Pedro Griff Marcincowski
Compensation partner who builds AI agents that prepare comp analysis
I work in compensation at Google. After hours I build AI agents that prepare comp analysis, and they never make the pay decision. I test where a language model should stop, where plain code should take over, and what a person has to decide.
Models write the words. Code computes the money. People decide.
How I build
1. Keep identifiers out of the model
Before any model call, names, national IDs and salaries become placeholders. They map back after the model answers, so the prompt doesn't hold the real values.
2. Let a person approve
Comp decisions wait on approvals that can take days. The workflow saves its state at each step and waits for a named approver before anything reaches a manager.
3. Code computes the money
The model doesn't do arithmetic. Range math, compa-ratios and vesting run in plain code with tests. The model explains the results.
Projects
comp-insight-advisor
An agent I'm building that reads a synthetic company's pay, equity and offer data and drafts a retention brief. A comp partner reviews the draft and decides what goes to a VP.
VestingSim Engine
Equity valuation engine modeling 48-month front-loaded enterprise schedules with exact share conservation (Largest Remainder Method) and parallelized Brownian Motion price path simulations.
Relata: CVM Regulatory Reporting & CPC 10 Accounting
Deterministic open-source regulatory compliance platform for Brazilian public companies (Companhias Abertas). Features mathematical invariant auditing across FRE Items 8.1–8.7, continuous dividend Black-Scholes option pricing with monthly accrual schedules, Hamilton-Hare largest remainder share distribution (∑ s_i ≡ S), LGPD Art. 18 surrogate tokenization, CVM Sistema Empresas.NET XML generation, and Model Context Protocol (MCP) server for Claude Desktop & Cursor.
NextComp Salary Range Workbench
Algorithmic salary band modeling engine enabling interactive midpoint and spread % adjustments with real-time recalculation of range minimums, maximums, and Incumbent Impact Summary telemetry (Green-Circle correction budget, Red-Circle governance, and Department Compa-Ratios).
PeopleAgentMesh: Enterprise Multi-Agent Orchestrator
My first multi-agent system. Human approval steps, PII tokenization and evals on synthetic HR cases. I'd build it differently now.
CompFlow Total Rewards Platform
Enterprise compensation calibration microservice automating employee review planning (merit %, formula bonuses, equity refreshes, department budget burn) and new hire offer approvals with location tiers and VP exception escalation.
AlgoCore CS Lab
Computer science laboratory implementing contiguous memory dynamic arrays, ring buffers, open-addressing HashMaps with Tombstone recycling, and augmented AVL trees for sub-millisecond percentiles.
Homelab Kubernetes Cluster
Declarative, immutable bare-metal Kubernetes cluster infrastructure powered by Talos Linux. Automated certificate renewals, resilient container orchestration, and GitOps-driven delivery.
PeopleAgentMesh: Enterprise Multi-Agent Orchestrator & HITL Governance
Agent platform for sensitive People Operations, Compensation Planning, and Talent Promotion Reviews across Brazil 🇧🇷, United States 🇺🇸, and Canada 🇨🇦. Features a stateful DAG with pause/resume execution, zero-retention PII tokenization, statutory labor verification, and interactive executive review checkpoints.
Relata: CVM Regulatory Reporting & CPC 10 Accounting Platform
Deterministic open-source regulatory compliance platform for Brazilian public companies (Companhias Abertas) under CVM Resolução nº 80/2022 Seção 8 and CPC 10 (R1) / IFRS 2. Integrates continuous dividend Black-Scholes option pricing, exact integer share conservation (Hamilton-Hare ∑ s_i ≡ S), zero-retention LGPD Art. 18 surrogate tokenization, and a Model Context Protocol (MCP) server that exposes these tools to Claude Desktop and Cursor.
NextComp: Salary Range Workbench & Incumbent Impact Telemetry
Interactive compensation modeling workbench engineered with Next.js 15 App Router, PostgreSQL + pgvector, and Deterministic Mathematical Formulas. Allows compensation leaders to adjust midpoints and spreads with immediate recalculation of band bounds and real-time incumbent circle telemetry.
