Skip to content
Sam
autodidact · ai-native · building since feb 2026

I build things that run on their own.

I'm Sam — Dr. Inker online. An autodidact, AI-native solo developer. In about five months I've shipped production trading infrastructure, real-time multiplayer games, and AI products, using Claude Code, subagents, and persistent memory as my toolchain. Crypto and trading bots were my way into AI — I'm now turning that same production discipline toward everyday AI solutions for real-world businesses. Everything here runs live on mainnet, not in a slide deck.

See my work Get in touch GitHub open to work
commits, Feb–Jul 2026
3,700+
projects, 4 domains
25
services live on one host
46
automated tests
5,000+
live surfaces shipped
10
persistent agent memory
1.7 GB
selected work

Ten I'd put my name on

The strongest of two dozen projects — live trading infrastructure, real-time games, an autonomous AI character, and shipped products. Strategy parameters and P&L stay private; these describe capability and architecture.

Berthalla Online — a cozy 2D top-down multiplayer town with shops, trees and a pond
2D · Phaser
Berthalla Online — the 3D world in-play with a skills panel, chat and live multiplayer
3D · Babylon
one authoritative server · three clients
2D Phaser→ 3D Babylon→ Unreal 5.8 · C++
clients send intents → Colyseus binary state patches → tween

Berthalla Online

Live · native client in dev

A real-multiplayer cozy MMO — now with a second, native client in Unreal 5.8.

The flagship in progress: a native Unreal Engine 5.8 client on the same live server as the web build. I reverse-engineered the @colyseus/schema v4 delta wire format, ported it to C++ (conformance-gated), built the outbound msgpack / ROOM_DATA path, and validated it against production — the client sends a move, the live server moves the avatar. Zones generate data-driven from headless Python, characters run a Tripo auto-rig → textured-biped pipeline, and the web build (2D Phaser → 3D Babylon.js) is live and playable today.

Unreal 5.8 · C++ Colyseus · msgpack reverse-engineered wire protocol headless Python tooling Babylon.js
Play the web build →
Farmhalla — Mayor Bert stands at the edge of his watered crop rows: ripe strawbeagles, collieflowers, taters and pawsnip sprouts on dark tilled soil, with the clock plaque reading Spring 9 and harvested crops in the toolbar
the farm · spring, day 9
Farmhalla — the Berthalla town plaza at 9:30 am: Deputy Bruno the St. Bernard stands by the stone fountain on his morning schedule, with market stalls, lamp posts and tulip beds around the square
town plaza · 9:30 am, deputy bruno on his rounds

Farmhalla

Live

A Stardew-style farming life-sim that runs entirely in the browser.

Mayor Bert comes home to fix up his grandpaw's overgrown farm — and the town's derelict dog shelter with it. Under the cozy surface is a full simulation: 28-day seasons on a live clock (a day ≈ 7 real minutes), farming, a fishing minigame, a hazard-telegraphed mine, NPC schedules with heart events, festivals, and an overnight shipping economy — all built from a frozen design spec that survived a 4-judge adversarial panel before a line of code was written. The 32px art is AI-generated but never trusted: every asset passes an art gate, maps are composed by headless tools, and the render is pinned by 17 deterministic visual gates, each proven red on the known-bad build first. Saves are local-first, with optional X sign-in cloud saves behind an archive-on-wipe guard.

vanilla JS · Canvas generated 32px pixel art deterministic render gates 371 tests 249 commits
Play it →
berthalla.io — the Berthalla ecosystem homepage: chrome wordmark over the hero film with live ecosystem status below
berthalla.io · ecosystem hub
Bert Studio — the free web generator page where visitors type a prompt and get an on-model Bert image
Bert Studio · public AI generator

berthalla.io

Live

The Bert ecosystem's home — a brand site with a public AI art studio inside it.

One page ties the universe together: live ecosystem status refreshed on a cron, and a hero film hand-tuned for mobile autoplay (the hard-won iOS rule: never auto-unmute on a touch event — sound is strictly opt-in). Behind THE STUDIO section sits Bert Studio, a free public AI image generator — which in practice is an abuse-resistance problem: a per-visitor daily cap plus an atomic daily money cap booked before the safety classifier ever runs (no unpaid-for generations, even under race), a fail-closed, injection-guarded LLM safety gate, and per-IP rate limits that trust Cloudflare's real client IP. Every image stays on-model via a frozen character sheet.

