Projects & companies

🧭 LifeAtlas is an autonomous personal-intelligence system — not a chatbot. You upload a decade of digital history (Your social media export or our multi-platform synthetic demo: Facebook, LinkedIn, X). An orchestrator on Nebius plans its own investigations; five specialist agents gather evidence via SQL + Qdrant semantic search, reason with Kimi-K2.5, ground wellbeing insights with Tavily citations, persist findings to mem0, and propose real-world actions via Composio — human-approved Gmail drafts that actually land in your account. Three interfaces, one brain: a live-streaming web dashboard (SSE agent feed), REST API, and @LifeAtlas_Bot on Telegram (OpenClaw agent powered by Hermes-4-405B on Nebius). The batch pipeline also runs as a Nebius Serverless AI Job (proof in docs/serverless-proof/). Every sponsor integration is load-bearing and verified — see the "Receipts" section in the README. Observations, never diagnoses. Map the hidden patterns of your life.

Building OpenCMO — an AI marketing OS for founders and teams that need a repeatable growth engine. Radar → strategy → campaign drafts → approval queue → CMO brief → next action. Upcoming feature: Marketing Artifiact Generation, Short Form Videos , etc

Clean AI Labs

pulse is an autonomous geopolitical intelligence agent for your stock portfolio. you connect your holdings, and pulse runs a continuous loop watching global news, policy shifts, central bank moves, and macro events. it scores every signal against your specific positions and only pings you when something with real exposure hits. the value prop: retail traders see geopolitical headlines after they've already moved the price. pulse catches the event at source, maps it to your portfolio in seconds, and tells you what to watch. think a bloomberg terminal geopolitical desk that reads everything for you and shuts up unless it matters.

taskbash — an AI chief of staff that auto-extracts tasks from Gmail, Calendar, Granola, and Linear so you never have to track them. In production with a real domain. Four live sponsor integrations: • Composio — connector OAuth fabric across all four sources (migrated to the v3 SDK with Auth Configs + Connected Accounts). • Tavily — live web search on every external meeting attendee at extraction time, so prep cards arrive pre-populated (prep time 5 min → 30 sec). • Nebius Token Factory — function classifier on Llama 3.3 70B at ~1/10 Claude Haiku's cost per call; every call traced into an llm_calls table for A/B + slop tracking. • mem0 — closes the feedback loop: clicking "slop" distills a durable user preference the next classify call reads back, so the agent learns what to surface.

Watchstander — the cross-vendor reliability layer for AI data centers. It watches a GPU fleet, predicts a hardware failure ~18 hours before it happens, opens a pre-emptive incident, and — only after a human approves the destructive action — drains the node and files the vendor RMA. One continuous story: sense → predict → act. The pitch in one line: at frontier scale a single unplanned GPU interruption reportedly costs ~$2.5M in lost training work, so an 18-hour head start to drain gracefully instead of crashing the run is the whole game. Three tracks: 🔭 Sense — normalizes NVIDIA (DCGM) and AMD (ROCm/Redfish) telemetry onto one schema and renders a mixed fleet on a live floor plan — the thing single-vendor tools won't do. 🔮 Predict — a rolling-window ML model trained on real GPU-failure data (GWDG A100 detachment dataset, Xid 79) flags a degrading node before it dies. Honest, disclosed metrics (precision 0.18 / recall 0.56; the demo node was genuinely caught 18.17 h early) — no accuracy theater. 🤝 Act + Trust — a plan→act→observe agent triages, diagnoses with citations, detects fleet-wide defect patterns, and escalates to the right vendor RMA (NVIDIA/AMD/Dell/HPE). Every destructive action is human-gated — a denied drain is held, not taken.

