Maxim Berg

I have spent about twenty years building payment infrastructure and running the teams behind it. Over the last year I shipped AI agents that do real work in production. HRPulsar is the third company I have started, and the first one built in the open.

Now

2026 –

HRPulsar

Open-source talent and competency management platform. Founder, one engineer, building in public.

Companies pay for AI tools and still cannot tell who gets value from them. Peer reviews say everyone is doing great, self-assessments say everyone is an expert, the work says something else. HRPulsar measures skills from real work instead: competency frameworks, 360° assessments, development plans, grade ladders and exams in one place.

Open core: the community edition is AGPL and self-hosted. The cloud version is how the company earns money.
Multi-tenant, API first, LLM-based competency generation and semantic search on pgvector.
Python and FastAPI on the backend, Next.js and React on the frontend, PostgreSQL and Redis underneath.
Public changelog, roadmap marked as a roadmap, and a white-label version for the German market.

hrpulsar.com Source on GitHub

Selected work

2026

LetAgentPay

Spend control for AI agents. Own product.

Sits between an agent and the payment system. Eight deterministic policy checks run in milliseconds and return approved, pending or rejected before the agent spends anything. Policies can be written in plain English and are compiled into rules by the Claude API.

FastAPI, SQLAlchemy async, PostgreSQL, Redis, Next.js, TypeScript, Claude API, x402 and USDC on Base

Source on GitHub

Async-first backend with Redis-backed spending counters on time-bucketed keys. Pending requests hold funds, so parallel agents cannot double-spend.

Policy engine: agent status, category, per-request limit, schedule, daily, weekly and monthly limits, total budget
Live dashboard over server-sent events, no polling
Four ways to integrate: REST API, Python SDK on PyPI, TypeScript SDK on npm, MCP server
USDC micropayments over x402 with depeg protection
Notifications by email, Telegram and Web Push

2026

DevBot

Autonomous software development pipeline.

Takes an open Jira task, plans it, writes the code with Claude Code, watches CI, gets a review, merges and releases. Fully stateless: all state lives in Jira and GitLab, so it restarts safely and scales without a database.

Python, Claude Code CLI, asyncio, Jira and GitLab APIs, MCP servers, Redis, Kubernetes

Claude Code runs as a subprocess rather than through the SDK. That allows tool restrictions, credential isolation and per-invocation MCP server injection.

21-state configurable workflow with human approval gates via stop labels
Global and per-group circuit breakers, exponential cooldown, pipeline retry with CI log analysis
Comment-driven directives: users inject instructions mid-workflow with an @mention
Release windows by day, hour and timezone, with a holiday API
Three LLM providers supported

2026

KnowledgeBot

Enterprise RAG for Slack.

Answers employee questions in Slack strictly from the company's Confluence, scoped to what each user is allowed to see, with citations back to the source page. Also audits documentation quality and files Jira tasks with duplicate detection.

Python, Claude API, ChromaDB, sentence-transformers, SQLite, Slack Bolt, Confluence and Jira APIs, Docker

Multi-tenant vector store with one ChromaDB collection per Confluence space and Slack group-based access control.

HyDE search: generates a hypothetical answer first to improve embedding matches
Zero-downtime reindexing: builds the replacement collection in the background and swaps it in one operation
Jira task creation with duplicate detection via vector similarity plus an LLM check
Documentation audit with six parallel checkers and LLM-suggested fixes
Per-feature API cost attribution with overspend alerts in Slack

2025 – 2026

SalesBot

B2B client onboarding run by an AI agent.

A Telegram bot that takes a business client from first message to a live widget: negotiates commercial terms in natural language, runs KYC against government business registries, generates and signs the contract, then verifies the deployment.

Python, aiogram, FastAPI, Claude and GPT-4o with fallback, MySQL, SQLAlchemy async, Redis, Kubernetes

Clean Architecture with a composable 17-state FSM. Regional subclasses define base states, product flows inject theirs at runtime. One LLM provider abstraction with fallback chaining and per-call cost metering.

Claude collects, negotiates and extracts structured contract data from the conversation
KYC risk score from 0 to 1: auto-approve or route to moderation
Contract generation from Jinja2 templates per region, electronic signing through EDO
Widget deployment check: HTTP, HTML parsing and JS loading
Redis distributed locks against race conditions from rapid multi-message input

2017 – 2026

Multi-region payment platform

Gift cards, loyalty and payments across several countries. Nine years in production.

Distributed platform for the full gift card lifecycle, hierarchical commission calculations and financial reporting, with a bidirectional sync to the 1C ERP. Custom metaclass-based ORM with thread-safe caching and temporal queries.

Python, Celery, Flask, MySQL master/slave, Redis, Memcached, RabbitMQ, Pandas

Inventory forecasting from 168-bin hourly profiles (hour by weekday), monthly seasonality and IQR outlier filtering, predicting depletion up to 30 days ahead.

Seven commission types across partner hierarchies with retroactive recalculation
CAS-based distributed task chain tracking in Redis, beyond Celery's native primitives
1C ERP sync in both directions: payments, agreements, invoices, tax agent data
Sixteen report types in PDF, Excel and CSV, scheduled monthly for partners
Alert escalation from Slack to Telegram, SMS and email with Redis cooldowns

Contact

For HRPulsar, consulting on AI agents, or anything else. I read and answer every message myself.