Garage

A crash-durable task store for AI agents that guarantees a side effect happens exactly once, even under kill -9 at any point. 500 randomized crashes produced 0 double charges; a naive implementation under the identical harness produced 121.

Why: Built because the MCP Tasks spec defines a store interface and ships only an in-memory implementation, leaving persistence to implementers — and an agent that charges a card twice is a real failure rather than a theoretical one. Peak 4,949 tasks/s, with the ceiling traced to Postgres's commit path by elimination and confirmed by a Go port that hit the same limit.

Semantic search over primary-source philosophy that classifies whether a passage argues for or against a claim, not merely whether it is about the topic. On a held-out set written after the fix, 20 of 21 queries surfaced the philosopher who actually holds the position.

Why: Built after finding that embedding search matches vocabulary and imagery rather than the conclusion an argument reaches — so a query about free will returns everyone who mentions it, on either side, undifferentiated. Fixed with HyDE query expansion over a from-scratch HNSW index, and validated on queries written afterwards so the result could not be tuned into existence.

A metrics ingest pipeline that survives the label explosion that OOM-kills a naive one — proven by exit code. Same 20M-sample workload in the same 256 MB container: the naive ingester dies at 8M samples; Peras completes all 20M at 1.67M samples/s with ~3 MB of heap.

Why: Built after measuring, while instrumenting Hapax, what one high-cardinality label costs: adding task_id took an export from 22 series to 30,019 and the payload up 1,488x. Peras is the ingest-side defense — it attributes a breach to the exploding label dimension at runtime and aggregates it away instead of crashing, and CI re-proves the die-vs-survive contrast on every push.

A Next.js portfolio platform that unifies project case studies, in-repo writing, a curated Shelf, and an optional Mind Map in one personal system.

Why: Built to turn a personal site into a living ecosystem for projects, writing, experiments, and tools instead of a static resume page.

Full-stack subscription tracking MVP for monitoring recurring spend, upcoming renewals, and user-reviewed subscription data.

Why: Built to solve a real personal-finance workflow while practicing authenticated data ownership, database-backed CRUD, and product-quality dashboard design. Gmail discovery is experimental and still being hardened.

Streamlit prototype that extracts messy AI collaboration transcripts into reusable context modules and assembles transfer prompts under token and privacy constraints.

Why: Built to make long-running AI workflows easier to resume, audit, and hand off without copying entire conversations or leaking private context.

Flask blogging platform with voice-assisted drafting, Markdown publishing, authentication, comments, likes, bookmarks, and a public posts API.

Why: Built to explore how a standalone writing tool can integrate cleanly into a broader portfolio ecosystem instead of living as an isolated app.

Public practice journal with solved interview problems, final Python solutions, pitfalls, complexity notes, and reusable mental models.

Why: Built to make algorithm practice inspectable: not just final answers, but how mistakes, patterns, and interview triggers get turned into repeatable judgment.