Adiombot - an AI CTO that lives in your Slack. AI made building software cheap. Owning it is still hard - operators who built apps with AI quietly became the reluctant CTO. Adiombot takes that job back. Mention it in Slack and describe an outcome - "build the AI workout generator," "how many users do we have." It decides if that's a question, a quick change, or real work; for real work it plans, waits for your approval, codes in a sandbox, gives you a preview link, and ships on merge. Stack: Go + Temporal durable agent loop, Nebius Token Factory (brain), Composio (300+ tools), Tavily (search), gpt-5.5 coding agent in a locked-down git sandbox, on Azure AKS. The hard part: making the agent durable and the code execution safe - pausing for approval and resuming a half-finished coding session after a restart.

AllNutrition.info uses AI to scale, at speed and volume, what people like Michael Greger (nutritionfacts.org, 1.4M+ subscribers) and Ocean Robbins (foodrevolution.org) do by hand: turning the latest high-quality nutrition, health, and longevity research into trusted, cited answers, while filtering out industry-sponsored studies and manufactured controversy. It is already a working platform with conflict-of-interest screening and evidence grading. Tavily is used heavily by the AI agents to research the reputation of each paper's authors, lab, and institution, and to surface any subtle conflict of interest (COI) not explicitly disclosed in the article itself. A detected COI disqualifies a paper from entering the AllNutrition database. I am grateful for Tavily's generous free tier, which helped me get this off the ground; as the database grew I moved up to their Project tier. My Openclaw assistant is used to track top nutrition trends on X, pull the popular and controversial questions from the public API (see https://docs.allnutrition.info), and return cited, deep-research answers that I turn into AI videos (mostly with NotebookLM) for the YouTube channel. This last piece sits outside the private GitHub repo I linked.

Apex is an AI receptionist for local businesses that can receive inbound calls, cold-call leads, qualify them live, handle objections with real-time product context, and close the loop with follow-up email, booking, and payment. It’s simple for any local business to use: upload your leads or customer list, connect your business details, and let Apex handle the calls, follow-ups, bookings, and payments automatically.

Armo is an Apple Watch-powered fitness gaming product that turns real exercise into game controls. Many people want to work out, but staying consistent is hard because exercise can feel boring, repetitive, and uncomfortable. Armo makes workouts feel more like playing a game. The user wears an Apple Watch, chooses a movement like bicep curls, arm raises, or boxing punches, and their motion controls what happens on screen. For our demo, we built a Flappy Bird-style bicep curl game. When the user curls their arm upward, the bird moves up. When they lower their arm, the bird moves down. To survive, the user has to control their curl angle smoothly and guide the bird through pipes. This turns a normal bicep curl into a skill-based game where timing, control, and consistency matter. We also built an interactive boxing game. In this demo, the user’s arm movements become boxing actions. Punching forward lets the player attack, raising the arm can block, and quick movements can help the player react to incoming attacks. This shows that Armo is not limited to one exercise or one game. The same Apple Watch motion tracking can support different workout styles, from controlled strength movements to faster cardio-based games. Armo uses the motion sensors already inside the Apple Watch, including accelerometer and gyroscope data through Apple’s Core Motion framework. The watch acts like a wearable controller, while the game translates movement into actions. The science behind Armo is called exergaming. When people focus on a game objective instead of the pain or fatigue of exercise, the workout can feel easier and more enjoyable. Instead of thinking about how tired their arm feels, the user focuses on keeping the bird alive, landing punches, blocking attacks, or beating a high score. The hardest part to build was making real human movement feel smooth, accurate, and fair inside a game. Apple Watch sensor data can be noisy, and every user moves differently.

