Lucio Durán
C.T.O. at LOi
Lucio Durán is C.T.O. at LOi and works from Uruguay. His work combines applied artificial intelligence, product architecture and engineering leadership. He publishes open-source projects on AI agents, language model evaluation and retrieval-augmented generation (RAG), with source code, evaluation methods and explicit limitations.
Invite me for an AI talk or training · Explore projects and results
Technical talks and training
Proposals for communities, universities and teams: AI agents, model evaluation, RAG and inference. Sessions draw on open-source projects, engineering decisions and results that can be inspected.
Artificial intelligence expertise
AI agent architecture
AI agents with tools, authorization and failure recovery. Lucio Durán’s projects include source code, tests and explicit limitations.
LLM and AI agent evaluation
Reproducible model and agent evaluation: regressions, citations, structured output and prompt injection. Source code and results by Lucio Durán.
RAG: retrieval, search and evaluation
Lucio Durán’s RAG projects: hybrid retrieval, chunking, reranking, bilingual evaluation and incremental indexes with reproducible evidence.
AI inference: latency, memory and quality
Local AI inference and response delivery: token latency, quantization, streaming, caching and model routing. Projects by Lucio Durán.
Featured projects
WebGPU Local Inference
Inspectable browser ticket classification with native WebGPU, real WebAssembly fallback and recorded cached startup during a local asset outage.
Quantization Quality Lab
A real MiniLM CPU quantization study separating weight size, process memory, retrieval quality and latency, with raw regressions and output drift preserved.
Agent Policy State Machine
An executable authorization state machine with finite TLA+ safety checks, compensation and replay of the actual TLC transition graph.
Agent Replay Runtime
Durable tool execution with event replay, idempotency keys and recovery after real process crashes. Includes effect traces and cases that stop instead of retrying automatically.
RAG Evaluation Workbench
Reproducible RAG evaluation with BM25, dense and hybrid retrieval, local CPU inference, and separate retrieval and answer metrics.
Contents
Experience
Roles, companies, responsibilities and contributions.
Expertise
Areas of work, experience and related technologies, with individual records and references.
Education
Courses, institutions and credential links.
Biography
Career and personal history.
Blog
Technical publications, ordered by date.
Projects
Public software, source code and reproducible results.
Architecture decisions
Architectural context, decisions and consequences.
CV
Complete curriculum vitae, with print support.
AI talks and training
Proposed topics, audiences, formats and material for an AI talk or workshop.
Contact
Contact information and public profiles.
Complete index
Every page, including individual records.
Content access
Formats, sources and how this site works.
Privacy
Privacy policy.
Links
Built for machine reading
46 implemented controls help agents discover, retrieve and cite my content. Each control links to its implementation or technical documentation.
01 / Discovery06 controls
- Implemented: D01Public crawling
robots.txt allows public content and advertises the sitemap. robots.txt
- Implemented: D02XML publication map
The sitemap is built from the published catalog, including individual records. Sitemap
- Implemented: D03Index without JavaScript
Every page has an HTML link in an index organized by content kind. HTML
- Implemented: D04Paginated JSON catalog
Canonical URLs, kinds, languages, available dates and representation links. HTTP / JSON
- Implemented: D05llms.txt guide
Index for clients using this proposal. Generated from the same catalog. llms.txt
- Implemented: D06Discovery on every page
HTML and Link headers advertise alternate formats and the API catalog. HTML / Link
02 / Content and representations07 controls
- Implemented: F01Server-delivered content
Main text is in the initial HTML response, before JavaScript runs. HTML
- Implemented: F02Markdown by URL
Every page supports .md. The English root uses /index.md. Markdown
- Implemented: F03JSON by URL
Content, headings, sections, links, language and structured data in one document. HTTP / JSON
- Implemented: F04HTTP content negotiation
Accept selects explicit Markdown or JSON; Vary: Accept separates representations. Accept / Vary
- Implemented: F05One source for three formats
Markdown and JSON derive from HTML main, without a separate biography for bots. HTTP / JSON
- Implemented: F06Structure preserved
Lists, tables, code, links and image references survive serialization. Markdown
- Implemented: F07Decoration excluded from content
Navigation, controls and decorative seals are excluded from Markdown and JSON. HTTP / JSON