FastAPI · Python Grok image gen LLM safety classifier nginx · Cloudflare 73 tests
berthalla.io → Try Bert Studio →
BERT Widget on an Android home screen, showing BERT price, market cap, volume, liquidity and optional holdings value

BERT Widget

Android · live

A privacy-first native home-screen market widget, shipped as a signed Android app.

Two Jetpack Glance layouts put BERT's price, 24-hour move, market cap, volume and liquidity on the home screen, with an optional locally stored holdings value. The companion app adds a native market desk and setup flow. A dependency-free Node quote service normalizes upstream data, validates token identity by Solana mint rather than ticker, caches good responses and serves a clearly marked stale quote through short upstream outages. Holdings never leave the device; the app requests no wallet connection and disables Android cloud backup and device transfer. The signed v0.3.6 release supports Android 8.0 and newer, with source, permissions and release history public before install.

Kotlin · Jetpack Glance WorkManager Node.js quote API privacy by design iOS scaffold
See the release → Source →
promotion lifecycle
paper→ shadow→ canary→ live
16 collectors → Postgres feature store → anti-lookahead backtest trading

Polymarket 5-minute system

Live

A full-lifecycle quant system for 5-minute BTC/ETH prediction markets.

Every idea moves paper → shadow → canary → live, each promotion gated on out-of-sample evidence. A Postgres feature store is fed by ~16 always-on collectors (cross-venue liquidations, funding/OI, order-book snapshots); a custom backtest engine with point-in-time, anti-lookahead snapshots closes the live-vs-backtest gap; and a counterfactual ledger keeps logging the trades a demoted strategy would have made. Live execution is wrapped in regime-decay alarms, a bet-size tripwire, and an auto-kill watchdog.

Python PostgreSQL CLOB · Polygon systemd 1,714 tests 894 commits
private · trading flagship
A post from the verified @Berthalla_bot account on X, labelled 'Automated by @drinkerlabs', reading: 'woofmornin. the red line on the screen went down a little bit. i am currently looking at a very small crumb under the table to keep my eyes from wandering.'

Berthalla

Live on X

An autonomous "living" AI character that posts in-character.

A mood engine colours a constant voice; an LLM drafts every post, and a fail-closed adversarial editor-LLM firewall (pass / soften / veto, schema-locked) vets it — and re-screens its own rewrites — before anything publishes. Pure-logic vs thin-IO architecture; commits are blocked unless the suite is green. An anti-repetition system cut mode-collapse from 92% to 28%.

Python Claude + Gemini + Veo ~485 tests
On X → berthalla.io → 238 commits
polymarket-tax landing page — the same Polygon wallet computed side-by-side under three IRS frameworks (capital gains $4,287, gambling + OBBBA $11,490, ordinary income $5,914) with a $7,203 spread

polymarket-tax

Live · pre-launch

A Polymarket-native US tax tool: on-chain decode → IRS forms.

Decodes any Polygon wallet's full Polymarket history on-chain, runs FIFO cost-basis matching, and shows tax under three IRS frameworks side-by-side — then generates audit-ready Form 8949, Schedule D, a TurboTax CSV, and a methodology PDF. Paid in USDC via a self-hosted on-chain payment watcher with crash-recoverable fulfilment.

FastAPI Next.js 16 web3 ~250 tests
Open it → 91 commits
facelift before/after — the cluttered 1997 Heaven's Gate site (green Times New Roman on black with broken images) on the left, beside facelift's modern reskin on the right: a bold 'Leave This Planet' hero over a generated cosmic background with clean nav and a call-to-action

facelift

Live · public

Point it at any site; it ships an accessible reskin.

Reads the computed styles of every visible element to distill a site's design language, then has a trio of generator models each ship a WCAG-AA-repaired reskin. Every candidate must survive render-integrity and brand/imagery gates — logo detection, distortion audit, closed-loop object-fit auto-repair — then a judge-scored quality loop refines the winner before it ships. (The research behind this page came from facelift's own roadmaps.)

Node 22 Playwright multi-model + judge 300 tests
Try it → Case study →
preregistered gate
deflated Sharpe· purged CV· placebo· sign-stability
passes → paper-only quant research

Hyperliquid perps pipeline

Paper

A multi-asset perps research pipeline with a preregistered statistical gate.

