As an AI Quality Assurance Specialist, I'm currently responsible for:

• - Conduct quality assurance evaluations of linguistic data and AI model outputs, ensuring compliance with project guidelines;
• - Evaluate AI agent performance using structured rubrics and predefined quality standards across conversational and NLP tasks;
• - Identify recurring error patterns, providing actionable, structured feedback to improve model performance;
• - Collaborate with cross-functional stakeholders to clarify requirements, resolve ambiguities, and refine QA and evaluation procedures;
• - Maintain detailed documentation of evaluation criteria, decisions, and edge cases to support auditability and consistency.

As an Applied NLP Engineer, I'm currently responsible for:

• - Design and deploy multi-LLM evaluation pipelines to improve reliability, factual accuracy, and consistency of AI-generated educational content;
• - Deploy structured prompt workflows and evaluation pipelines that coordinate multiple LLMs to validate, refine, and improve AI-generated outputs through iterative consensus and rubric-based assessment;
• - Expand hierarchical competency models and label structures used for automated evaluation, enabling scalable and auditable assessment pipelines;
• - Build automated workflows for multi-model comparison, including sequential refinement and consensus-based architectures for robust output validation;
• - Develop evaluation frameworks using hierarchical competency taxonomies and adaptable scoring rubrics for AI-generated educational content;
• - Implement hallucination detection and correction systems, leveraging structured taxonomies and evaluator-based validation strategies;
• - Engineer prompting strategies and structured outputs (JSON) to ensure deterministic, auditable, and scalable model behavior across tasks;
• - Adapt prompt templates, evaluation criteria, and structured label systems to ensure consistent behavior across different LLMs and multilingual educational datasets;
• - Integrate pipelines with GitHub Actions and external data systems (e.g., Airtable), enabling automated runs, experiment tracking, and longitudinal benchmarking.

As a Quality Controller, I was responsible for:

• - Lead quality assurance for language data projects, ensuring accuracy, consistency, and alignment with client specifications;
• - Collaborate with project managers and subject matter experts to refine task guidelines, documentation and resolve quality issues;
• - Conduct audits and performance reviews of annotators and reviewers;
• - Analyze error patterns and generate insights to improve data quality and team efficiency;
• - Support the onboarding and training of new team members, ensuring alignment with quality standards;
• - Create, update, and edit annotation and QA guidelines to reflect evolving project needs and improve clarity for annotators;
• - Use quality tracking tools and dashboards to report on KPIs, monitor SLA compliance, and drive continuous improvement.

As a Senior Linguist, I was responsible for:

• - Support the Project Manager on management tasks;
• - Perform review of work authored by Linguists;
• - Elaborate and conduct training activities of Linguists and new-hires;
• - Localize prompts, response templates, evaluation instructions, and linguistic resources for Brazilian Portuguese, ensuring linguistic accuracy, cultural appropriateness, and consistency with global AI quality standards;
• - Develop complex grammars for Natural Language Understanding;
• - Create complex templates for Natural Language Generation;
• - Evaluate current system outputs;
• - Annotate and review linguistic data;
• - Follow and fulfill project KPIs.

As an NLU Language Developer, I was responsible for:

• - Developed NLP components: write rules, build grammars/FSTs;
• - Expanded annotation schemas and semantic label structures to support new conversational AI capabilities while maintaining compatibility with existing language resources;
• - Used modeling tools to bootstrap and test new functionalities;
• - Designed Portuguese-specific linguistic mappings and expanded semantic label structures to improve NLU coverage and speech technology performance;
• - Created annotation systems and mappings for Portuguese;
• - Analyzed system performance and identified areas for improvement;
• - Collaborated with other Engineers and annotators in data collection and development efforts;
• - Developed specific guidelines for Brazilian Portuguese phonetic/phonemic transcription;
• - Supported script design used to train speech synthesis models;
• - Worked on localization (en-ptBR-en).

As an Associate Linguist, I was responsible for:

• - NLU Grammar development;
• - NLG development;
• - Developed annotation guidelines, semantic label definitions, and taxonomy documentation to support scalable NLP data creation and model training;
• - Annotate and review linguistic data;
• - Provide linguistic analysis and support for NLU and NLG projects;
• - Response modeling;
• - Structured linguistic annotations using standardized semantic labels for disambiguation, normalization, and language understanding tasks;
• - Annotating lexicon entries according to guidelines;
• - Deriving NLP data.

Part of my responsibilities as a Computational Linguist:

• - Supported the expansion and organization of scientific terminology taxonomies for Portuguese language resources;
• - Data cleaning and wrangling;
• - Flexional paradigm script design.