03 / Identity, attribution and citations06 controls
- Implemented: C01Canonical document URLs
HTML, Markdown, JSON and search results refer to the same publication. HTTP / JSON
- Implemented: C02Anchored sections
Headings have identifiers for linking a specific passage. HTTP / JSON
- Implemented: C03Consistent identity
Person uses a stable identifier and links to the author’s public profiles. JSON-LD
- Implemented: C04Content-specific structured data
Profile, articles, projects and credentials publish structured data alongside visible text. JSON-LD
- Implemented: C05Declared author and dates
Articles identify the author and preserve available publication and modification dates. JSON Feed
- Implemented: C06Project evidence and limitations
Projects link repositories, revisions and published results with their scope. Fuentes / Sources
04 / Retrieval and context05 controls
- Implemented: R01Search over published text
BM25 searches passages extracted from build HTML. Scores measure lexical relevance. HTTP / JSON
- Implemented: R02Explicit filters and pagination
Language, kind, offset, limit and next support bounded, repeatable traversal. HTTP / JSON
- Implemented: R03Results with citations and versions
Every result includes an excerpt, citationUrl and the indexed document’s contentHash. HTTP / JSON
- Implemented: R04Read individual sections
section selects a heading and its subsections, with its own hash and source hash. HTTP / JSON
- Implemented: R05MCP context windows
Reading supports limits and continuation by Unicode characters. The hash identifies the complete source. MCP
05 / Contracts and interfaces07 controls
- Implemented: I01RFC 9727 API catalog
A Linkset advertises endpoints and links their contracts and documentation. Linkset
- Implemented: I02OpenAPI 3.1 contract
Operations, parameters, limits, responses and errors are documented for HTTP clients. OpenAPI
- Implemented: I03Document and passage JSON Schema
The two response types have distinct schemas and a describedby header. JSON Schema
- Implemented: I04Read-only MCP
Tools to list, search and read using the official SDK and HTTP transport without persistent sessions. MCP / HTTP
- Implemented: I05RFC 9457 structured errors
Document, search, catalog, inventory and feed APIs return problem+json for handled errors. Problem Details
- Implemented: I06WebMCP in supporting browsers
Two read-only tools use document.modelContext when available. They validate inputs and support cancellation; experimental browser support is required. WebMCP / experimental
- Implemented: I07ARD discovery
An ARD manifest links the OpenAPI contract and RFC 9727 catalog with representative queries. ARD v0.91 proposal
06 / Freshness and integrity05 controls
- Implemented: V01SHA-256 content identity
Documents explain which fields are hashed. It identifies a version, not the truth of its claims. HTTP / JSON
- Implemented: V02Build inventory
Paginated URLs and hashes, collection version and generation time for comparing snapshots. HTTP / JSON
- Implemented: V03Conditional reads
ETag and If-None-Match allow 304 responses without retransmitting unchanged documents. ETag / 304
- Implemented: V04Metadata without a response body
HEAD is available on documents, search, catalog, inventory and feeds. HEAD
- Implemented: V05Publication feeds
JSON Feed and Atom include article text; RSS publishes summaries. Dates and links come from source files. JSON Feed / Atom / RSS
07 / Reading and access05 controls
- Implemented: A01Explicit languages
HTML declares the content language; JSON and document responses preserve it. HTTP / JSON
- Implemented: A02Linked EN and ES versions
hreflang links available versions and declares x-default. HTML head
- Implemented: A03Native HTML semantics
main, headings, lists, tables and links express the content structure. HTML
- Implemented: A04Photo references, decorative seals
Photos are retained. Project seals are empty elements with aria-hidden. HTML / aria-hidden
- Implemented: A05Public APIs without sessions
HTTP read APIs support CORS and OPTIONS. No account, cookies or key required. CORS / OPTIONS
08 / Consistency and boundaries05 controls
- Implemented: B01Versioned profile
Pages and APIs use the same versioned professional profile snapshot. HTTP / JSON
- Implemented: B02Index generated from HTML
The build rejects missing or duplicate heading anchors and expected pages absent from the index. Build
- Implemented: B03Site-scoped reading
The reader accepts known public paths and bounds source size, time and redirects. Contrato / Contract
- Implemented: B04Content treated as source material
MCP tools identify document text as untrusted material for agent instructions. MCP instructions
- Implemented: B05Machine-readable checklist
This list and its JSON use the same registry. Each control has an identifier and evidence link. HTTP / JSON