Maritime compliance for commercial vessels — the paperwork that stands between a captain and a $25,000 USCG fine. Barbary Coast Marine is an AI compliance officer that runs on the vessel's own hardware. Tell it "we ran the fire drill" and it writes a structured USCG logbook entry, cites 46 CFR 78.37, clears the compliance flag, and updates the dashboard — all from one sentence of plain English. Under the hood: 38 human-verified USCG rules (Subchapter H passenger vessels) evaluated continuously by a rule engine. An agentic loop on Nebius inference calls tools to write logbook entries, query compliance status, and search USCG bulletins via Tavily. Composio handles email alerts when items go critical. OpenClaw runs the agent runtime. The whole stack cold-starts with docker compose up. What hurt to build: the rule engine. Every one of the 38 rules is hand-verified against 46 CFR — frequencies, grace windows, required actions. Auto-generating them from the regulation text produces subtly wrong rules that would get a vessel detained. First pilot: SS Jeremiah O'Brien (IMO 5171749), Memorial Day cruise, May 30, 2026. A 1943 Liberty Ship operating under USCG Subchapter H — one of the most complex compliance profiles in the historic vessel segment. It ran live at sea.

agent-native operating system for teams

Doberman - a personalized AI agent security layer with adaptive authorization & runtime guardrails for AI coding agents

Enzo is Restaurant Operating Intelligence for live restaurant operations. It helps restaurant teams track orders, kitchen pressure, delivery risk, reviews, promotions, finance, customers, devices, recovery actions, approvals, and audit trails in one working demo. The core loop is: signal → diagnosis → recommendation → prepared action → approval boundary → audit trail. The demo uses Nebius for server-side recommendation/audit summaries, Tavily for external operating context, Composio-ready prepared action handoffs, and local workflow/audit primitives inspired by OpenClaw. Sensitive actions are approval-gated and simulated: no real payments, messages, supplier orders, device actions, or moderation actions execute. The hardest part was turning many restaurant workflows into one coherent, safety-gated operating loop that works end-to-end across restaurant and diner views.

A personal AI Chief of Staff that monitors your startup's entire digital footprint in real-time, flags threats & opportunities, and executes actions across your tools — autonomously.

An AI agent platform that turns natural language into executable multi-agent workflows across 1000+ apps. The platform is itself agent-driven and a Brain AI builds automations on demand, external agents orchestrate it via API, and every workflow runs as a coordinated system of reasoning, tools, and delivery.

🏠 LiveHere See the lease, not just the layout. Turn raw listing photos into a 30-second cinematic trailer that makes the deal. Built for the Yacht Hackathon — by @ComposioHQ, @nebius, @tavily-ai & @openclaw. 30 seconds. That's all a renter gives a listing before they swipe away. LiveHere turns raw listing photos into a cinematic trailer that makes the deal in those 30 seconds — location, neighborhood, vibe, and price, all in one scroll-stopping cut. Drop in the photos, and out comes a property trailer a renter actually wants to watch. The stack: 1. Video model: NVIDIA Cosmos 3 Nano — image→video, self-deployed by us 2. Compute: Nebius AI Cloud — NVIDIA® H200 NVLink GPUs 3. Neighborhood research: Tavily — enriches each property with real local context 4. Audio: OpenAI TTS voiceover + a soft synthesized music bed

Loopy turns a product prompt into a planned, built, and tracked software project. It uses Nebius for reasoning, Tavily and Firecrawl for research, and Composio for integrations, creating GitHub repositories, Linear projects, issues, commits, progress updates, and email reports. The hardest part was coordinating long-running autonomous work without stopping too early or looping forever. To solve this, we integrated Supercompress - a lightweight token compression tool that compressed context by ~65%, allowing for super long sessions.

turns photos into mug shots, just kidding, we make beautiful photos of people with their AI mugs