~28 strategy modules across a 9-asset perp universe, each run through a gate that stacks deflated Sharpe, purged/embargoed cross-validation, multiple-testing correction, placebo controls, and sub-period sign-stability. Deployed as paper-only systemd timers with a live positioning collector; the live order path stays structurally disengaged. A clean, well-powered null is treated as a successful, capital-preserving outcome — research rigour over alpha-chasing.

Python NumPy / SciPy PostgreSQL FastMCP risk-gate 511 tests 322 commits
private · deep quant research
Sourcerer output — a generated fit-brief for Andrej @karpathy scored fit 0.98 and grounding 1.00, listing grounded claims that each carry a clickable source citation

Sourcerer

Live · public

An AI technical-sourcing agent that writes grounded, cited candidate briefs.

Give it a role; it finds an engineer on GitHub, researches them from public sources, and writes a fit-brief where every claim must cite real evidence — anything it can't ground is demoted to an "unverified" list in code. Each dependency (GitHub, web, LLM) sits behind a typing.Protocol + mock, so it's unit-tested with zero network; the fetch is robots-aware and SSRF-guarded. Built test-first via subagent-driven TDD, with a whole-branch review that caught a real stored-XSS before launch.

Python · async LiteLLM · GLM-5.1 fail-closed grounding 30 tests · CI
Try it →
Flappy Bert mid-run — Bert flaps toward a pipe gap and a shield power-up under a full moon, night city skyline below, score counter reading 1 Flappy Bert at an 8x streak — the streak banner flashes as Bert lines up a floating coin pickup above a pipe, night skyline behind
in-game · streaks & power-ups — tournament post on X ↗

Flappy Bert

Live

A flap-to-earn arcade you play right inside Telegram.

A from-scratch HTML5 Canvas game that runs as a Telegram Mini App. Scores are server-authoritative — the backend, not the client, is the source of truth, so leaderboards and tournament payouts can't be spoofed — and finished runs are shared as server-rendered score cards.

HTML5 Canvas · vanilla JS Telegram Mini App server-authoritative anti-cheat tournaments
Play on Telegram →
more work

The rest of the fleet

Trading bots, on-chain tooling, generative bots, and web — most running live, a few honestly parked.

Plus: a Polymarket weather-bot project (open-source base, extended with a leakage-free validation harness + a 7.2pp attribution fix), the earlier Telegram arcade era (Bert Tap Attack, Bert Runner NYC), a Solana→Printful merch bridge, and a design-phase cNFT sticker album.

origin · primary sources

Where it started

day zero

It started with a cartoon dog. In late January 2026 I was hand-making AI art of Bert — a memecoin community’s mascot — one image at a time in a web UI, and I had never written a line of code. On February 5 I installed VS Code and saved my first file: a Telegram bot meant to automate the art. My entire shell history from that week: where python · python --version · pip install — the last one with no arguments.

Four days later I installed Claude Code — the config backup still shows first start Feb 9, 14:08 UTC. Everything on this page was built AI-native from that moment on. And the same cartoon dog now fronts a live arcade, a 3D MMO with a native Unreal client, and an autonomous X presence.

Hand-made AI art from January 30, 2026 — Bert in a Kraken bucket hat beside a raccoon mascot in a cartoon football stadium The original Bert character reference sheet the first bot was built around

left: hand-made art from jan 30, 2026 — the week before the first line of code · right: the character reference the first bot was built around

bot.py — feb 5, 2026 · the first file

Verbatim, recovered from the home PC — tutorial-grade comments and all. The promotion gates, adversarial reviews, and 5,000 tests came later; this is what they grew from.

import logging
from telegram import Update
from telegram.ext import Application, CommandHandler, MessageHandler, filters, ContextTypes
from google import genai
from google.genai import types

# 1. SETUP: Fill in your secret keys here
TELEGRAM_TOKEN = "PASTE_YOUR_TELEGRAM_BOT_TOKEN_HERE"
GOOGLE_API_KEY = "PASTE_YOUR_GOOGLE_AI_STUDIO_KEY_HERE"

# 2. SETUP: List your character image filenames (make sure they are in the same folder!)
CHARACTER_IMAGES = ["char_front.jpg", "char_side.jpg"]

async def start(update: Update, context: ContextTypes.DEFAULT_TYPE):
    await update.message.reply_text("Hi! Send me a prompt, and I'll draw your character doing it.")
the proving ground · feb–mar 2026

Fourteen prototypes in the first month — dated folders straight off the home-PC desktop, most of them dead ends. That was the tuition: the survivors run in production today.

