AI translates everyday questions into candidate designs, explains trade-offs and notices missing variables, but it cannot redefine the outcome after seeing inconvenient data. Retrieval provides evidence about measurement validity and known risks from trusted sources. A protocol agent proposes alternatives; a sceptic agent identifies confounding, expectancy effects and underpowered designs; a safety layer blocks interventions outside the product's permitted scope. Statistical computation is separated from language generation so the narrative must reflect the calculated result. Personalisation concerns feasibility and explanation, not unsupported medical inference. The user retains control of data, sharing and deletion, and professional judgement remains essential wherever consequences become clinical.
Personal Science
“What if everyone could run rigorous experiments on their own life?”
The beginning
Personal Science is a private research companion for turning questions about health, learning, work and habits into careful small experiments. People already change routines and collect measurements, but the process is usually guided by anecdote, generic advice or correlations discovered after the fact. I imagine a system that helps someone formulate a question before looking for an answer, design a safe protocol, understand what one person's data can and cannot establish, and preserve negative results. It is not a machine for optimising every hour of life. It is a way to replace vague self-tracking with modest, inspectable learning while protecting the intimacy of the data involved.
Wearables and personal apps generate abundant measurements but rarely create valid personal evidence. Baselines shift, interventions overlap, outcomes are chosen after results appear and normal variation is mistaken for causation. Commercial incentives reward engagement and positive findings, while failed experiments disappear. Health-related self-experimentation can also cross quickly from harmless curiosity into unsafe behaviour or delayed professional care. Generic AI advice intensifies the risk by sounding personalised without using a valid design or clinical context. The missing layer is not more data; it is protocol, uncertainty, privacy and a clear boundary between personal exploration, clinical research and medical decision-making.
What it could become
The product would guide a user from an informal question to a preregistered N-of-1 protocol. It helps define the intervention and outcome, establish a baseline, identify confounders, choose feasible measurements, schedule phases and state stopping conditions. During the study it captures adherence and context without constantly exposing interim results that could change behaviour. At completion, statistical modules estimate effects and uncertainty, test sensitivity to missing data and generate a plain-language report that distinguishes observation from conclusion. Users can share a protocol or result selectively with a coach, clinician or research cohort while keeping raw data local or encrypted. Unsafe topics trigger boundaries and professional guidance rather than experimental instructions.
For whom
- Curious individuals and quantified-self communities
- Coaches and learning specialists
- Clinicians supporting low-risk behavioural experiments
- Researchers designing decentralised studies
Core capabilities
- Conversational protocol design
- Private multimodal data integration
- Within-person statistical analysis
- Bias, safety and confounder checks
The project becomes meaningful only when a new technical possibility is translated into a clear human advantage, an experience people can understand, and a system capable of earning trust over time.
Intelligence and mathematics
N-of-1 trial design, interrupted time series and Bayesian hierarchical models allow careful inference from small, noisy personal datasets. Randomisation and counterbalancing reduce time and order effects when appropriate; sequential methods define in advance when enough evidence exists; state-space models separate slow baseline change from short intervention response. Measurement-error models acknowledge that wearable proxies are not the underlying phenomenon. Results are expressed as posterior distributions, effect ranges and decision thresholds rather than binary success. When cohorts voluntarily combine compatible protocols, partial pooling can improve estimates without pretending participants are interchangeable. The mathematics is designed to resist overclaiming, not to make a personal chart look scientific.
How it might live
A consumer subscription could support low-risk questions about learning, focus, routines and wellbeing, with professional plans for coaches, educators and researchers designing supervised protocols. Clinical applications would require separate regulated development and should never be smuggled into the consumer product through suggestive language. Trust depends on local-first or strongly encrypted data handling, transparent analysis, no advertising and no sale of behavioural profiles. Research partnerships could fund opt-in decentralised studies where participants understand what is shared and receive results back. The durable advantage would be a library of validated protocol patterns and an architecture that preserves privacy and statistical integrity—not the accumulation of intimate raw data.
For me, a venture is more than an interesting technology. It needs a narrow first user, a repeated problem, a distribution path, a credible advantage and a reason to improve as more people use it. I would test those conditions before deciding whether this idea should become a company, a product, an open technology or an ongoing research programme.
Rules for making it real
- 01
Private by architecture.
- 02
A personal result is not a universal claim.
- 03
Safety boundaries come before curiosity.
- 04
Show uncertainty in language people understand.
From question to company
- 01Frame
Launch question-to-protocol experience.
- 02Prototype
Integrate manual and wearable data.
- 03Prove
Pilot learning and low-risk habit experiments.
- 04Build
Add professional collaboration and research cohorts.
What could go wrong
Serious imagination includes the possibility that an idea should change radically—or should not exist. These are the tensions the project would need to resolve:
- Users experimenting with unsafe health interventions.
- Overinterpreting noisy or confounded results.
- Personal data creating insurance or employment harms.