engineering lab
Experiments, with the results and the failures.
A journal that can grow, organised into numbered tracks. Entries are real repositories or documented builds; forks and adapted work are labelled as such.
track 01
AI engineering
Agents, guardrails and evaluation.
track 02
Backend architecture
APIs, data models and delivery pipelines.
- Acquisitions API: a hardened backend with a real delivery pipelineA solo Node and Express backend on Drizzle and Neon Postgres, with Arcjet rate-limiting and bot detection, Better-Auth and JWT authentication, and multi-arch Docker CI/CD.ShippedNode.jsExpressDrizzle
- AI analytics pipeline: ingestion serviceA FastAPI ingestion service over PostgreSQL and MongoDB with scripts for ETL and a Dockerised CI/CD setup: a starter structure for analytics pipelines.PrototypePythonFastAPIPostgreSQL
track 03
System design
Realtime, scale and failure modes.
- Video backplane: realtime across stateless replicasA live-streaming MVP built to exercise WebSockets at scale: Socket.IO with a Redis adapter fans chat out across stateless API replicas, with an RTMP to HLS video pipeline beside it.PrototypeFastifySocket.IORedis pub/sub
- PicknPay: an event-sourced order modelA solo full-stack e-commerce monorepo (Turbo and pnpm, React 19, Express) built around an 11-table event-sourced Postgres schema with order and address snapshots and OTP-verified delivery.ShippedReact 19ExpressTurborepo
track 04
Experiments & prototypes
Small builds that answer a question.
- Sumi: a mascot that shows what the AI is doingI built Sumi for Sumeru Tech (formerly Age of AI): three rounds of design exploration for a learning-platform mascot: from an owl, to a flame-shaped spark, to a glowing character whose body language acts out each step the AI takes. A concept, not a shipped product.ConceptDesign systemsAI interactionMascot
- Pronunciation trainer: Whisper in a containerAn AI pronunciation trainer using Whisper for speech recognition, packaged in Python with a multi-stage Docker build.PrototypeAdaptedPythonWhisperPyTorch
track 05
Technical research
Studying benchmarks, protocols and patterns.
- τ-bench: studying tool-agent-user evaluationA fork of Sierra's τ-bench, which scores language agents on dynamic conversations with a simulated user, domain tools and policy documents, and reports pass^k reliability.StudyFork · studiedAgent evaluationTool-callingBenchmarks
- Merkle claim: an ERC-1155 airdrop that stores one hashAn ERC-1155 token contract that mints only to addresses proven to be on an allow-list, using a Merkle proof. The contract stores a single 32-byte root; the list lives off-chain and each claimant supplies their own proof.PrototypeSolidityMerkle treesERC-1155
- Code review: Aavegotchi's GHST staking diamondA long-form written review of Aavegotchi's GHST staking contract, an immutable EIP-2535 diamond: how it is structured, how its storage and facets work, and what each function does.StudySolidityEIP-2535Smart contract review
- Diamond-standard contractsSolidity explorations of the EIP-2535 diamond pattern: an auction, an upgradeable staking contract and a Foundry and Hardhat diamond setup.StudySolidityEIP-2535Foundry
track 06
Engineering notes
Short write-ups. The first ones are on the way.
Nothing published in this track yet. Entries will be added as they are finished.
track 07
Open-source contributions
Fixes and features sent upstream.