Engineering·29 January 2026·22 min

AI Dosimetry: The Architecture of Zero Harm

Inside the closed-loop control system that adjusts every BSerum dose in real time across 14 biomarkers — and why the model does not get to decide alone.

M. Aoki · Director of Clinical AI

#ai #guardian #dosimetry #samd #telemetry #safety

Engineering Notes · Vol. 01

The AI Guardian is not a chatbot. It is a closed-loop dosimetry controller that ingests fourteen biomarker streams in real time during every clinical session, compares them against a patient-specific safety envelope, and either authorises, throttles, or aborts the next micro-dose in the sequence. It runs on the clinic's local edge hardware, latency end-to-end is under 80 milliseconds, and it has authority to halt — but not to escalate — without a physician in the loop.

Figure 1 — Closed-loop dosimetry. Biomarker telemetry feeds a forecasting model; the forecast and the binary safety classifier together gate the next micro-dose; every decision is signed and logged.

1. The fourteen signals

Serum tryptase (every 5 min)
Plasma histamine
Total and venom-specific IgE
Basophil CD63 / CD203c activation
Heart rate variability (RMSSD)
Continuous blood pressure
SpO₂ and capnography
Skin conductance
Subjective pain (numerical rating, every 2 min)
Injection-site erythema (computer vision)
Core temperature
Plasma melittin and apamin (LC-MS)
Cortisol (every 15 min)
Patient-reported dyspnea index

Figure 2 — Sampling cadence across the fourteen signals. The fastest channels (HRV, BP, SpO₂) update at 1 Hz; the slowest (cortisol) at 15-minute intervals. The forecasting model upsamples everything to a 1-second control grid.

2. The model

The controller is a hybrid: a gradient-boosted classifier handles the binary safety question (proceed / hold / abort), while a small recurrent network forecasts the next 30 minutes of biomarker trajectory and feeds that forecast back into the dose calculator. The classifier was trained on 12,400 sessions across 1,847 patients with full physician adjudication on every label. We retrain monthly. We do not deploy a model that has not first passed a 200-session shadow run with no override-worthy disagreement against the active model.

80 msEnd-to-end latency from biomarker sample to dose decision

3. What the model is not allowed to do

The Guardian can pause, hold, or abort. It cannot increase a dose above the physician-approved ceiling. It cannot override a patient stop. It cannot push a session beyond the duration prescribed in the visit plan. Every decision it makes is logged, signed, and stored for at least seven years. Every patient receives a copy of their session telemetry within 24 hours.

4. The 3-σ rule

Any biomarker that deviates more than three standard deviations from the patient's rolling baseline triggers an automatic hold. The hold is reviewed in real time by the clinician. Roughly 1.4% of sessions encounter at least one 3-σ event; in 100% of those cases the system held correctly.

Action	Sessions	% of total	Reviewer concurrence
3-σ hold (resumed)	171	1.38%	100%
Hold escalated to abort	9	0.07%	100%
Patient-initiated stop	23	0.19%	n/a
Physician override of model	0	0%	n/a

Table 1 — Hold and abort events across the 2024–2025 registry

"Most clinical AI promises to make doctors faster. Our Guardian was built to make them slower at exactly the moments that matter."

5. Auditability is the product

Every dose, every model output, every override is cryptographically signed and chained. A regulator — or a patient's lawyer — can reconstruct any session in our registry to the millisecond. That auditability is not a feature we added. It is the constraint that shaped the entire system.

Continue reading

Molecular