
Evaporative cooling towers are responsible for billions of litres of freshwater use a
year. Radiative sky cooling is a water-free alternative for heat rejection, but its
capacity depends on the weather, which makes it hard to control.

A deep RL framework for adaptive control of hybrid radiative-evaporative cooling with
thermal storage. DDPG, TD3 and SAC agents learn to jointly optimise radiative panel
utilisation and storage dispatch from real-time weather. Evaluated on 1 MW datacenter
simulations across three US climate zones — Phoenix, Houston and Seattle — using TMY3
data with 3-seed validation.

**Results**

- SAC reaches 94.4% ± 0.5% water savings in Seattle
- TD3 reaches 70.8% ± 4.3% in Houston
- DDPG gets the highest electricity savings in Houston, at 10.8%
- Every RL agent beats both fixed baselines and a rule-based heuristic given perfect forecasts

The ablation is the more interesting result: forecast integration is worth +10.2
percentage points in Seattle, +2.7 in Phoenix, and −1.1 in Houston. Forecasts help where
the weather actually varies, and do nothing where it is consistently humid.

Known limits: this is simulation, so there is a sim-to-real gap; there is no MPC baseline
yet; the ablation assumes perfect forecasts; and 48-hour episodes may be too short to
learn seasonal patterns.

**Next:** Decision Transformers (offline RL) and Dreamer-style world models on the
EnergyBench dataset (1.2 billion observations), plus carbon-aware optimisation and
multi-agent coordination across datacenter clusters.

**Stack:** Python, Sinergym, Stable-Baselines3, DDPG/SAC/TD3, PyTorch

Related: "The Digital Dust Bowl" on Medium connects datacenter cooling to the Aral Sea crisis.


Deep RL for hybrid radiative-evaporative datacenter cooling. Up to 94.4% water savings across three US climate zones.

Datacenter Cooling with Deep Reinforcement Learning


Red team agents (ReconAgent, ExploitAgent, ReportAgent) attack web applications.
Blue team agents (MonitorAgent, DefenderAgent, ForensicsAgent) defend in real-time.
The thesis: speed is the difference between caught and breached. Cerebras inference
at 1000-1700 tok/s means the blue team can actually react before attacks land.

The demo runs two scenarios - Cerebras-powered blue team blocks 70-80% of attacks;
simulated slow inference (3-5s delay) catches only 10-20%.

Targets OWASP Juice Shop (100+ documented vulnerabilities). Event bus for real-time
agent communication. Security Genome for cross-session learning with embeddings.
Monochrome frontend for live attack/defense visualization.

Built for the Cerebras hackathon. 24-hour coding sprint.

**Status:** Tier 4 in development (FastAPI + WebSocket backend). Shipping this month.

**Stack:** Python 3.11, Cerebras inference, FastAPI, WebSocket, Next.js, Docker


Autonomous AI pentesting platform. Red team attacks, blue team defends, in real-time.

SENTINEL


Peltier modules are the standard for VR thermal feedback. They're also heavy (50g),
slow (70ms response), and power-hungry (5-10W continuous). ThermoTouch replaces active
heat pumping with phase-change materials and resistive heating.

Microencapsulated n-alkane paraffin (melting point 32-34C) in a silicone matrix absorbs
heat from the skin through endothermic phase transition - no electricity needed for cooling.
Resistive heating handles the warm side. 5x5mm actuators match human thermal receptor
density (~1 per cm2) with air-gap isolation between cells to prevent crosstalk.

**Simulation results:**
- Cooling response: ~5-10ms (10-15x faster than Peltier)
- Heating response: ~250ms (matches human perceptual asymmetry - we're far more sensitive to cold)
- Weight: under 5g per actuator vs 50g for Peltier
- Power: 2W average vs 125W for a 10-module Peltier glove
- Battery life: 7.5 hours vs 18 minutes on the same 15Wh pack

FEA validation in Fusion 360 confirmed Peltier baseline (36-40C heating, 18-24C cooling)
matching published ThermalWear data within 4%.

**Limitations:** 250ms heating is borderline. PCM regeneration takes 8-10s after sustained
contact (>3 min). Encapsulation durability over 10,000+ flex cycles is untested.

**Next:** transient thermal simulation, flexible PCB fabrication, single-cell prototype
with IR camera validation, then 4x4 array and 20-participant user study.

**Status:** Phase 1 complete (FEA validation), Phase 2 in progress (prototype design)

**Stack:** Fusion 360, FEA thermal simulation, flexible PCB design


Passive thermal haptics for VR. Phase-change materials replace Peltier modules - 3.5x faster, 25x lighter, 7+ hour battery life.

ThermoTouch


An open-source builder community for creative engineering, run on four rules: build
every day, start with what you own, if it runs it counts, and read other people's code.
Everything is free and MIT licensed. 1,500+ members.