Mathematical Rigor & Formula Identities
Exact formula reciprocal invariants verified via Vitest:
• $\text{Min} = \text{Mid} / (1 + \text{Spread} / 2)$
• $\text{Max} = \text{Min} \times (1 + \text{Spread})$
• $\text{Green Circle Budget} = \sum_{\text{Base} < \text{Min}} (\text{Min} - \text{Base})$
Incumbent Circle Impact Summary
Instantly flags underpaid incumbents below range minimum (Green Circle) and overpaid incumbents exceeding range maximum (Red Circle), quantifying the exact dollar budget required to restore internal equity.
Live Interactive Simulation
Test market inflation bumps (+3%, +5%) or normalize spreads across 18 career bands across Engineering, Product, and Design without page reloads.
CompFlow: Total Rewards Planning & Offer Orchestration Platform
Enterprise compensation calibration engine managing both Current Employee Annual Review Planning (Merit %, Formula Bonus, Equity Refreshes & Department Budgets) and New Hire Candidate Offers with location-tiered salary bands and automated VP exception escalation.
📝 Employee Merit & Bonus Proposal
Cycle: FY2026⚡ Agentic Calibration Stream
READY🤝 Candidate Offer Package Builder
Recruiting Workflow📄 Offer Approval Lifecycle
OFFER DRAFT📊 FY2026 Departmental Budget Depletion Gauges
Atomic Redis CountersReal-time budget pool governance tracking merit salary burn, performance bonus reserves, and equity allocation ceilings across engineering departments.
🔌 CompFlow Microservice REST API Surface (`/api/v1`)
⚡ K8s Service OnlineExplore the live asynchronous FastAPI microservice endpoints backed by PostgreSQL 16, Redis 7 caching, and ArgoCD GitOps on Talos Kubernetes.
VestingSim: Equity Valuation & Monte Carlo Simulation
Calculates 48-month deterministic vesting schedules using the Largest Remainder Method and runs real-time Geometric Brownian Motion (GBM) Monte Carlo path simulations.
📊 Grant Parameters & Tranche Distribution
Deterministic Math🎲 Real-Time Monte Carlo Risk Engine
AlgoCore: Interactive Algorithmic Arena
Empirical proof of hardware cache lines, modulo wrap-around ring buffers, and balanced tree order statistics.
⚡ FIFO Queue Race: O(N²) Array vs O(1) RingBuffer
Microseconds MatterSimulate processing 15,000 compensation review approval queues in FIFO order.
🌳 Order-Statistic AVL Tree: O(log N) Percentiles
50,000 SalariesMove the slider to find exact salary percentiles in sub-millisecond O(log N) time without sorting.
Homelab Kubernetes: Declarative Bare-Metal Cluster
Immutable infrastructure-as-code powered by Talos Linux, GitOps workflows, and automated container orchestration.
☸️ Cluster Topology & Architecture
HEALTHY🛡️ Security & GitOps Invariants
Declarative IaC- CompFlow Platform: Distributed FastAPI microservice with dedicated PostgreSQL 16 StatefulSet (`local-path-storage`) and Redis 7 cache.
- Immutable OS: Talos Linux provides no SSH, zero shell access, and API-driven declarative configuration.
- GitOps Sync: ArgoCD operator continuously reconciling declarative Kustomize applications (`apps/comp-flow-platform/`).
- Automated TLS: Ingress-NGINX with cert-manager cluster issuer certificates.
- Observability: Prometheus metric scrapers on `/metrics`.
Interactive Prototypes & Simulations
Experimental browser-based game prototypes and interactive tools created in the lab.
Design Thinking for Compensation
An interactive simulation demonstrating human-centered design thinking methodologies applied to organizational total rewards.
Aegis Core
Fast-paced defensive arcade prototype testing HTML5 Canvas physics and spatial collision detection.
Neon Drift
Retro-futuristic responsive driving mechanic experimenting with game loops and input handling.