Independent Research Laboratory

Computational science, healthcare, and local AI architectures.

Officina develops deterministic scientific software in Rust, studies tool-augmented local language models, and explores alternative statistical and neural architectures.

01 / Core Architecture

Pharmacometrics & Verifiable Local Agents

A modular workflow separating numerical execution from linguistic reasoning, ensuring every reported figure is traceable and reproducible.

Who does what, in order Six steps across four lanes. You ask a question. Apothicaire plans. Caladrius computes every number and seals it. Apothicaire explains. The Gate checks that each number has a source; if one has none, Apothicaire rewrites its answer once, and what remains is shown in a badge. You read the answer. 1. Aska question aboutyour data 2. Planchoose thecommand to run 3. Computeevery number,then seal it 4. Explainthe result,in plain words 5. Checkeach numberhas a source 6. Readan answer youcan re-derive You Apothicaire Caladrius The Gate at your own PC words · local 27B numbers · Rust engine check · no model MCP call sealed result A number without a source: one retry, then a visible badge.

Caladrius

Rust · Pharmacokinetics (PK)
Open Source

Pharmacokinetic engine written in Rust: non-compartmental analysis, compartmental models and weighted fitting, built for strict cross-platform reproducibility.

  • Every answer carries a SHA-256 seal, and a saved project can be replayed step by step.
  • The same digits on Linux, Windows, macOS and WebAssembly, compared in CI on every build.
  • 48,955 of 48,955 reference values reproduced against PKNCA, 256-bit closed forms and R fits.

Apothicaire

Local 27B LLM · Agent
Research

Local 27B language model agent interfaced with Caladrius via the Model Context Protocol (MCP) for domain-specific pharmacokinetic workflows.

  • Operates entirely on local hardware with zero external API dependencies.
  • Investigates parity between tool-augmented 27B local models and frontier models on structured PK tasks.
  • A prototype tested on synthetic exercises. Not for a real patient or a regulatory decision.

Verification Layer: The Gate

Python

Deterministic check, with no model inside, run on every answer of Apothicaire. A number that comes neither from a Caladrius result nor from the user's own message is flagged: the model regenerates once, and whatever remains is listed in a visible badge.

Output Guard
02 / Ongoing Research & Internal Stack

Fundamental AI & Infrastructure

Experimental decision architectures alongside the internal retrieval and orchestration systems powering Officina.

OLS Prime

Decision Model · Statistical Memory
AI Research

Decision model exploring a Softmax-free attention mechanism formulated through Ordinary Least Squares (OLS).

  • Replaces standard Softmax attention weighting with closed-form OLS projections.
  • Paired with Matrix, a hierarchical statistical tree memory architecture.
  • Hypotheses are registered before they are tested. In Report n°2, H3 does not hold and is published as such.

Kit

RAG & Agent Harness
Internal Stack

Proprietary retrieval-augmented generation (RAG) pipeline and agent harness developed internally to support Officina's research and tooling workflows.

  • Powers Officina's internal knowledge retrieval, indexing, and agent evaluation loops.
  • Closed-source infrastructure currently in active internal R&D.
Status: Active Internal Development · Not Commercially Distributed
03 / Roadmap

Caladrius: Foundations First, Then Features

Finish what is on the branches, then lay three foundations before any new model or estimator. There are no dates. A phase ends on a check against an independent tool, not on a deadline.

Phase 0Current
Close iteration 3
Three compartments, the maximum-likelihood fit, FOCEI with diagnostics and population simulation, merged on main.
Ends whenThe conformance table on main covers them and the digest is equal on the four targets.
Phase 1
Numerics kernel
Nothing visible. Maths functions, linear algebra, quadrature and random streams gathered in one numerics layer that every later phase calls, starting with the caladrius-math crate.
Ends whenEach function gives the same digest on the four targets.
Phase 2
Events, outputs, groups
Custom dosing regimens and several doses per subject. Multiple-dose, steady-state and urine NCA. Ratios and summary tables.
Ends whenEach new parameter has a PKNCA oracle case and regimens agree with rxode2.
Phase 3
Model interface and likelihood
Covariate models, the censored likelihood, post hoc estimates, VPC and NPDE, and design evaluation.
Ends whenFits agree with nlmixr2, VPC statistics with the vpc package, FIM with PFIM.

After phase 3, two tracks are independent and can run side by side.

Track · Phases 4 and 5

Models

  • ODE solver. Michaelis-Menten elimination and turnover models. Ends on deSolve and diffsol cross-checks.
  • Model language. User-written ODE models and target-mediated disposition. Ends when every built-in model, rewritten in the language, reproduces its conformance values.
Track · Phases 6 and 7

Estimators

  • More estimators. Iterative two-stage, SAEM, importance sampling, Laplace and quadrature. Ends on agreement with saemix within Monte Carlo error.
  • Resampling and search. Bootstrap, likelihood profiling, covariate and structure search. Ends on comparisons with Pharmpy and saemix on public datasets.
Alongside, at any time: XPT files, CDISC result datasets, a report with the run record · Out of scope: bioequivalence, IVIVC
04 / Publications & Talks

Papers, Preprints & Presentations

Working manuscripts, technical notes, and conference presentations from Officina members.

2026In Progress
Benchmarking Tool-Augmented 27B Local Models via MCP on Pharmacokinetic Workflows
Manuscript in preparation · Caladrius & Apothicaire
Dec 2026Poster
Provenance is cheap, intent is not
ELLIS UnConference, Paris · Session: Next Generation of Discovery Loops
2026Note
Cross-Platform Floating-Point Reproducibility and SHA-256 Verification in Rust
Technical note · Caladrius
Read →
05 / Researchers

Team & Collaborators

Officina brings together independent researchers working across healthcare, applied mathematics, and AI systems.

Founder · Principal Researcher
Abdallah Ragued

Pharmacometrics (DMPK), deterministic Rust computing, local LLM agents, and alternative decision architectures.

Collaborating Researcher
Ilias El Badaoui

Applied mathematics and inverse problems, with a focus on MRI super-resolution medical imaging.

Additional contributors and interdisciplinary research tracks will be listed here as ongoing projects are published.