I build production tooling where machine learning meets smart contracts.
- AI/ML engineering — zero-shot models at the edge, offline-first ML pipelines, agent frameworks, Hugging Face ecosystems (datasets, transformers, Hub CLI).
- Blockchain engineering — Solidity + Foundry smart contracts, on-chain provenance, audit-ready security tooling, Sepolia deployments, ethers.js/web3.py.
- Full-stack glue — Python, TypeScript, React — shipping tools that humans actually run.
Everything below is open source, tested, and under active development.
| Repo | What it is | Stack |
|---|---|---|
| model-ledger | On-chain provenance for ML models — hash a Hugging Face manifest, register it, let anyone verify it. No trusted third party. | Solidity · Foundry · Python · ethers.js |
| slither-chat | Smart-contract audit copilot: explains every Slither finding in plain English (offline KB, on-device HF zero-shot model, or any LLM), with patches, SVG reports, and a benchmark vs. real audited contracts (recall 0.97). | Python · Slither · Hugging Face · transformers |
| trait-gpt | Rarity score + AI-written descriptions for any NFT collection — trait statistics, CLIP zero-shot attribute tags, BLIP captions, Streamlit gallery, all CPU, zero keys. | Python · transformers · Streamlit |
| hf-hub-lint | Lint Hugging Face Hub repos: model-card completeness, license declaration, metadata hygiene, config sanity — weighted 0–100 score with actionable fixes, live Hub or offline fixtures, zero keys, zero runtime deps. | Python · pure stdlib |
| inject-scout | On-device prompt injection & jailbreak scanner for LLM apps — 22 deterministic rules (leaks, role-escape, encoded payloads, indirect injection), 0–100 score, --strict CI gate, 29-attack golden corpus, no API keys. |
Python · pure stdlib |
| pocket-eval | The LLM eval harness that fits in your pocket — CPU-only, keyless, offline: perplexity + a 12-item QA mini-benchmark over a bundled DeFi/ML corpus, deterministic n-gram & seeded random baselines, optional tiny HF models (tiny-gpt2, Qwen2.5-0.5B). Real-run results table, 23 offline golden tests. |
Python · pure stdlib · transformers (optional) |
| embed-playground | Compare lexical (TF-IDF) vs dense (sentence-transformers) vs hybrid (RRF) search on a bundled 36-doc DeFi/Web3 corpus — 16-query recall@5/MRR benchmark (lexical 75% vs dense 94%, real runs), Streamlit playground, zero keys. | Python · pure stdlib · sentence-transformers · Streamlit |
| card-forge | Forge complete Hugging Face model cards from config facts — architecture → task inference, honest parameter estimates, license detection, A–D documentation score, config diffing between revisions (live Hub fetch optional). 26 offline golden tests. | Python · pure stdlib |
| chain-chat | Ask on-chain history in plain English ("which token moved the most yesterday?") — schema-aware LLM → SQL over a bundled DuckDB/parquet snapshot. Zero keys, offline demo, 65 tests. | Python · DuckDB · Streamlit · NL→SQL |
| agent-wallet | AI agent for safe testnet transactions (built on my own agent-lab): inspects a wallet, checks allowances, dry-runs the tx via eth_call, and explains risk in plain English before anything is signed — Sepolia only, mainnet default-off, 67 offline tests. | Python · agent-lab · web3.py · click |
| chain-scout | AI wallet risk scanner: paste any Ethereum address → plain-English risk report — rug-pulled tokens, dangerous approvals, holder concentration, estimated PnL — transparent weighted score, keyless demo, testnet-first. | Python · Streamlit · web3.py · CoinGecko |
| position-guard | DeFi health monitor: watches Aave v3 / Compound v3 positions via The Graph subgraphs + CoinGecko, computes health factors, and pushes LLM-written plain-English alerts ("your ETH collateral is at 1.12 HF…") — Telegram optional, keyless web-preview demo by default. | Python · The Graph · httpx · SQLite |
| agent-lab | Zero-dependency AI agent framework in pure Python — tool use, persistent memory, retries, guards. | Python |
| whale-watch-tg | Telegram whale tracker — subscribe to any wallet, instant alerts with LLM summaries ("3,000 ETH to Binance — likely sell pressure"), zero-key demo. | Python · python-telegram-bot · web3.py |
| KARYX | Military-grade edge AI model optimization & deployment suite. | Python · ONNX |
| hive | Multi-agent harness for production AI. | Python |
| mobile-profile-shell | A phone-OS style portfolio — React + Vite + Tailwind + shadcn. | TypeScript · React · Tailwind |
Newest ships: model-ledger (Solidity registry + Foundry tests + Python CLI + web viewer, Sepolia-ready), slither-chat (56 tests passing, benchmarked against a 1,748-contract HF corpus), agent-wallet (AI agent for safe testnet transactions — agent-lab core, dry-run + plain-English risk notes, mainnet default-off, 67 offline tests), position-guard (DeFi health monitor — 52 tests, golden health-factor math, keyless demo), hf-hub-lint (HF Hub repo linter), inject-scout (prompt-injection scanner — 20 tests, 27/29 attack-corpus recall), pocket-eval (keyless CPU LLM eval harness — real-run table: n-gram 83% vs tiny-gpt2 17% on the bundled mini-benchmark), embed-playground (lexical vs dense search — recall@5 75% vs 94% on the bundled DeFi/Web3 benchmark) and card-forge (model-card generator + config diff — 26 tests, A–D checklist score) — all feature-complete, documented, and contribution-ready.
- Hugging Face zero-shot classification in production pipelines — DeBERTa-v3 vulnerable-class tagger running on-device (CPU, no API key) inside slither-chat.
- Benchmarking ML pipelines — precision/recall/F1 scoring against real-world ground truth from the HF Hub (recall 0.968 on Slither-audited contracts).
- Agent frameworks — tool use, function calling, memory, retries (agent-lab, hive).
- Edge model optimization (KARYX).
- Smart contracts with Foundry — zero-dependency Solidity 0.8.26, 18/18 tests incl. fuzzing and golden cross-language hashes (model-ledger).
- AI + Web3 intersection — content-addressed model provenance on-chain, AI-assisted security auditing (slither-chat).
- Contract security — Slither-based static analysis, patch-diff generation, severity triage.
- Tooling — web3.py, ethereumjs/ethers.js, Sepolia deploy + Etherscan verify scripts, local anvil chains.
- Small shippable deltas — I land changes in increments that are easy to test and review, and every repo carries a roadmap of what's next.
- Tests before claims — every numeric claim in my READMEs is reproduced by a command in the repo (benchmarks, golden fixtures, live demo transcripts).
- Honest engineering — no empty commits, no exaggerated badges.
- 💼 Open to Software Engineering / AI-ML / Web3 roles — fully remote or EU relocation.
- 📖 Read the code: github.com/pxlcrtiv
- ✉️ Best way to reach me: open an issue or GitHub discussion on any repo — or star one of the projects above, it genuinely helps.
Liked what you see? ⭐ Star a repo. It's the best feedback a solo builder gets.