I ship production apps that businesses run on,
and AI agents and models that operate real software.
Prague · ČAIO 2026 finalist · open to contract work
Pripoj.to LAI Kestrel VoltPlán Virlya felix.cumarav@gmail.com
| Project | What it is | Built with | |
|---|---|---|---|
| 🚐 | Pripoj.to | trailer rental with keyless pickup through smart locks | React Native · Next.js · Firebase · Stripe · Igloohome |
| 🍱 | Meal-box ordering | orders, menus, delivery zones, payments and invoicing for a food-delivery company | web shop · admin · CRM · payments · bank import |
| 🎮 | Multiplayer game | real-time iOS / Android game for a client | Expo · Rust engine · WebSockets · Redis |
| 🐧 | LAI | AI agent that drives the whole Linux desktop through the accessibility tree | Python · AT-SPI · X11 · MCP |
| 🦅 | Kestrel | desktop agent with memory and its own screen, built on opencode | TypeScript · Bun · LAI |
| ⚡ | VoltPlán | forecasts EV-charging demand for every zone in Prague | LightGBM · scikit-learn · Streamlit |
| 🪄 | Virlya | ten agents turn one brief into a full social-media campaign | multi-agent · image generation · Telegram |
Agents are half of it. The other half is software real businesses run on: payments, auth, scheduling, invoices, hardware, store releases.
Pripoj.to trailer rental · full product
Rent a trailer from your phone and pick it up with nobody there. React Native app with phone-number sign-in, a live map of trailers and availability-checked booking; Stripe with 3-D Secure; Firebase Cloud Functions that turn a paid booking into a time-limited Igloohome lock PIN, extend rentals and clean up expired codes; a Next.js admin with the fleet's real-time state, pricing and stats.
%%{init: {'theme':'base','themeVariables':{'primaryColor':'#161b22','primaryTextColor':'#e6edf3','primaryBorderColor':'#e0af68','lineColor':'#e0af68'}}}%%
flowchart LR
U[app<br/>map · booking] --> A{availability<br/>check}
A --> S[Stripe<br/>3-D Secure]
S -- webhook --> F[Cloud Functions<br/>auto-extend · PIN cleanup]
F --> L[Igloohome PIN<br/>keyless pickup]
F --> M[admin<br/>live fleet · pricing]
Meal-box ordering system client · in development
Replaces a meal-box delivery company's web shop and spreadsheets with one system: weekly menus built from a dish database with allergens, an order cut-off every Thursday 23:59 for Mon–Fri delivery, delivery zones drawn on a map with manual approval at the edges, online payments, bank-statement import that matches payments to invoices automatically, accounting export, and a CRM over all of it. Specified down to the domain model, API, security, migration and running costs before the first line of code.
Multiplayer mobile game client · in development
Real-time iOS / Android game with a Rust rules engine, WebSockets, matchmaking and bots; load-tested at 300 concurrent players.
Also: Forksy, a nutrition tracker, and PhAuth, phone-number sign-in, both React Native / Expo.
Desktop agents mostly fail on grounding: a model looking at a screenshot guesses coordinates and guesses wrong. LAI reads the AT-SPI accessibility tree, the desktop's DOM, fuses it with four other sources of truth and invokes widgets through their own accessible actions. Pixels are the fallback, not the plan.
$ lai do "open the text editor, write a haiku about X11, save it to ~/haiku.txt"
▸ app_open {"name": "Text Editor"}
+ ✓ window up after 0.6s
▸ ui_snapshot
+ ✓ 57 accessible elements
▸ ui_type {"ref": 12, "text": "Pixels in a grid…"}
+ ✓ set via a11y_set_text
▸ ui_click {"name": "Save"}
+ ✓ invoked via a11y_action, not a pixel click
▸ file_read {"path": "~/haiku.txt"}
+ ✓ effect verified on disk
@@ completed: 8 steps in 34s @@%%{init: {'theme':'base','themeVariables':{'primaryColor':'#161b22','primaryTextColor':'#e6edf3','primaryBorderColor':'#7aa2f7','lineColor':'#7aa2f7','clusterBkg':'transparent','clusterBorder':'#30363d'}}}%%
flowchart LR
subgraph perceive [perceive]
A[AT-SPI tree<br/>role · name · state · bounds]
W[window tree]
P[process tree]
S[screen capture]
O[OCR fallback]
end
A & W & P & S & O --> F{{grounded<br/>state}}
F --> ACT[act<br/>a11y action first,<br/>pixels last]
ACT --> V[verify<br/>re-read state]
V -- effect seen --> F
V -. "same failure ×3<br/>or no-op success ×6" .-> R[refuse & replan]
R --> F
Kestrel puts LAI's hands under sst/opencode's agent loop and adds what a plugin can't: memory of the machine it lives on, its own X server so it never takes your mouse, and loop discipline. Same tools over MCP took opencode 18 calls; Kestrel took 7. The tools were identical. The whole gap was the loop.
Python · TypeScript · Bun · X11 / AT-SPI · MCP · ~47k lines in LAI
Forecasting · VoltPlán
2030 EV-charging demand for every grid zone in Prague, a station-type recommendation per zone, and a V2G layer that resolves grid overloads by moving charge windows instead of building hardware.
%%{init: {'theme':'base','themeVariables':{'primaryColor':'#161b22','primaryTextColor':'#e6edf3','primaryBorderColor':'#9ece6a','lineColor':'#9ece6a'}}}%%
flowchart LR
D[(70+ features<br/>population · housing<br/>transit · grid)] --> L[leakage guard<br/>drop target-derived cols]
L --> M[LightGBM<br/>P10 · P50 · P90 heads]
M --> H[517-zone holdout<br/>MAE 4.7 · P@50 0.84]
M --> C[station type<br/>AC small · AC mid · DC hub]
M --> G[V2G load shifting<br/>delay charge windows]
ČAIO 2026 AI Startup track · 2,378 training zones · trained on CPU (Lunar Lake, no CUDA) with sklearnex · Streamlit demo
Multi-agent · Virlya live product
Ten specialised agents turn one brief into a multi-platform campaign. A human signs off on the plan before anything is generated, and approves or rejects finished posts from Telegram.
%%{init: {'theme':'base','themeVariables':{'primaryColor':'#161b22','primaryTextColor':'#e6edf3','primaryBorderColor':'#bb9af7','lineColor':'#bb9af7'}}}%%
flowchart LR
B[brief] --> R[web research] --> S[strategy] --> H{human<br/>approves plan}
H --> C[on-brand copy] & I[image generation]
C & I --> P[posts<br/>IG · FB · X · LinkedIn · TikTok]
P --> T{Telegram<br/>approve / reject}
Tools for working with agents
claude-session-dots: a panel dot per Claude Code session that turns red when one is waiting on you.
learn-claude-code: an AI tutor where answer keys are committed before the question is shown, so the model can't grade itself generously.
claude-widget: 5-hour and weekly usage limits in the Cinnamon panel.
1 beat a dumb baseline before believing a smart model
2 fix grounding before tuning prompts; most agent failures are perception failures
3 verify the effect, not the call; a click "succeeds" the moment the event is sent
4 build on a strong base and spend the effort where the gap actually is
Agents tool use · MCP · RAG · vector databases · multi-agent pipelines · evals
Systems TypeScript · Rust · Node · Bun · PostgreSQL · Redis · Docker · Caddy · Linux (X11, AT-SPI)
Product React Native / Expo · Next.js · Firebase · Stripe · Spring Boot