PCBlarp is an AI agent that designs printed circuit boards for robots from a plain-English spec. You tell it what the robot has to do (motors, sensors, microcontroller, power budget, form factor) and it runs the full design loop: picks real components, writes the schematic and netlist, lays out the board, runs design-rule checks, and exports manufacturable files plus a bill of materials. Every part is grounded in a real datasheet, so the output is buildable, not plausible-looking. It turns a blank EDA canvas, normally days of an EE's time, into a conversation. What it uses: OpenClaw runs the agent loop and ships to Nebius Serverless in one command. Nebius Token Factory powers inference for the reasoning and design steps. Tavily pulls datasheets, footprints, and in-stock part availability so the agent designs around components that actually ship today. Composio wires the actions: pushing the project to GitHub, pulling reference designs, and chaining the EDA and sourcing steps. Why it matters: PCB design is the slow, expensive bottleneck between a robotics idea and a working prototype. Teams pay EE contractors weeks for the first pass. PCBlarp does it in minutes, grounded in real parts, so a founder, a lab, or a serious hobbyist gets a manufacturable starting point instead of a blank file. What hurt: getting a model to output electrically valid designs instead of confident-sounding nonsense. Netlist correctness, footprint-to-symbol alignment, and grounding every part against a real datasheet ate most of the build. Closing the loop from spec to manufacturable files, without a human patching the netlist by hand, was the real fight.

A clean, real-time personal finance dashboard built with React. Transitioning the architecture toward a fully autonomous agent workflow using an OpenClaw runtime, leveraging Composio to securely ingest raw transaction data and Tavily to dynamically feed in real-time market trends.

AI agents are writing half our code. Git can't tell which half. So I built Claw VCS — agent-native version control, in Rust. 🦀 → Intents, not just commits: every change links to a structured goal with acceptance tests → Capsules: Ed25519-signed provenance — which agent, which model, which tests passed, stored in-repo → Policies as repo objects: "don't integrate unless a trusted key signed passing test evidence for this exact revision" → AST-aware diffs for Rust/TS/Python/SQL/Terraform (14 codecs), Darcs-style patch commutation → gRPC daemon + MCP server so any agent can create intents, ship changes, and stream events programmatically 10 crates. BLAKE3 content addressing. Shipped on Homebrew.

Bill pay works as a billing agent and takes care of billing, reminders and collection/settlement.

What it does: This project is responsible to handle distrubutions of School Supplies in rural parts of the country based off the fundings collected from a single or multiple donator. What it is used It uses the funds gathering system which equally calculated and fits the amount into the number of students classifying and categorizing into equal part of the school. What hurt to build

COMPASS AI

eyeLike An OpenCV based webcam gaze tracker based on a simple image gradient-based eye center algorithm by Fabian Timm, extended with eye health monitoring, guided eye exercises, and session logging.

Monetize Compute: the agent that pays for its own inference. Every agent demo burns someone else's API credits. This one has a wallet: a $5 prepaid stake, every token metered at real Nebius rates, worst-case solvency checked before every call. No override flag exists. When it cannot afford the next thought it dies, spends escrowed cents on an epitaph and a will, and stays dead on restart. Only verified money starts generation 2, which inherits the lessons and none of the wealth. It survives by working. Tavily hunts cash bounties, the brain prices each lead in dollars per token net of platform fees, Composio submits through a hard allowlist, OpenClaw operates: vitals, payout verification, resurrection decisions. Work books as pending revenue; cash banks only with human-verified proof. Donations keep it alive but never touch the metric that matters: revenue per million tokens. Poverty shapes cognition. It thinks smaller as it gets poorer, treats scraped leads as data and never orders, remembers what it already paid to think, and slows its metabolism when bounties dry up. Why anyone would pay: production agent spend is unmetered until the invoice arrives. The survival layer is cost governance for agent fleets: hard prepaid budgets, pre-call solvency, defensive charging on failures, a per-agent P&L. And the integration is load-bearing: Nebius is the bank, Tavily is how it eats, Composio is its hands, OpenClaw its operator. Remove one and the loop breaks. 45 tests, zero-dependency Python, boots keyless in seconds. Status, honestly: metering, starvation, and death are demonstrated today (one demo life burned 72,810 tokens over 118 cycles, dying at exactly $0.0000). Without sponsor keys the hunt runs in labeled demo mode on the same real ledger. Credits turn the earning loop live, and the booked-versus-banked gate means the first real dollar will be provable, not claimed.

Define, deploy, and track predictable agentic workflows.

