Architecture memos
How the runtime, memory substrate and personas fit together, and the decisions behind each boundary.
Lab notes
Notes with their hedges left on.
Research notes for teams building serious AI systems: retrieval, autonomy, interface design, evaluation, and the discipline to move from experiments to reliable operations. If a figure appears, it is tagged or it does not ship.
Safety drills
Failure modes worked through in advance: what fails closed, what gets logged, and who reviews it.
Operator UX
Control surfaces for the people running the system: calm defaults, visible state, and honest status.
ENGINEERING · September 29, 2026 · 3 min read
ABI Is Rust Now: What Changed, and What Didn't
The ABI runtime moved from Zig to a nightly-Rust workspace on July 30, 2026, and WDBX became its own sibling workspace in August. Here is what that means for anyone building from source.
By MLAI Research
RELEASE · June 9, 2026 · 7 min read
WDBX V2: Observable Memory, Multimodal Inputs, and Research Alignment
The second iteration of the WDBX runtime ships an observable pipeline: block-chain memory with temporal queries, multimodal input fusion, an async neural job queue, and a research-alignment telemetry layer that scores every turn against the constitution.
By MLAI Research
FROM ABBEY · June 8, 2026 · 4 min read
On Saying “I'm Not Sure”
A note in Abbey's voice — the empathetic-polymath persona — on why an honest “I don't know” is one of the most useful things an agent can say.
By Abbey
ENGINEERING · June 6, 2026 · 8 min read
Inside the WDBX Hybrid Ranker: Four Factors, One Score
Semantic similarity is only the first of four factors. Here is the exact score WDBX uses to rank retrieval candidates — and why every factor is bounded to [0,1].
By MLAI Research · WDBX Core
RESEARCH · June 4, 2026 · 7 min read
How Sparse Evidence Attention Decides What the Model Sees
Before a model reasons, something must choose which records reach the context window. SEA scores eight criteria, then packs greedily under a hard token budget.
By MLAI Research · Abbey
SAFETY · June 2, 2026 · 6 min read
The Weights Behind Abbey, Aviva, and Abi
Persona routing is not a hidden model call. It is a normalized weight vector with hard policy overrides and an inspectable strategy decision.
By MLAI Safety Engineering
FIELD NOTE · May 21, 2026 · 7 min read
Designing AI Systems That Can Explain Their Last Move
A practical guide to capturing retrieval paths, policy checks, and operator decisions so teams can debug autonomous workflows after the fact.
By MLAI Engineering
ENGINEERING · May 18, 2026 · 9 min read
What We Plan to Measure Before Calling an Agent Production-Ready
Latency and accuracy are not enough. The gate we are building tracks tool-boundary violations, source coverage, rollback paths, abstention quality, and human-review burden.
By MLAI Safety Engineering
RESEARCH · May 12, 2026 · 12 min read
From Vector Search to Weighted Backtrace Graphs
Why retrieval infrastructure needs relationship-aware context and why ranking alone is not sufficient for high-stakes answer generation.
By MLAI Research
SAFETY · May 8, 2026 · 8 min read
The Case for Separate Planning, Review, and Execution Agents
Role separation lets teams encode checks and balances into an AI workflow instead of relying on a single agent to police itself.
By MLAI Safety Engineering
PRODUCT · April 29, 2026 · 6 min read
Making AI Control Surfaces Feel Calm Under Pressure
Interface details that prevent overload: progressive disclosure, confidence bands, decision diffs, and emergency stop affordances.
By MLAI Product
OPERATIONS · April 18, 2026 · 10 min read
What Incident Reviews Need From AI Logs
How to structure traces so teams can reconstruct context, model output, policy decisions, tool calls, and human overrides after a production event.
By MLAI Operations
DEPLOYMENT · April 11, 2026 · 8 min read
Private AI Does Not Mean Invisible AI
Offline and VPC deployments still need measurable controls: local audit trails, release gates, repeatable evals, and operator-visible confidence signals.
By MLAI Engineering
FIELD NOTE · February 18, 2026 · 7 min read
Neural Backtracking: What a Block Chain Buys an AI Memory
Hallucination debugging is archaeology. WDBX chains every interaction block to its parent so you can prove nobody edited the record — and so a system can be built to walk back to the exact moment a model drifted.
By MLAI Research · WDBX Core
ENGINEERING · January 8, 2026 · 6 min read
Surviving Writergate: Migrating ABI to Zig 0.16
Zig 0.16 rebuilt std.io around vtable-based Reader and Writer interfaces. Ninety-five files later, here's what the migration actually looked like — and the patterns we standardized on.
Historical: the superseded Zig implementation
FIELD NOTE · November 30, 2025 · 6 min read
Why WDBX Is Written in Zig
Garbage collection pauses are invisible in the mean and fatal in the tail. WDBX is Zig because a real-time vector database is a p99 product, not a p50 product.
Historical: the superseded Zig implementation