too low. He quoted $30,000 for a second client — no negotiation, signed immediately. By his fourth and fifth engagement, he understood: very few people in the market could handle this level of complexity. Scarcity is pricing power.
He raised prices. Fewer clients said yes, but he needed only one or two high-value contracts to run for three to four months.
The ROI math for enterprises is stark: 50 researchers spending 2 hours/day on document search at $100/hour = $200,000 wasted per month. A $50,000 RAG system cutting that by 80% pays for itself in days. You're not selling technology — you're selling a measurable return on investment.
The Real Moat: After the Demo Works
Nearly every enterprise client had already tried building their own RAG system. The demo ran fine. It collapsed the moment real enterprise documents were loaded.
Raj identified six failure modes that caused enterprise RAG to break at scale:
① Document quality is "garbage-tier": Enterprise document libraries mix 1990s scanned images, image-format PDFs, and modern electronic files. Dirty data in, garbage answers out. You need automatic quality detection that routes documents to the right pipeline — OCR for scanned files, structured parsers for electronic ones.
② Fixed-size chunking catastrophically fails: Cutting every 500 words severs complete arguments mid-thought. Documents have hierarchical structures. Hierarchical chunking that preserves headings, sections, and context relationships is essential for coherent retrieval.
③ Wrong embedding model loses accuracy: General-purpose embeddings work for everyday language but fail on pharma clinical terminology and banking regulatory language. Domain-specific fine-tuning or vertical-specialized embedding models are required.
④ Single retrieval strategy is insufficient: Pure vector similarity fails for precise lookups of regulation numbers or drug compound names. Hybrid retrieval combining keyword search with semantic search is necessary.
⑤ Multi-hop reasoning hits a wall: Questions requiring cross-document reasoning ("compare the 2019 and 2023 versions of this regulation") can't be answered with single-pass retrieval. Multi-step reasoning chains are required.
⑥ Hallucination and missing attribution are fatal: Enterprise clients in medical and financial contexts cannot accept unverifiable answers. Every generated response must trace back to specific document passages and page numbers.
These six failure modes were things Raj discovered the hard way — through real deployments with real clients. Each one became his competitive moat.
Three Months. $60,000.
The person who had a few thousand dollars in the bank six months ago — three months later had pharma companies and banks lining up to work with him. $60K in, $20K+/month run rate.
No flashy product. No consumer app. No chasing the latest model release. Just the dirty work of making enterprise documents actually useful in the two highest-paying verticals he could find.
Source: 大黑AI (Xiaohongshu)
Thinking
Why pick the dirtiest work instead of the coolest product? When everyone chases the sexy AI applications, the unsexy corners have far less competition and far stronger customer willingness to pay — because the pain is real, ongoing, and quantifiable.
Why pharma and banking specifically? Two conditions both met: ① deepest document accumulation with the most chaotic quality; ② highest compliance pressure making manual search most expensive. Clients who have money, pain, and urgency are the ideal combination.
On pricing: Frame it not as "what my system costs" but as "what my system saves." When $50K becomes a 3-day payback period investment, it's not an expense anymore.
Action
- Find the document gold mine in one vertical: Medical, legal, finance, manufacturing, energy — any sector where employees waste hours daily searching documents.
- Use the opener: "How many hours does your team spend reading documents per day?" Let them calculate their own pain.
- First three clients from your network: Higher tolerance, real feedback, willing to work through early problems with you.
- Charge first, perfect later: If they won't pay $5K for a working MVP, they're not worth pursuing.
- Anchor pricing to ROI: Calculate their monthly waste, price your solution at 20-30% of that number. It almost sells itself.