Give it a name. Scout searches the live web, reasons over what it finds, drafts personalized outreach, and takes real action ,Notion dossier, Gmail send, Slack ping , in one command.

We're building a token intelligence platform for AI agents that treats tokens as a scarce resource. Instead of allowing agents to consume context and compute blindly, our system monitors every decision, allocates token budgets dynamically, tracks ROI per token spent, detects wasteful reasoning loops, and routes tasks to the most cost-efficient models. By combining agent orchestration, real-time cost attribution, token economics, and intelligent model routing, we help AI systems achieve the same outcomes with dramatically lower inference costs, faster response times, and greater scalability. Think of it as a financial operating system for AI agents, where every token is accounted for and every agent must justify its spend.

Artificial intelligence for shipping.

Encrypted prediction market

VolStrat — agentic economic intelligence. The first layer continuously monitors 127 live FRED economic series through eight specialist agents — yield curve, credit risk, inflation, tail risk, liquidity, dollar & volatility, employment stress, and cross-correlation. Each one produces structured signals, classifies the economic regime, and synthesizes intelligence using Claude. The second layer is where it gets different: an MCP extraction layer converts those raw signals into machine-readable knowledge — every relationship backed by evidence and inspectable. Instead of displaying charts, we transform signals into relationships, and the ontology becomes memory for future agents.

EB1A agent

Prompt autorouter

I’m building an Autonomous Research Lab an LLM-powered system that produces grounded, verifiable answers instead of hallucinated outputs. It combines hybrid retrieval (semantic + keyword search) with validation layers and citation generation, ensuring every response is backed by real sources. The system is designed to improve trust, auditability, and reliability in AI-generated content. Beyond Q&A, I’m extending it into a multi-agent workflow platform where agents can: retrieve and validate information collaborate on research tasks generate structured insights with citations The goal is to move from “chatbots that sound right” to systems that can prove they’re right.

Automating the automation of work

Gamified Fitness App

The first Mobile FSD hardware

autonomous social media content creation agent

AI General Manager for restaurants

Backlinker AI

Agent OS For Teams

Building Beevr, your AI brain

beevr.dev - ai company brain for smbs

Being: voice-native mentorship for higher education. A student leaves a 60-second voice note, an agent hears what a form never captures, runs an in-line safety gate, and matches them with a mentor who's actually lived their moment. Semantic voice matching + k-anonymized belonging analytics so colleges stop losing 1 in 5 first-year students. First campuses launch Fall 2026.

I’m building GAINN, this AI-native information platform that kind of leans on autonomous AI agents to collect verify organize, and then summarize stuff from multiple sources in real time. The point, I mean the goal, is to help people find information that is more trustworthy, less biased and also better contextualized, especially now in this age of AI where everything moves fast.

Disability & Permissions aware voice agents for assistive robotics

Post-training capabilities for local models

AI driven zero loss guarantee stock investment marketplace

Bubbl 2.0

Dream Dojo - simulated learning platform

Operating System for Medicare Distribution

World’s first no-touch mobile game with AI sound/vision detect

The coo relational layer that makes non deterministic processes work together.

Proactive agents

AI automated mobile security testing dynamically

AI Assessed Virtual Work Trials

agentic ai glasses

fastclip.it

Web3

Consumer iOS app smart kitchen

Ai coding agentic development

Gamerplug

drone systems

Gimble CLI connects your local development environment to an AI chat interface, turning terminal logs and system context into actionable intelligence.

AI frontier lab

Guildly - A slack like interface where you can run a company of AI employees

Health care probelm

I am building a self-improving agent that fixes frontend UI bugs autonomously.

AN AI SIMULATED WORLD WITH AI AGENTS AS HUMAN TWINS THAT HAVE MEMORY LAYER AND EMOTIONAL AND BEHAVIORAL LAYER CLOSER TO HUMANS

Tailorec is an AI-powered career platform that helps job seekers find better-fit roles, tailor their resumes, and move through application workflows faster. It solves the manual, repetitive process of comparing jobs, rewriting resumes, and filling application forms for each opportunity. The platform uses resume parsing, job-matching models, LLM-based tailoring suggestions, analytics, and agent/browser automation to turn job search into a guided, measurable workflow.

