Convexia Wants to Buy Drugs, Run Trials, and Sell for Profit Using AI Agents

The Macro: Drug Development Is Broken in Obvious Ways

I am going to say something that everyone in pharma knows and nobody in pharma wants to fix: the drug development pipeline is absurdly inefficient. The average cost to bring a new drug to market is $2.6 billion. The average timeline is 10 to 15 years. The success rate from Phase 1 clinical trials to FDA approval is around 10%. These are not problems at the margins. These are systemic failures at the core of an industry that generates $1.4 trillion in annual revenue.

The reasons are well documented. Drug discovery is manual and slow. Preclinical evaluation relies on wet lab experiments that take months for results that computational models could approximate in hours. Clinical trial design is conservative by necessity but inefficient by practice. Regulatory navigation is a specialized skill that most drug developers learn through expensive failure.

What is less well documented is how many viable drug candidates get abandoned along the way. A compound that works for one indication might be perfect for another, but the original developer dropped it because the commercial opportunity in the first indication was too small. Academic labs produce promising preclinical data and then cannot find pharma partners to take it further. Biotech companies run out of funding with assets that have real scientific merit sitting in their IP portfolios.

The AI drug discovery space has attracted serious capital. Recursion Pharmaceuticals, Insilico Medicine, and Isomorphic Labs are all working on using AI to discover new molecules. But discovery is only one piece of the puzzle. Sourcing existing overlooked assets, evaluating them computationally, designing trials, and navigating regulatory pathways are all separate problems that are equally ripe for AI.

The market structure creates an interesting opportunity for a new kind of pharma company. One that does not discover drugs from scratch but instead finds and acquires promising assets that others have missed or abandoned, uses AI to evaluate them faster than traditional methods, and runs leaner clinical programs. This is basically the venture capital approach applied to drug development: source deals, do rapid diligence, move fast, and sell winners to strategic buyers.

The Micro: Stanford Founders, Agent-Based Drug Evaluation

Convexia calls itself “the world’s first AI-maximalist pharma company.” The model is straightforward. Source overlooked drug candidates from academia, biotech, and abandoned pharma pipelines. Evaluate them using computational biology models. Run them through clinical planning. Sell the winners.

Ayaan Parikh and Rahul Vijayan founded the company. Both are Stanford grads. Ayaan studied CS and Biology. Rahul studied CS and Economics. That combination matters because this is as much a financial arbitrage play as it is a scientific one. They are based in San Francisco with a four-person team, part of Y Combinator’s Summer 2025 batch.

The platform runs on a series of specialized AI agents. The Sourcing Agent mines databases to find preclinical candidates that have been overlooked. The Scientific Agent runs computational models including ESM-3, RFdiffusion, Boltz-2, and AlphaFold to assess safety and efficacy. The Commercial Agent analyzes FDA incentives, pricing dynamics, market size, and competitive landscape. The Clinical Agent simulates digital twins and builds trial plans. The Probability of Success Agent evaluates factors that impact clinical trial likelihood.

What makes this different from Recursion or Insilico is the scope of ambition. Those companies use AI to discover new molecules. Convexia uses AI to do everything: find the asset, evaluate the science, model the market, plan the trial, and assess the probability of success. The claim is that this full-stack approach runs 10x faster and 20x leaner than traditional pharma.

The Asset Discovery Agent is available for $199 per month, which is aimed at investment groups, VCs, and biotech companies evaluating potential acquisitions. The rest of the platform components are available through custom licensing. The long-term play is to operate the full drug lifecycle: acquire assets, run trials, and sell to strategic buyers for profit.

They are using over 50 custom ML models for binding analysis, toxicity screening, ADME prediction, and immunogenicity assessment. The computational stack is impressive on paper. The question is whether computational evaluations are accurate enough to replace or meaningfully accelerate the wet lab work that pharma has relied on for decades.

Human oversight is built into the process. PhD scientists validate the biology. Key opinion leaders sit in a review roundtable for final go/no-go decisions. This is not a fully autonomous system and it should not be. Drug development has real consequences for real patients, and AI hallucinations in this context are not embarrassing but dangerous.

The Verdict

Convexia is making an audacious bet. The idea that AI agents can replicate or replace the institutional knowledge of pharmaceutical companies is bold, possibly premature, and exactly the kind of swing that could produce an outsized outcome if the computational models prove reliable.

The risk is obvious. Computational biology models are improving fast but they are not yet reliable enough to eliminate wet lab validation. If Convexia sources an asset, evaluates it computationally, and advances it to clinical trials based on models that turn out to be wrong, the financial and human costs are significant. The pharma industry is conservative for a reason. Drugs that look good in silico fail in vivo all the time.

The $199 per month entry point for the Asset Discovery Agent is clever because it generates revenue and user feedback while the more ambitious parts of the platform mature. If that tool proves useful to VCs and biotech scouts, Convexia builds credibility and data that feeds back into the full pipeline.

In thirty days, I want to see how many paying users the Asset Discovery Agent has attracted. Sixty days, I want to know whether they have identified and acquired their first drug asset. Ninety days, the question is whether the computational evaluations are producing results that hold up when compared against traditional evaluation methods. The vision is compelling. The execution requires getting biology right, and biology does not care about your pitch deck. If the models work, Convexia could compress years of drug development into months. If they do not, this is an expensive lesson in the gap between computational predictions and biological reality.