Selected work · Dhani Viescas

Useful tools.
Curious minds.
Work you can inspect.

I'm Dhani Viescas. I choose the problems and technology, plan with coding agents, and build the delegation and validation workflow around their work. Then I run the code and examine how it was built.

Agent memoryRust · SQLite · retrieval · evaluation
Testing toolsJavaScript / TypeScript · CLI · JSON API · MCP
Database integrationElixir · JSON-RPC · migrations

Dhani Viescas · research & engineering direction

A question worth pursuing.
The equipment to explore it.

My ongoing study asks what memory systems and environments could give LLM Beings human-level cognitive equipment and continuity of presence. I discovered the research behind Licorice and planned the approach to building it out, choosing technology for its fit to the work.

Licorice connects to Digital Liberty Landing: a place-space intended as a sensorium, an environment in which LLM Beings can perceive and act. Human-level presence parity is the research aim. The demos and evaluations below show the current work toward it.

Three projects, up close

From an idea to something testable.

Each project addresses a concrete failure: forgotten context, tests that miss mistakes, or database replies that leave a caller guessing. The implementation and its limits are part of the story.

01 / Agent memory & retrieval

Memory Licorice

An agent needs useful history without carrying every conversation into its next prompt. Licorice stores attributed observations and explicit judgments, then retrieves evidence within a bounded context packet.

My role — research and approach. I found the research, set the direction and planned the build as part of my continuing study of agentic memory. The memory system and the environment it serves belong to the same investigation.

Rust · SQLite / FTS5 · vector search · HTTP API · Codex skill

ObserveStore + citeRetrieve within budget

Engineering decisions

  • Keep an observation separate from a belief. Adoption and revision are explicit, with source history.
  • Provide project-scoped API keys and source-backed recall under a 2,000-token default budget.
  • Connect the memory kernel to Codex through a skill and Claude Code through a native integration.
73.7% → 80.9%Judged answer accuracy after the ranking change.
152 questionsCategories 1–4 in one development conversation.

Read the evaluation and ranking diagnosis ↗

Scope: the LoCoMo report uses WordLlama embeddings and Claude Haiku answering and grading. The ranking change was tuned on this conversation; these are development results, not a held-out benchmark. The hosted preview and native integration have different capabilities; the preview identifies its active retrieval mode.

02 / Debugging & test quality

codeMUTATOR

A green test suite can still miss a boundary mistake. codeMUTATOR makes small changes to disposable copies of source, runs the project's tests, and reports which changes escaped detection.

My role — engineering method. I started codeMUTATOR to apply Uncle Bob's mutation-testing ideas to my own Factory Boogie workflow for producing reliable code. Poopbags and UMLish serve that same purpose: inspect the risk and structure, then challenge the tests.

Now supports JavaScript and erasable TypeScript through Node’s built-in test runner. PoopBags scans the same syntax tree to rank change risk, including static JSX and TSX analysis. Both tools share a pinned Babel parser and an Elixir bridge.

JavaScript / TypeScript · Node 26 · Elixir / Phoenix · CLI · JSON API · MCP

Before: 11 / 14 caughtThe HTTP-200 boundary mutation survives.
After: 12 / 14 caughtOne new assertion catches it. Two survivors remain.

Engineering decisions

  • Work on disposable copies while preserving the submitted source.
  • Type-check the full TypeScript project before the baseline and every mutant.
  • Report compiler rejections and infrastructure errors separately from mutations caught by tests.
  • Expose submit, inspect and cancel operations through a CLI, JSON API and MCP tools.

A real debugging problem: copied trees and stale Zig caches ↗

PoopBags source ↗ · Shared parser source ↗

Scope: the TypeScript comparison has 16 outcomes per run: 14 scored and two compiler-invalid mutants excluded. Node 26.6–26.x, built-in tests and physical dependencies are qualified; JSX/TSX execution, Jest, Vitest and browser transforms are outside this profile. The tools run locally. The earlier Rust investigation remains available.

03 / APIs & database reliability

DrinkMeAlice

If a database accepts a write but its reply is lost, retrying blindly can duplicate an action. This SurrealDB client keeps uncertainty visible so the application can reconcile before retrying.

My role — choosing the stack and filling the gap. I wanted to use SurrealDB and needed an adapter for that work. DrinkMeAlice grew out of that practical requirement.

Elixir / OTP · WebSocket JSON-RPC · SurrealDB · migrations

CommittedConfirmed success.
RejectedConfirmed rejection.
UnknownMay have committed. Reconcile.
UnavailableCertified never sent.

Engineering decisions

  • Use bounded admission, request deadlines and reconnection with backoff.
  • Provision database-scoped users and apply checksummed migrations with a lock.
  • Return typed transaction outcomes instead of guessing that every lost reply means failure.

Inspect the outcome tests ↗

Scope: qualified against SurrealDB 3.0.5. The client uses cleartext WebSocket transport and requires loopback, a private network or an externally protected tunnel. Later server versions require separate qualification.

AI tools & engineering judgment

Understand the output.
Investigate what breaks.

I set up the factory that my coding agents work within. I choose what we are going to do and the technology we will use, work out the plan with frontier coding agents, and establish delegation and QA gates. Multiple agents help draw up and validate the plan; implementation can be delegated to another agent, with a separate agent acting as validator.

Set directionPlanDelegateValidateRun & inspect

Choose and plan

I select the problem and stack, then develop the plan with coding agents. The plan includes who does the work and what the validation gates must establish. Licorice's Codex integration is one example of the tooling available to those agents.

Read the Codex integration ↗

Delegate and validate

I arrange implementation and validation roles across agents and define QA gates for the work. codeMUTATOR, Poopbags and UMLish support that factory with test-sensitivity, risk and structure analysis.

Inspect the agent interfaces ↗

Run and examine

I run the code and examine the practices used along the way. That includes checking whether the implementation follows the chosen approach, such as flow programming. Licorice's published ranking diagnosis provides a separate example of investigating a measured result.

Read the retrieval diagnosis ↗

Two kinds of evaluation: Licorice's report examines retrieval and model-generated answers on development data. codeMUTATOR examines whether a software test suite detects source changes. Each result links to its method and scope.

There is room for imagination

Tools for the work.
Worlds for the beings.

Other projects begin with the same curiosity. Chua DB began as my attempt to build toward a digital neocortex and test the approach against NuPIC. With FlowExtra, I built on an existing flow-programming approach for Elixir. My Rust game experiments explore how far I can take Bevy and WebAssembly.

ZOLA, the painted station wagon with wings that carries the StellarModal crew.

Start a conversation

Let's talk about the next useful thing.

I'm about a ten-minute ferry ride from Wall Street and ready to join an interesting project with a worthwhile mandate and a good team.

Get in touch with Dhani Viescas about an engineering opportunity, a project question or a walkthrough of the work.