Triomics' $22M Says Narrow AI Beats General AI
Triomics raised $22M for oncology-specific AI, not another general chatbot. Here's why vertical AI is the smarter bet for a small Sri Lankan team.
A vertical AI startup just raised real money for doing the opposite of what most AI demos chase. Triomics pulled in $22 million in a Series B to build AI that only does one thing: oncology. No general assistant, no "ask it anything." Just cancer-center workflows, modelled deeply. I think that focus is the actual lesson here, and it's one a small team in Colombo can act on without a $22M cheque.
I'm reacting to TechCrunch's report, Triomics nabs $22M to bring oncology-specific AI to cancer centers. The facts below are theirs. The argument is mine.
🔍 What Triomics actually built
According to the report, Triomics (founded 2021, co-founded by CEO Sarim Khan and CTO Hrituraj Singh) trains its models specifically on oncology data and plugs the output into the tools clinicians already use, so nobody has to switch apps. The tasks it automates are unglamorous and specific:
- Clinical trial matching for patients
- Appointment preparation
- Patient chart summarization
- Tumor report submission to government registries
"We have seen medical records [with] thousands of pages of information" — Sarim Khan
That quote is the whole pitch. The problem isn't that doctors lack a chatbot. It's that a single patient file can run to thousands of pages, and a generic model has no idea which pages matter for an oncology decision. Depth in one domain is the product.
📊 The funding tells you what investors believe
Here's what the round looks like, pulled straight from the report:
| Detail | Value |
|---|---|
| Round | Series B |
| Amount | $22 million |
| Lead investor | Battery Ventures |
| Other investors | Nexus Venture Partners, Lightspeed, Y Combinator |
| Prior round | $15M Series A (mid-2024) |
| Enterprise customers | Grew 4× over the past year |
| Annualized recurring revenue | 10× increase |
| Named customers | Memorial Sloan Kettering, Yale Cancer Center |
A 10× jump in recurring revenue and a 4× customer count is what convinces a fund to write a follow-on cheque less than two years after the Series A. Notice what's not in the story: no claim of a smarter base model than the big labs. The moat is the domain, the data, and the fact that the output drops into existing clinical software.
Key takeaway: Investors didn't fund a better model. They funded a team that owns one workflow end to end. That's a moat a two-person team can build; training a frontier model is not.
💡 Why "narrow" is the opening for small teams
The generative-AI gold rush trained everyone to think big: one model, every task, every user. But the big labs already own that. You will not out-general OpenAI, Google, or Anthropic. What they cannot do is know your niche better than you.
Triomics didn't beat anyone on raw model quality. It won by going where the giants won't bother:
- A narrow domain with messy, high-stakes data (oncology records).
- A specific, repeated task people already pay humans to do (chart summaries, trial matching).
- Verifiable output that slots into existing tools instead of demanding a new app.
That third point is the one builders skip. A summary a doctor can't trust is worse than no summary. The report stresses Triomics produces verifiable patient summaries inside the clinician's existing workflow. Trust and integration, not novelty, are what got bought.
For a Sri Lankan engineer or small team, the same shape applies to far smaller markets. Think:
| Generic idea (crowded) | Narrow version (winnable) |
|---|---|
| "AI assistant for business" | Garment-export compliance doc checker |
| "AI for healthcare" | Sinhala/Tamil discharge-summary translator |
| "AI legal helper" | Sri Lankan land-deed clause extractor |
| "AI for accountants" | IRD APIT filing reconciliation |
None of those need a custom model. They need someone who understands the domain, owns the messy data, and integrates into the tool people already open every morning.
🛠️ How to test a narrow-AI idea this weekend
You don't need funding to validate this. You need a domain you understand and a wrapper around an existing model. A rough sketch:
1. Pick ONE document type you know well (a payslip, a lab report, a tender).
2. Collect 20 real examples (anonymise anything sensitive first).
3. Prompt a hosted model to extract the 5 fields that actually matter.
4. Have a domain expert grade the output. Where it's wrong is your moat.
5. Build the verification + integration layer. That's the hard, valuable part.
The extraction is the easy 20%. The verification and the fit into someone's daily workflow are the 80% that no general tool will ever do for your niche. If you're working with documents, the small utilities I've built — like the in-browser tools on this site for OCR, PDF handling, and text extraction — are the kind of plumbing you'd reach for before any model sees the data.
A warning worth repeating: anything touching medical, legal, or financial data carries real liability. "Verifiable" in the Triomics story is doing heavy lifting. Don't ship an unchecked summary into a decision that affects someone's money or health.
What this means for you
The headline number is $22 million, but the transferable lesson costs nothing. Pick a domain you genuinely understand, find the one repeated task inside it that's drowning in messy data, and build the verification layer that makes a general model trustworthy for that task. That's a business a small Sri Lankan team can start this month.
You can't win the general-AI race. You were never supposed to. The opening is in the thousand-page files the giants can't be bothered to read carefully, and there are plenty of those right here.