bot.pyfeb 05→ BananaBertBot, live today
bert-flex-botfeb 12parked
clickerfeb 13→ Bert Tap Attack
BertCSfeb 18parked
dr-inker-chart-scannerfeb 18the first trading itch
BertNYCfeb 19→ Bert Runner NYC
solana-wallet-trackerfeb 20reached the server; retired
flappy-bertfeb 23→ the arcade flagship, live
Bert_X_botfeb 25first deploy to the server
bert-shop-botfeb 26parked
bert · unitymar 01Mayor Bert town-sim; the idea became Berthalla
bert-survivormar 01parked
bert-tycoonmar 03parked
MMmar 05first market-maker sketch
lineage runs in production parked — tuition paid
how i work

Autodidact, AI-native by default

Claude Code, subagents, and persistent memory aren't an accessory to the work — they're the toolchain that lets one person ship like a team. But the interesting part isn't that AI writes code; it's the discipline wrapped around it so that moving fast doesn't mean moving carelessly.

I build features with subagent-driven TDD: a spec, a plan, then implementation reviewed by independent spec- and code-quality passes. Risky claims go to adversarial verification panels — multiple agents whose job is to refute, not agree. They've caught real, verdict-changing bugs: a pairing flaw that could have promoted the strictly-worse strategy, a lookahead leak in a feature builder, a prompt-injection bypass — each killed before it merged. That discipline is now open source as refute — a skeptic-panel CLI + GitHub Action on npm, with a published eval that's honest about where the panel wins (100% recall on hard flawed claims a single reviewer missed) and where it doesn't (it over-flags, so CI mode defaults to comment-only).

Durable context is the other half. Two dozen operating playbooks hold the manual for each system; a 1.7 GB memory store keeps hard-won lessons across sessions; and a custom MCP server lets an agent touch a live exchange only through a hash-verified preview-then-execute gate — the gate, not a string scrub, is the trust boundary.

Promotion discipline

paper → shadow → canary → live, each step gated on out-of-sample evidence. Strategies die on data, not on a story.

Preregistration

Hypothesis, endpoints, kill thresholds — plus the RNG seed and a code+config hash — fixed before any run, so fast iteration can't become optional-stopping.

Adversarial review

Multi-agent red-teams as a merge gate — two security audits hardened the live arcade; an 8-agent panel guards the LLM-trader's invariants.

Honest by construction

No fabricated numbers — live surfaces show a dash when a bot is offline; triggers are verified against a clean readback before any outward action.

Fleet ops & cost

46 services on one self-managed host across promotion tiers, fed by 16 collectors into PostgreSQL — with spend caps and RPC retry budgets throughout.

stack

What I work with

Languages

Python 3.12 · TypeScript / Node.js · Rust→WASM · C++ (Unreal 5.8) · SQL · vanilla JS / HTML5 Canvas

Quant & markets

Polymarket CLOB + UMA · Hyperliquid perps · order-book imbalance / CVD · funding & open-interest · liquidation modelling · maker-rewards simulation · FIFO cost-basis

Research statistics

shadow→canary→live gates · preregistration & power analysis · matched-pairs block bootstrap (BCa CI) · Holm correction · Wilson intervals · deflated Sharpe · purged/embargoed CV · Brier calibration

DeFi / on-chain (Solana)

Meteora DLMM · Raydium CLMM · Jito block-0 bundles · Streamflow vesting · pump.fun · Jupiter routing · Helius RPC · Anchor IDL decoding · Colyseus wire-format reverse-engineering

LLM / AI engineering

MCP server authoring (FastMCP) · agent risk-gating (preview→execute) · prompt-injection hardening · adversarial editor-LLM firewalls · multi-model orchestration (Claude + Gemini + Veo) · vision LLMs

Backend & full-stack

PostgreSQL 16 schema & indexing · SQLAlchemy / Alembic · FastAPI · asyncpg · Redis · live ingest pipelines · web3 / JSON-RPC decoding · React / Vite · Next.js · Privy + ES256 JWT · on-chain payment watchers

Game development

Colyseus authoritative server (600ms tick) · Babylon.js 3D · Phaser 2D · from-scratch Canvas engines · Web Audio chiptune synthesis · procedural pixel-art pipelines · Unreal Engine 5.8

AI-native workflow

Claude Code subagent-driven TDD · adversarial verification panels · spec→plan→build · operating playbooks · persistent memory · authored Claude Code skills

timeline

Jan → Jul 2026 · zero to production in ~5 months

  1. The spark. No code on the machine yet — just a Documents folder filling up with hand-made AI art of Bert, a memecoin community’s cartoon-dog mascot, generated one image at a time. Automating that felt worth learning to program for.