As a teacher's assistant, I:

• - Supported Professors during lectures;
• - Held office hours;
• - Elaborated and corrected students' assignments.

As an English teacher, I taught Business English courses to professionals in São Paulo.

As an English teacher, I taught English language courses in São Luís, Brazil.

Brazilian Portuguese (BP) has an inventory of seven vowels in stressed syllable /a, ɛ, e, i, ɔ, o, u/. The contrast between lax /ɛ, ɔ/ and tense vowels /e, o/ is lost in unstressed contexts. In southern dialect, reduction favors tense vowels [e, o] (eg. p[ɛ]le ‘skin’; p[e]lada ‘naked’, (underline indicates stress)), and lax vowels are, in general, blocked outside the stressed syllable. However, this is not the case for northern dialects, where lax vowels [ɛ, ɔ] are licensed in pretonic (eg. p[ɛ]lada ‘naked’) and non-final postonic syllables (eg. vesp[ɛ]r[a] ‘eve’), but also alternate with tense (eg. b[e]leza ‘beauty’; corr[e]go ‘stream’) and high vowels (eg. [i]ntrega ‘delivery’; tip[i]co ‘typical’). As noted by Crosswhite (2004), reduction via tensing fits in much better with the theoretical approaches to vowel reduction. In this dissertation, I argue that reduction via laxing in northern dialects of Brazilian Portuguese result from a interaction of two types of neutralization - contrast enhancement and prominence alignment. Throught contraints ranking, it is shown how this pattern is derived while also capturing the standard case of reduction via tensing in southern dialects.

See my PhD dissertation (in Portuguese) here

Grant: Conselho Nacional de Desenvolvimento Científico e Tecnológico. Process n. 140139/2015-0

Previous studies of external sandhi in Brazilian Portuguese have proposed the existence of three categorical process targeting adjacent vowels in word boundaries: elision (car[a] [i]doso ~ car[i]doso ‘old guy’); dipthongation (car[aj]doso) and degemination (caz[a] [a]marela ~ caz[a]marela ‘yellow house’)/ The general picture of what has been claimed is: (i) external sandhi in BP is blocked when there is an intonational prominence in the context; (ii) dipthongation only occurs if at least one of the vowels is high; (iii) in a V1+V2 context, elision occurs only if V1 is [back]; (iv) it is only possible for elision to occur with a front vowel if the vowel is in V1 position and preceded by a coronal fricative (rceb[i] [a]mor ~ *receb[a]mor ‘recieves love’; conhe[si] [a]lguém ~ conhe[sa]lguém ‘knows someone’). This project is analyzing the patterns in a V1+V2+V3 sequence (eg. conhec[i] [a] [i]greja ‘knows the church’) in different dialects of Brazilian Portuguese. I am also investigating if the unexpected patterns attested by our data are better explained through an articulatory phonology perspective.

Brazilian Portuguese has a seven-vowel system in stressed position /a, ɛ, e, i, ɔ, o, u/, which is neutralized in unstressed syllables. In northern dialects, lax-mid [ɛ, ɔ], tense-mid [e, o] and high vowels [i, u] alternate in non-final postonic context. In southern dialects, alternation is only observed between tense-mid [e, o] and high vowels [i, u], as lax-mid vowels are, in general, not licensed outside the stressed syllable. This projected analyzed this alternation in both dialects. I have argued that both lax-mid and high-vowels are fully predictable in non-final postonic context, resulting from height harmony triggered by the vowel in final postonic position.

See my master’s dissertation (in Portuguese) here

Research grant: Conselho Nacional de Desenvolvimento Científico e Tecnológico. Process n. 130194/2013-1

Syncope in BP can be defined as a variable process that deletes the vowel in non-final postonic syllables (eg. [ˈah.vo.ɾɪ] ~ [ˈah.vɾɪ] ‘tree’). Note that the process has a direct effect on stress position: when syncope occurs, stress shifts from the antepenultimate (marked stress pattern in BP) to the penultimate syllable (standard pattern).
At first, the sociolinguistic character of the phenomenon was targeted: syncope in Brazilian Portuguese is stigmatized, viewed as a characteristic of speakers with low level of formal education. However, the process only occurs if deleting the vowel results in well-formed syllable (eg. ˈti.pi.ku ~ *tip.ku, *ti.pku) and this is ultimately what determines its occurrence. More recently, I have analyzed its phonological character: as syncope has a direct effect on syllable stress, I argue that the it is motivated by the reestablishment of the canonical foot of the language, a syllabic trochee at the right edge of the word.