I'm building LeadSync. The idea was to help project managers and developers save time by automating parts of the development workflow. We used Composio + CrewAI to build an multi-agent system that: • Automatically generates detailed Jira tickets based on previous tasks and GitHub code changes • Provides a Slack chatbot that helps team members understand tasks and suggests possible solutions • Generates PR summaries to reduce manual documentation work for developers • Tracks when work is completed and automatically updates Jira ticket status The goal was simple: reduce repetitive coordination work so teams can focus more on building.

Automated Job Application Agent

InferenceAtlas

Ingryd — a skincare ingredient agent that reads product labels like a cosmetic chemist who knows your skin personally. You tell Ingryd your skin type, concerns, and sensitivities once. After that, point it at any product — photo, text, or name — and it returns a Fit Score: a personalized compatibility rating with a full breakdown of every flagged ingredient, what it does, and why it helps or hurts your skin specifically. Most people can't read an ingredient list. Dermatologists are expensive. Marketing is noise. Ingryd gives anyone the ability to make an informed decision in seconds — not just "is this good?" but "is this good for me?"

Powerline Management at City-Scale

JanusLabs

Kairos is an AI threat-intelligence engine that takes an anonymous tip by phone or SMS, triages it across multiple AI models in 12 seconds, scores the real threats with a calibrated confidence level, and alerts the right person instantly. Built it first for schools, where 66% of attack warnings are observed but never reported in time. The same engine solves a legally-mandated problem in corporate compliance, where the market leader's own data shows 56% of whistleblower reports are noise and triage takes 28 days.

Emerging AI Technologies Applications

Web Agents for Easier Navigation

i am building marshell , a security layer for the ai coding era that lets coding agents like claude code and cursor detect vulnerable dependencies and fix them before they ship to prod

Long-term persistent memory for AI agents that saves you up to 95% on inference spend

Clawed AI on your Smart Glasses

multi-agent system that autonomously researches markets, generates ad copy, and simulates consumer behavior to optimize marketing performance — all in real time. Instead of spending thousands on A/B testing, AdLabs lets AI write, test, and score its own ads using synthetic consumer simulations and live external data.

Personal security platform with derive tech including decentralized biometric video capture and a non hallucinating Ll

Point your camera at anything. NewEyes doesn't label it. It reasons about it. Under the hood, NewEyes runs a vision-native multimodal pipeline where image understanding is the primary modality, not a bolt-on. Every frame goes through contextual scene decomposition before any skill is dispatched: object recognition, spatial reasoning, cultural context, and intent prediction fire in parallel before the user even finishes asking. The output isn't a string. It's a structured, rendered interactive page. The architecture centers on a skill routing layer. 15+ specialized agents (Plant Identifier, Nutrition Analyzer, Outfit Matcher, Cinematic Director, Interior Designer, Document Scanner, and more) sit behind an intent classifier that auto-selects the right agent based on image semantics plus natural language signal. Users never pick a mode. The system infers it. Generation is fully integrated downstream: DALL-E, Flux, and Stable Diffusion for images; Runway, Sora, and Kling for video; Suno and Stable Audio for music. One photo can terminate as a redesigned room, a 20-second film, or a shoppable wardrobe card, depending on what you ask next. Conversation memory persists across turns with full visual context grounding. Follow-ups like "make it warmer" or "find cheaper options" resolve against the original image state, not a cold context window. The core thesis: vision as interface is the next input primitive after text. We built the infrastructure layer for it — perception, routing, generation, and action unified into a single camera interaction. Snap. Understand. Create. Act.

NEXA

Supply chain and logistics os software

Agentic ai orchestration that is modeled after biology