Algorithmic trading platform
Nine event-driven services with paper and live execution on perpetual futures, three-level risk control with a kill switch, and a decision log that explains every trade taken or refused.
567 merged PRs in 8 months
AI-native product and functional architect
Sergei Kobe
I design the product and the system. AI agents build it. Review gates decide what ships.
Seven months, about three thousand pull requests across 21 repositories: fintech, PropTech, retail and my own agent infrastructure. I run the product and design work and write the plans, contracts and quality gates. Agents and teams write the code.
≈3,000pull requests since March 2026
88%written by AI agents under my plans
348multi-agent workflows run
2,613review findings logged and tracked
21repositories in active work
5MCP servers designed
How I work
The same loop runs on every project, from a trading platform to a hiring tool. It is written down, enforced by hooks and checked by reviewers who know nothing about the session.
Start from the person who will use it
Every product starts with the people who will use it: their day, their pain and the decision they need to make. The answer becomes a short brief with what we promise, what we refuse to promise and how we will know it works.
a decision log on every projectThe plan is the contract
Every feature starts as a work package: data model, API contract, edge cases, test plan. When the tests and the code disagree, the plan decides.
264 architecture decisions on recordBlind build
Two agents work from the same plan in separate worktrees. One writes the tests, the other writes the code, and neither sees the other's work until they meet.
405 test writers · 408 implementersReview on both ends
An architect agent reviews the plan before any code exists. Reviewers with no session history read the code before merge, and refuters try to kill each finding.
≈875 review agentsShip by runbook
Production moves only from main, through a written runbook, with a smoke test after every deploy and a rollback path decided in advance.
44 projects with living roadmapsIncidents become gates
Every incident ends as a rule, and every rule that can become a hook or a CI check does. Agents once broke one rule nine times in a night; a pre-commit gate ended it.
19 lessons · 50 feedback rules
Selected work
Clients stay unnamed. Each case carries one number that can be checked.
Real-time market data
A streaming pipeline for 400+ perpetual futures. One-minute bars become signals within seconds, and ingestion heals itself after gaps.
p50 2.2 s · p99 3.8 sAI analyst for traders
An MCP server with 21 tools over dashboards, time series and a backtester: strategy correlation and portfolio Monte Carlo on request.
16 days · 436 testsMarket analytics and alerts
Margin flows, open interest, liquidations and order-book depth across 400+ tokens, with dashboards and subscriber alerts.
400+ tokens trackedAI platform for a villa-rental company
Thirty-five internal AI tools under one sign-on: sales chat, listing quality, hiring, owner reports and staff operations. I set the product direction and decided what each tool had to do first.
448 live sales conversations in 3.5 monthsWhat guest reviews say about revenue
Reviews from two booking platforms turned into themes, revenue impact and a list of the best and worst units, so the next product decisions start from evidence rather than opinion.
744 guest reviews analysedVilla landing factory
A listing link becomes a landing site on its own subdomain, with a booking calendar and a lead form, assembled from one admin screen.
1,810 testsOne app for four audiences
A single mobile app for buyers, sellers, owners and guests, built one audience at a time, with the guest cabinet first.
737 tests · 94.7% coverage on the first nightOwner cabinet for a beauty franchise
Bookings, accounting and analytics in one place, with MCP servers over the booking and accounting systems.
≈7,000 testsMulti-tenant cabinet on a reusable kernel
A copy-shop chain's cabinet built on a kernel that I reuse to start new products in days instead of weeks.
2,427 testsAI platform for an event agency
A staged interface built around the agency's real production flow, one use case at a time.
2,218 testsTwelve specialised agents
An operations system of topic-routed agents, each in its own sandbox with egress control and git-tracked memory. When hosted CI stalled, the same setup ran a self-hosted CI by evening.
12 agents · 4 runtimesSkill set
Twelve areas I work in every week.
AI-native product management
Product & UX design
Functional architecture
Multi-agent orchestration
Trading & market-data systems
LLM applications for operations
MCP servers & integrations
Internal platforms
PropTech & hospitality
Data & analytics
Content & search
Self-hosting & operations
Where I can help
For partners, investors and teams who want to build with AI agents at production quality.
From a customer's pain to a product in use
Discovery with the people who will use it, a spec with acceptance criteria, an interactive prototype within days, then a roadmap of blocks that each ship whole. After launch, the numbers choose the next block.
Interfaces people can use on day one
UX flows, design systems and interactive prototypes, checked on real phones and contrast-audited on every page before launch.
Set up a team where agents ship production code
The playbook, blind test-and-code pairs, review gates, CI and runbooks. Your team runs the same loop from the second week.
From an idea to a running product
Spec, contracts and a plan of blocks that each ship on their own and are useful the day they land. I lead it through launch.
Execution, risk and market data
Signal-driven trading infrastructure that explains every decision it makes, with analytics and an AI analyst on top.
AI tools behind one sign-on
Sales, operations, hiring and analytics tools with role-based access, an access log and usage analytics.
AI that can read your business
Read-only MCP servers over CRM, accounting and analytics, with scoped tokens and access denied by default.
Agents on your own hardware, safely
Sandboxed hosts, egress control, self-hosted CI, observability, and a rule from every incident.
Let's talk.
Partners, investors and future colleagues: a short message with what you are building is the best start.