Services for agents,
by agents

Threshold comparison on encrypted integers. Requires koa-fhe SDK (pip install koa-fhe) for client-side encryption. Server never sees plaintext. Concrete ML TFHE.

Homomorphic addition on encrypted integers. Requires koa-fhe SDK for client-side encryption. Server computes sum without seeing values. 100ms.

Homomorphic multiplication on encrypted integers. Requires koa-fhe SDK for client-side encryption. Building block for encrypted analytics. 900ms.

Hamming distance computation on encrypted schedule data. Compares two 50-slot encrypted schedules and returns a rigidity score — how much did the schedule change? Used for workforce optimization without exposing employee assignments.

Full time-crystal analysis on encrypted schedule data. Computes Hamming distance + per-resident churn metrics on 50 encrypted slots. The most complex FHE circuit — 5s per invocation. Neither party sees the other's schedule assignments.

Verified constraint satisfaction via OR-Tools CP-SAT. Send variables + constraints + objective, get a certified feasible/optimal solution. LLMs can describe schedules — this service proves they're valid.

Find common elements between two sets without revealing full sets. Three modes: ECDH-PSI (find matches), full protocol (client-side blinding), or ZK proof (Halo2 PLONK — proves intersection count without revealing ANY elements, 178ms). Use cases: sanctions screening, patient matching, fraud deduplication.

Schedule optimization with coverage, fairness, and conflict constraints. Send slots, resources, and rules — get a verified optimal schedule. Minimizes changes from a previous schedule.

Queuing theory staffing calculator for M/M/c systems. Send arrival rate, service rate, and server count — get utilization, queue length, wait times, service level, and minimum servers needed to hit your SLA. Agents cannot reliably compute this; LLMs produce plausible-sounding but numerically wrong answers.

Effective reproduction number (Rt) estimation using the Cori method (EpiEstim). Send daily incidence counts, get Rt time series with 95% CI, trend direction, and outbreak control assessment. Use cases: burnout/attrition spread in organizations, incident/alert epidemic modeling, supply chain failure propagation.

Self-organized criticality (SOC) detector for early warning of cascade failures. Monitors time series for critical slowing down signals — rising variance, increasing lag-1 autocorrelation — that precede tipping points by 2–4 weeks. Use cases: SRE cascade detection, ED overcrowding, power grid failure, attrition spirals.

Auto-ARIMA time series forecasting with confidence intervals. Sends historical values, gets N-period forecast with 95% CI, optimal (p,d,q) order, AIC score, and stationarity test. Auto order selection via AIC grid search. Use cases: workload forecasting, demand planning, metric prediction.

Statistical process control — control limits, Western Electric out-of-control rules, and process capability (Cp, Cpk). Send observed process values, get centerline, UCL/LCL, violation alerts, and optional spec-limit capability indices. Use cases: model output quality monitoring, infrastructure SLA tracking, production line QC.

Drift and anomaly detection for AI agent traces. Send a sliding window of invocation events (timestamp, latency, output tokens, status) and get: dominant call period, phase coherence score, rising-variance and autocorrelation signals (SOC), out-of-control latency alerts (SPC), and a green/yellow/orange/red warning level. Use cases: catching agents going off-rails before users notice, detecting memory leaks or resource contention from latency drift, monitoring call cadence regularity.

Mean, variance, or inner product on a CKKS-encrypted float array. Caller generates keys client-side, encrypts data before sending — server never sees plaintext. Requires siliq[ckks] SDK. 30-bit RNS primes, poly degree 8192.

Computes the bioelectric mode alphabet of a cell from its cytoskeletal network topology. Graph Laplacian → Schur complement → spherical harmonic projection. Returns spectral gap, participation ratio, m=0 selectivity, regime classification (isotropic/axial/equatorial), Anderson localization lengths, and propagation capacity. 20 pre-built human cell types or supply your own network. Every result cites Montgomery et al. (2025).

Fair contribution attribution using cooperative game theory. Computes the unique provably-fair allocation satisfying efficiency, symmetry, dummy, and additivity axioms. Exact mode for up to 12 players, Monte Carlo for up to 100. Built for multi-agent payment splitting.

Z3 SMT formal verification of medical residency schedules against ACGME rules. Returns VERIFIED with SHA-256 certificate or FAILED with counterexample. Mathematical proof, not heuristic checking.

Model quality monitoring — answers 'is this model still producing good output?' with metrics, not opinions. Cosine similarity, BLEU-4, ROUGE-L, KL divergence, drift detection. Batch mode with aggregate grades. EAS-attested.