About lakecode.
The coding agent with a governed, compounding memory.
What we're building
Models are rented; everyone gets the same ones. What an organization actually owns is the knowledge that decides correctness in its systems — why a constant has the value it has, which IDs are valid, what failed last quarter, how one table derives from another. Almost none of that is written anywhere a model can see.
lakecode is a coding agent built around that asset. It learns your systems once and persists what it learns as structured claims with provenance — so every later session, by anyone on the team, starts already knowing. The result is measured, not vibes: cheaper where you'd already be right, correct where you'd otherwise be confidently wrong.
Four rules we don't break.
Measured claims only
Benchmarks are pre-registered before we know the result, model-pinned, dated, and shipped with caveats. When a run misses the target, we publish the measured number, not the target. We publish our cost model for the same reason.
Glass box, not black box
Everything the agent knows is inspectable: claims carry provenance and confidence, write-back is gated, and knowledge climbs a promotion ladder with review before it becomes org truth.
Own the loop
Capture at commit time and governed write-back can't be bolted onto someone else's agent. That's why lakecode is a full agent, not a plugin.
Your data is yours
Ingestion does store knowledge: your code and docs become claims in your org's substrate (a dedicated database per beta org). We say that plainly because the flip side matters: the substrate is exportable at any time and deleted on exit — terms that live in the beta agreement, not just this page.
The company history is a list of runs.
We mark progress the way we ask you to judge it — dated, pinned, caveated.
- May 2026Stage-1 pipeline economics land: $0.033 vs $0.113 per question at matched accuracy — same model both sides, so the gap is the compiled context.
- 2026-06-04The efficiency A/Bs: ~27% cheaper at identical correctness, warm vs cold substrate (20 runs, 0 failed).
- 2026-06-08The grounded-coding causal matrix: one substrate claim flips correctness from 0/3 to 3/3.
- 2026-06-09The flywheel validates causally: a convention taught once is honored by fresh, independent sessions — cold 0/2, warm 2/2, nothing fabricated.
- 2026-06-10lakecode.ai ships — every number on it dated, model-pinned, and caveated.
the name · and the duck
Why “lakecode”? Because the substrate behaves like a lake: knowledge sinks in, settles, and stays — and the enterprise edition lives on the lakehouse. The code part you already know.
And the duck? Rubber-duck debugging, completed. You explain your system once — to a duck that remembers — and the next engineer never has to explain it again. The duck's name is still in review; promotion to org-wide truth pending.
Founder-led, deliberately small.
Design partners are onboarded personally, one team at a time — the person who runs the benchmarks is the person who answers your email.
founder
Sachin
Building lakecode — the agent, the substrate, and the benchmark culture that keeps both honest.
support
support@
Beta partners also get a named feedback channel — it's in the agreement.
Prefer email? Reach the founder directly.