**Orcas v1.5** — a 30-day software cohort where 10 builders each shipped one project a
day using Python, locally-run models and webcam input. All 30 are open-source on GitHub.

**Orcas v2.0** — an in-person hardware cohort at Turin Polytechnic Lyceum in Tashkent,
15-24 August 2026, with hardware kits through a partnership with Nazarx Robotics.

**Stack:** Python, Ollama, webcam and gesture input, robotics hardware, GitHub


Open source for creative engineering. A builder community running cohorts where people ship something every day.

BuildCored


My automation business. Workflow automation for clients — CRM pipeline configuration,
intake automation, AI-assisted document drafting with human review gates, and task and
reminder systems tied to deadlines. Current work includes a Canadian consulting firm.

**Stack:** Pipedrive, Google Workspace, Zapier/Make, DocuSign API, Google Apps Script


My automation business. Workflow and CRM automation, intake pipelines, and AI-assisted document drafting.

Flowcored


Building a development environment designed for deep, locked-in engineering sessions.
Arch Linux + Hyprland (tiling Wayland compositor) as the base. Cursor, Claude Code,
Obsidian, vLLM, and custom automation tools wired into the desktop workflow. No Mac -
need proper support for engineering tools (SolidWorks, ANSYS) and full control over
the system.

The goal is an environment where AI assistance is ambient, not app-switched. Terminal,
editor, browser, notes - all connected through custom integrations and scripts.

The intelligence layer absorbs what was originally a separate JARVIS project - features
that only make sense as part of the OS, not standalone:

- **Meeting listener**: live transcription, speaker diarization, automatic action items
  via LiveKit + VAD
- **Voice pipeline**: Silero VAD + Faster-Whisper + Kokoro TTS - speak to the system,
  get spoken responses
- **Temporal knowledge graph**: Qdrant + Graphiti - everything ingested into a graph
  where facts expire and update automatically, no stale context
- **Unified messaging bus**: Telegram, WhatsApp, Discord, email all routed through a
  single interface
- **Document intelligence**: NuMarkdown-8B for complex OCR and document understanding
- **Autonomous execution**: microsandbox, Windmill for multi-step workflows

**Status:** Designed, ready to deploy. Blocked on hardware (need a proper laptop).

**Stack:** Arch Linux, Hyprland, Cursor, Claude Code, Obsidian, vLLM, LangGraph,
Qdrant, Graphiti, LiveKit, Faster-Whisper, Kokoro TTS, Windmill, Runpod

If you want to fund a laptop for this, I'll open-source the entire config.


Custom AI-integrated Linux workstation. Arch + Hyprland with AI tools wired into the desktop workflow.

Agentic OS


50+ years of satellite data - Landsat (1972-present), Sentinel-1/2, MODIS, VIIRS,
GRACE-FO gravimetry, ERA5, MERRA-2, and 150+ years of Uzhydromet records from 87
synoptic stations - all analyzed with methods from 2015 or earlier. Random Forest,
SVM, Markov-Cellular Automata, traditional indices. No foundation models. No vision
transformers. No deep learning. One DL paper exists (2023, Korean group, radiometric
normalization only).

Fine-tuning Prithvi-EO-2.0 (NASA/IBM's 600M param geospatial foundation model) and
Chronos-2 for six application modules: dust storm early warning, irrigation optimization
(deep RL on AquaCrop), water allocation optimization, crop yield prediction, salinity
risk mapping, crop disease detection.

Targeting 200+ dust alert users and 10+ farmer pilots in Karakalpakstan by mid-2026.
2-3 papers planned.

**Status:** In development (data pipeline and model fine-tuning phase)

**Stack:** Google Earth Engine, Prithvi-EO-2.0, Chronos-2, PyTorch, AquaCrop-OSPy,
Gymnasium, SAC, FastAPI, Telegram Bot API

**Data:** Sentinel-1/2, Landsat, MODIS, VIIRS, GRACE-FO, ERA5, MERRA-2, CA-discharge,
Uzhydromet archives

Needs GPU compute for foundation model fine-tuning. Looking for sponsorship.


First modern ML pipeline for the Aral Sea disaster zone.

Aral Basin Environmental Platform


Telegram-based knowledge sharing platform. Book cataloging, resource sharing between
communities, reading list management, integration with external book databases.

**Stack:** Python, Telegram Bot API


Telegram bot for the Agora Library knowledge sharing platform.

Agora Library Bot


## Overview

This website serves as a digital notebook, inspired by Leonardo da Vinci's codices. It showcases research proposals, projects, reading lists, and writing.

## Technical Stack

- Next.js 13 with TypeScript
- Chakra UI for styling
- MDX for content management
- Vercel for deployment


Personal website built with Next.js and Da Vinci aesthetics

Projects