  2. The first file. VS Code goes onto the home PC and bot.py gets saved — a Telegram bot meant to draw Bert on demand. The whole shell history that week: where python · python --version · pip install, no arguments.

  3. Claude Code — day 4. First start 14:08 UTC, four days after the first file. Everything below was built AI-native from here.

  4. drinkerlabs is born. A folder with that name appears on the home-PC desktop, and the same afternoon the first bot moves to a rented Hetzner server. The server era starts here.

  5. The arcade era. Flappy Bert — born on the home PC on Feb 23 as prototype #8 — goes live with server-authoritative anti-cheat, tournaments, and server-rendered score cards; Bert Tap Attack and Bert Runner NYC follow.

  6. The Polymarket 5-minute quant system begins — Postgres feature store, a custom backtest harness, and the paper→shadow→canary→live promotion lifecycle.

  7. DeFi build-out. The Hyperliquid research pipeline, the first Solana market-maker, the atomic pump.fun launcher, and the Meteora DLMM engine (later run live with real capital).

  8. polymarket-tax SaaS started (on-chain decode → IRS Form 8949 / Schedule D); the cross-exchange hedged market-maker reaches observer phase.

  9. Berthalla goes live on X — an autonomous AI character posting in-character behind a fail-closed editor-LLM firewall.

  10. Berthalla Online ships its 3D flagship at drinkerlabs.info/mmo/, evolving 2D Phaser → 3D Babylon.js in place on the same authoritative server.

  11. A preregistered LLM swing-trader harness (BCa-bootstrap, Holm-corrected bake-off across four models) and a four-model generative-media bot with a news-driven anime-short pipeline.

  12. A native Unreal 5.8 client joins the live server. Reverse-engineered the Colyseus binary wire format, ported the decoder to C++, and validated outbound intents against production — one authoritative server now drives both a web build and a native client.

  13. Farmhalla ships overnight. A Stardew-style farming life-sim — design spec frozen July 2 after a 4-judge adversarial panel, v1.0 live at drinkerlabs.info/farm/ ten hours later, then hardened in the open to 371 tests and 17 deterministic render gates.

  14. BERT Widget ships on Android. Two native home-screen layouts, a companion market desk and a production quote API — signed, privacy-hardened and downloadable with the complete source and release history. In parallel, the Kraken↔Solana market-maker enters a read-only observer phase with conservative queue-aware paper fills and public telemetry.

An honest note. Live trading systems trade real capital on mainnet and can lose; several systems here are intentionally paper, shadow, or retired, and are labelled as such. Strategy parameters, wallet addresses, and P&L are deliberately withheld. Nothing here is financial advice.
hire me

What I can be hired to do

AI that survives production

Agents, automations and LLM features for real businesses — built with gated actions, evals, honest fallbacks and cost caps, not demo-ware. The same discipline that runs the systems above.

Trading & market infrastructure

Research harnesses, backtest → paper → canary promotion pipelines, exchange integrations, risk-gated execution. Engineering, not signals — I don’t sell performance promises.

Products, end to end

Full-stack apps, dashboards and games taken from spec to deployed, monitored and documented — solo, fast, with tests to show for it.

packages launch pricing
System audit
$1,900 fixed · 1 week
  • · Code + architecture review of a trading bot, AI agent, or production service
  • · Written findings report with prioritized fixes
  • · Walkthrough call at the end
start via email →
AI agent sprint
from $4,900 · 2–3 weeks
  • · A working agent or automation — deployed, tested, documented
  • · $4,900 single integration · $6,500 adds evals + deployment hardening
  • · Fixed scope, agreed in writing before work starts
start via email →
Fractional AI engineer
$2,950/mo
  • · One active request at a time, ~48h response
  • · Includes review of AI-generated code
  • · Pause or cancel anytime — no contracts
start via email →

Bigger builds — a full MVP from spec to deployed in 4–6 weeks — from $9,900. Email me with what you're building.

currently open to contract & full-time work · remote · fastest contact is email

Let's build something.

I'm open to interesting work — especially building practical AI solutions for real-world businesses, alongside trading, on-chain systems, AI agents, and games. The fastest way to reach me is email.

dr.inker@drinkerlabs.info GitHub X / Twitter