A blocking integration should not be the reason you miss your launch.

A mobile trading product needed prices from seven heterogeneous venues, on-chain and off, delivered to phones fast enough to trade against. Each venue fails differently, and a chart that freezes or lies is a support ticket at best and a bad fill at worst.

Sole engineer on the service: protocol specification, venue ingestion, storage design, deployment topology, and the operational tooling around it.

Independent Tokio tasks per venue with backoff reconnect and liveness watchdogs, funnelling into a bounded channel that deliberately drops stale ticks under backpressure — the correct failure mode for price data. Per-symbol transport precedence lets on-chain marks own a symbol while fresh and hand back to the venue socket when they age out, so an outage degrades quality rather than emptying a chart. A database-driven market registry means new markets go live without a deploy. Prices are served over a documented WebSocket contract with snapshot-first ordering guarantees, so the mobile client needs no buffering, deduplication, or staleness logic of its own.

Removed the single-transport dependency behind price outages, cut on-chain ingress roughly eightfold by slicing account frames server-side, and made market listing an operational action rather than an engineering one. 263 tests, none ignored, including wire-protocol assertions taken directly from the specification.

Real money moving through perpetuals, a hosted-perps router across four venues, prediction markets, and a high-leverage sixty-second game — with custodial wallets and no tolerance for a position ending in an unknown state.

Backend architecture and delivery: transaction construction, job execution, venue integrations, reconciliation tooling, and post-incident redesigns.

Every high-risk module states the invariant it protects. The market registry fails closed rather than defaulting to a wrong account slot. Withdrawal sweeps only write a terminal record after funds are confirmed home; on exhaustion the position stays open, ops are alerted, and a reconciler re-enqueues it. Protective stop and take-profit legs are placed and confirmed before the previous ones are cancelled, so a failed reset can never strip existing protection. Expiry and liquidation are made mutually exclusive by a database write-fence. A gasless-transaction incident was closed by re-specifying the budget in terms of confirmed sponsorships and deploying it under a new key namespace, so stale semantics could not survive the release.

Introduced deterministic recovery paths across every funds-moving workflow, replaced silent defaults with explicit refusals, and gave the team paging that fires once per incident instead of not at all. A real-funds mainnet harness with cent-level conservation assertions makes release verification repeatable rather than a judgement call.

Every vector database quantises embeddings to save memory, and every retrieval-augmented agent inherits that choice. Recall is easy to measure. What nobody measures is whether the codec changed the answer a given chat model gave. The two layers of compression in an agent, the retrieval key and the prompt payload, are routinely confused, and neither is benchmarked end to end.

Sole author: workspace design, the codec trait and distance kernels, model connectors, the compressed-domain store, the benchmark matrix, and the experiment programme with its write-ups.

A five-crate Cargo workspace built around one Codec trait, so f32, int8 and binary vectors are scored on their compressed bytes with no decode step. Anthropic, OpenAI-compatible and local ONNX models sit behind two traits. A SQLite store searches in the compressed domain and records every run: embedding model, codec, chat model, timings, tokens, retrieved chunks and the answer. The benchmark is an explicit matrix of embedding model × codec × chat model × dataset. Integration tests drive the real binary against an in-process mock of both provider APIs, so CI needs no key and no model download.

The first published experiment, on 8,036 SciFact chunks across two embedding models, confirmed that int8 is close to lossless at 3.9× smaller and that binary loses a third of the neighbours at 32× smaller, with the loss set by the embedding model rather than the codec. Every number ships with its commit, config and command, and negative results are kept. The roadmap runs through LLM-judged scoring, product quantisation, payload compression and an agent-memory loop.