What I am thinking about next

Coming maybe

Half-formed ideas, sketches in the wiki, problems I keep returning to. None of these are promised. None are dated. They live on this page because I would rather you see the shape of the thinking now than read a polished roadmap that is half fiction. The disclaimer applies until 14 February 2030.

6
Themes I am chasing
until Feb 2030
Bound by the disclaimer
honest, not roadmap
Status

The themes I keep returning to

Eight ideas, in roughly the order they sit on my desk. Each one is real enough to describe in a paragraph. None are real enough to promise.

Prototype

slipstream extensions

A second wave of subagents and skills on top of slipstream. Tighter Postgres tooling, a sharper review subagent, a release subagent that drives a tagged push from a clean working tree. Each one earns a place by replacing a manual step I do every week.

Researching

SarmaLink-AI paid tier

A hosted tier that wraps the open-source gateway with metering, usage analytics and a portal. Launch window is 14 February 2030, in line with the disclaimer. Until then the open-source gateway is the only path and stays MIT-licensed.

Sketch

lsmdb compaction work

A second pass at the compaction strategy in lsmdb. Tiered versus levelled is the obvious decision. The interesting one is what happens to bloom filters and the read amplification curve in the middle. A sketch of the trade-offs sits in the wiki.

Sketch

raftkv membership changes

Joint consensus for safe membership changes. The fault-injection harness is the right place to land this because it can prove the invariant under partition. Until that is wired up, joint consensus stays on the wiki, not in the codebase.

Researching

sandboxd capability ABI

A narrow capability ABI on top of the deny-by-default host. The point is to make capability passing explicit at the boundary so a sandboxed program can ask for, say, a single read-only file handle, and nothing else, audited at grant time.

Sketch

A real observability starter

A starter that bootstraps Prometheus, Grafana, Loki and OpenTelemetry with dashboards that read like a runbook, alerts wired to Telegram and service-level objectives baked into the histograms. The pieces already live in k8s-ops-toolkit. The missing layer is the opinionated wiring.

Researching

forge-infer speculative decoding

Speculative decoding with a small draft model alongside the main. Paged KV cache is in place. The interesting work is the acceptance schedule and the cost of the draft pass under a continuous batching loop.

Listening

Memory primitives, generalised

The signal-ranked recall in slipstream wants to be a standalone primitive that other agents can pick up. Open question whether the right shape is a library, an MCP server or a sidecar process. Listening to how slipstream gets used before I pick.

How to read this page

These are not promises

A theme on this page is a theme, not a commitment. Some will ship, some will turn out to be a bad idea once I prototype them, some will be cannibalised into something else.

No dates

Dated roadmaps from a one-person studio are mostly fiction. If something here ships, it will land in a repository and a blog post on the day it lands, not a quarter early.

The disclaimer applies

I am bound by the public disclaimer until 14 February 2030. Anything that becomes paid lives behind that gate. Everything else stays MIT-licensed.

Want to push an idea forward?

Open an issue on the repository that the idea touches. A real use case usually moves something from sketch to prototype faster than anything else.