Engineering work

Work and projects

Day-to-day engineering, independent products and research interests.

Professional work

Datafusion Systems

Senior Software Engineer · Dubai · June 2024–present

I work on Datafusion’s Intelligence Platform (IPF), a system for combining and analysing data for investigations. My role covers full-stack and applied-AI development in an air-gapped environment.

I develop full-stack workflows and services for document processing, entity extraction and source-grounded retrieval, so answers can be traced to supporting records. My model work includes parameter-efficient fine-tuning with PyTorch and LoRA/QLoRA, held-out evaluation against base models, and regression checks to measure the effect of model changes.

Technical work
  • Build Angular and Node.js applications with Python/FastAPI services for document processing and entity extraction.
  • Develop retrieval workflows with spaCy and Sentence Transformers, linking answers to supporting source records.
  • Fine-tune language models with PyTorch and LoRA/QLoRA; evaluate against base models using held-out datasets and regression checks.
About IPF

Alongside my day-to-day engineering work, I build independent products and explore ideas through research and experiments.

Independent products

Products I build, with a route to try or inspect each one.

Draftlane

Development preview

AI-assisted purchase-order review for Shopify. Check customer purchase orders against a Shopify catalogue, resolve discrepancies, and prepare private drafts for approval.

Explore Draftlane
Details and evidence

Variable layouts and scans depend on configured and authorised AI processing. Merchant access is not open yet.

Ocelin

Experimental preview

Find earlier coding conversations and follow current Codex and Claude sessions across projects. Preview a saved conversation before choosing where to reopen it.

Preview and setup
Details and evidence

The Windows companion and browser dashboard read local activity. The optional Ask feature sends a selected snapshot to a model service.

Calibration Explorer

Public workbench

Compare a model’s confidence with observed accuracy, inspect calibration and decision thresholds, and save an assessment record with its evidence. Use it in the browser, or run the same assessment from Claude Code through a local MCP server that reads files you select and saves reports in your project.

Open the workbench
Details and evidence

Imported predictions are processed in your browser. Temperature is fitted on calibration rows; a separate split measures what changes. The worked example retains the complete results table and released workflow capture.

Venture

Kordane

I co-founded Kordane to build applied-AI products and systems across private knowledge, documents, speech and conversational workflows.

Explore Kordane

Cortex

Assist, Extract and Voice · Access by enquiry

Find answers in private knowledge, extract information from documents, and work with speech.

Explore Cortex
Details and evidence

Further capabilities are planned.

Axon

AxonCX · Access by enquiry

Handle customer conversations, automate follow-up workflows, and hand off to a person when needed.

Explore Axon
Details and evidence

Live Agent is in development.

BoundSense

Browser demo · Access by enquiry

Protect information, control permitted AI routes, and retain reviewable records of decisions.

Explore BoundSense
Details and evidence

A browser demonstration is available. The Windows and Chrome tooling is being prepared for public distribution.

Forge

Training and evaluation · Access by enquiry

Evaluate and specialise models for a defined task, with recorded comparisons to support review.

Explore Forge
Details and evidence

Selection, runtime and distributed evaluation are planned. This is Kordane’s product, separate from the Training Forge reference below.

Research & experiments

Public implementations and recorded comparisons, with their inputs, methods and limits available to inspect.

Reference implementations

Careful Machine: checking answers against records

Public reference implementation

A public reference implementation showing how structured evidence and deterministic checks evaluate model-written claims, with synthetic examples and recorded runs.

Explore the reference
Details and evidence

Training Forge: reviewing model updates

Public reference · Synthetic executor

A TypeScript state machine for tracking lesson review, evaluation decisions, promotion and rollback. The example uses a synthetic executor.

Read the reference write-up
Details and evidence

Recorded studies

Do domain prompts help?

Recorded routing study

I compared generalist and domain-specific prompts around the same model, using generated questions with checkable answers and a transcript that can be replayed.

Under the strict first-token scoring registered before the run, the generalist prompt scored 82.75% and the routed domain prompts 67.75%. On every question the routed prompts lost to the generalist under that scorer, the reply showed its working first, answered "No." with a full stop, or was cut off at the 256-token limit; none was a finished wrong answer.

Scored on final answers, a rule chosen after reading the replies, the generalist scored 88.00% and the routed prompts 91.00%. That difference is not significant (exact McNemar p = 0.14).

Read what the strict scorer measured
Measurements and limits
Accuracy in the recorded run, under each scorer

Strict first-token scoring, registered before the run

Generalist prompt82.75%
Routed domain prompts67.75%

Final answer, chosen after reading the replies

Generalist prompt88.00%
Routed domain prompts91.00%

claude-haiku-4-5 · 2 September 2026 · 400 scored questions, 372 unique. Both conditions use the same model with different prompts, and both scorers read the same recorded replies.

The final-answer rule was chosen after the replies had been read, so it explains the strict result rather than replacing it. Every question the routed prompts lost to the generalist on final answers was a reply that hit the 256-token limit before it finished, so that result also depends on the token budget.

On final answers over the 372 distinct prompts, counting each repeated question once, p = 0.40. A +3.0 point gap is below what 400 questions can reliably detect, so not significant does not mean the two are equal.

Repeated identity questions are correlated, and this is a single recording of one model with one prompt set, not evidence about routing in general. Settling the question would need a new recording with the scorer declared in advance.

Open-source libraries

Evidence & verification

grounded-claimscareful-verifieru-pack

ML evaluation & calibration

calibratedab-significanceprobe-headsfrozen-evalsilent-zero

Runtime & process controls

careful-routergpu-quiescenceclean-room-guard

Install and contribution details
Install

npm install @m-sanchez/<name>

Eleven on npm under @m-sanchez · gpu-quiescence on PyPI · the reference implementation from a pinned git tag

Published in 2026. Feedback and bug reports are welcome through the repositories.

Full profile on GitHub