A physiological coaching agent built on Pison wearable data.
"Reads your biometric data. Thinks offline. Coaches you over iMessage."
MIT Media Lab · MAS.664 · Spring 2026 · Daniel Tecum-Puac & Jerry Chen
About Pison
Pison makes wearable devices that capture physiological signals — HRV, strain, recovery, sleep quality, and circadian rhythm data. The Pison AI Agent is a coaching system built on top of that data. It reads your biometric signals, analyzes them through a deterministic science pipeline, and delivers personalized coaching over iMessage.
The agent operates on two paths. An offline path runs every few hours: it ingests wearable data, runs an 18-module physiological analysis, builds a structured briefing from the results, and stores it. An online path fires in under a second: when you send an iMessage, the agent loads your briefing and conversation history, reasons about your current state, and replies — in plain text, the way iMessage expects.
The coaching relationship evolves over time. Early conversations report raw data clearly. As the agent learns your patterns, it starts identifying trends, making predictions, and communicating more compactly. The relationship stage is determined by message count and data freshness — the agent always knows whether this is message 3 or message 50.
ONLINE PATH (per iMessage, target < 1 second)
─────────────────────────────────────────────────────
User sends iMessage
→ Load briefing + conversation history from Database
→ AI Coach (reasoning over your current physiological state)
→ Plain-text iMessage reply
The offline path handles the expensive work so the online path stays fast. The analysis pipeline models your circadian performance rhythm, tracks HRV night profiles, computes strain-recovery balance, and aligns self-reported wellbeing against measured signals. This structured briefing — not a raw data dump — is what the coach actually reads when you text.
Research Findings
Experiments run at MIT Media Lab, Spring 2026
Context Architecture (HW7)
We evaluated how different context structures affect coaching quality — raw biometric snapshots vs. structured briefings, single-session vs. 90-day longitudinal history, and schema versioning with data freshness guards. All three experiments ran on real biometric data from two Pison users. Structured briefings produced measurably better coaching responses. Longitudinal history beyond 30 days added meaningful signal. Schema versioning proved necessary to prevent stale data from corrupting advice.
90-day READY% trajectoryHRV night profileStrain vs. recovery balance
Scale Experiments (HW8)
We stress-tested the system with 30 simulated agent instances sharing a single deployment, measuring concurrent throughput, offline sync scale, and prompt cache economics.
DB connection pool saturates at N>10 — one-line config fix
Offline sync scale
Serial vs. concurrent briefing generation for N=1–30
Concurrent 21× faster at N=30 (23s vs 8 min), zero rate-limit errors
Prompt cache hit rate
Cache economics across 30-user population
97% hit rate at N=30 — 81% cost reduction vs. no caching
API Reference
This API is not publicly accessible. Endpoints are authenticated or webhook-gated. Documented here for research transparency.
Method
Endpoint
Description
POST
/webhook/linq
Receives incoming iMessages and triggers the coaching response
GET
/auth/google/start
Initiates Google Calendar OAuth flow
GET
/auth/google/callback
OAuth callback — stores calendar access token
GET
/health
Service health check
The agent is a closed loop between Pison wearable hardware, the Linq iMessage platform, and Anthropic's Claude API. It is not open for external access.
