Interviewed by: Mehadi Hasan
Developer: Alice Severi Gonçalves
Role: Independent Data Scientist and Computational Ethologist
Location: São Paulo, Brazil
Featured Project: The Transparency Project — EthoPipe
Website: The Transparency Project
What if the biggest challenge in understanding animal behaviour isn't a lack of observations, but the way those observations are recorded?
A shelter worker might describe a dog as stubborn, aggressive, or guilty. A trainer might interpret a particular posture as evidence of confidence or disobedience. These descriptions may reflect human interpretation rather than measurable biological events. When such judgments become part of research datasets, the distinction between observation and opinion becomes difficult to recover.
For Alice Severi Gonçalves, an independent data scientist and computational ethologist based in São Paulo, Brazil, this problem sits at the intersection of science, software engineering, and animal welfare. Through The Transparency Project, she is developing EthoPipe, a data pipeline designed to transform unstructured canine behaviour observations into standardized, machine-readable records.
Her approach is built around a straightforward principle: scientific data should be reproducible, verifiable, and resistant to subjective interpretation. That principle also shapes how she uses artificial intelligence. Rather than allowing an AI model to make unconstrained judgments, Alice treats it as a tightly controlled component within a larger system of validation, biological boundaries, and auditable data processing.
Her journey into this work has been anything but conventional. It spans legal education, language, digital operations, hands-on experimentation with computers, and years of experience working with dogs. The result is a project that explores an important question for AI-assisted science: how can we make software more useful without allowing its uncertainty to become someone else's risk?
From Exploring Windows 95 to Building Data Systems
Alice's relationship with computers began in the late 1990s, with a beige Compaq Presario running Windows 95. Early exploration involved navigating desktop menus, experimenting with software, and learning through interactive programs.
One experience that particularly influenced her problem-solving approach was playing Zoo Tycoon. Because the original game was in English and localized walkthroughs were unavailable to her in Brazil, she had to work out how the simulation behaved through experimentation.
She changed environmental variables, adjusted exhibit boundaries, observed animal responses, and examined how those changes affected guests and animals. Instead of relying on instructions, she learned by forming hypotheses, changing inputs, and observing outcomes.
That experimental mindset continued through hardware modifications and operating-system testing, including participation in the Windows Insider Program. Long before she began developing software formally, she was learning to investigate complex systems by exploring their behaviour.
Her academic and professional path took several turns. She earned a Bachelor of Laws from UNIP and later spent three years at Elife working with high-volume consumer text auditing and digital customer operations for major retail platforms.
These experiences gave her exposure to structured rules, interpretation, consistency, and the consequences of handling information incorrectly.
Alice eventually recognized a connection between these disciplines and software engineering. Legal reasoning depends on rules, defined boundaries, and the interpretation of exceptional cases. Language follows grammatical structures. Reliable software similarly depends on explicit rules, controlled inputs, and predictable handling of errors.
Her transition into generative AI began in 2024 and early 2025 through a visual branding project for her local pet-care service. She was creating graphic assets based on photographs of her American Pit Bull Terrier, Morgan, and encountered recurring problems with image models producing inaccurate bully-breed anatomy.
Rather than accepting the results, she experimented with cleaner backgrounds, consistent square images, and more precise prompts informed by anatomical and art-historical parameters. One of the resulting designs was selected for the official Adobe Firefly gallery.
The project introduced her to a broader possibility: if AI could help translate detailed instructions into visual outputs, perhaps it could also help translate complex, non-linear ideas into executable Python code and structured data systems.
That realization helped establish the direction of her current work.
Why Animal Behaviour Needs Better Data
Alice's interest in canine behaviour and welfare developed through years of practical involvement with dogs. Since 2012, she has worked in dog walking, pet sitting, and canine facility care in São Paulo.
Through that experience, she saw how behavioural information was often recorded in narrative form. Shelter intake notes, training records, and observational reports frequently relied on subjective descriptions instead of standardized measurements.
Words such as stubborn, spiteful, dominant, and guilty can make a record sound meaningful while leaving important questions unanswered. What exactly did the animal do? What was its posture? How long did it take to recover from a stressful event? What environmental conditions were present?
Different observers may interpret the same behaviour differently. When researchers or organizations attempt to compare those records, inconsistent terminology can make the underlying evidence difficult to reproduce or analyze.
For Alice, this is not merely a data-cleaning problem. It affects how people understand animals and make decisions about their welfare.
Her own perspective on animal training also changed over time. After initially participating in traditional balanced-training circles, she moved toward force-free, positive-reinforcement methodologies by 2020, influenced by what she observed about the behavioural consequences and ethical compromises associated with aversive control.
Quantitative observation offers a way to examine animal behaviour without relying exclusively on human interpretation. Instead of treating a label as a biological fact, researchers can document observable movement, response latency, physiological measurements, and environmental context.
This distinction matters when behavioural assessments influence training decisions, animal-care practices, and welfare policy.
Alice's legal background adds another dimension. She sees standardized, reproducible evidence as a way to strengthen accountability and support evidence-based animal protection standards.
The challenge is to create a system that preserves the complexity of biological behaviour while making the resulting information consistent enough to analyze.
EthoPipe: Turning Observations Into Structured Evidence
EthoPipe is the central technical project within The Transparency Project, Alice's open-science initiative.
Its purpose is to transform unstructured canine behavioural observations into standardized records that can be validated, stored, and shared in formats suitable for scientific analysis.
The project approaches this challenge as a data-engineering problem. Rather than treating a behavioural note as a finished record, the pipeline processes it through a series of validation and normalization steps.
1. Replacing subjective labels with observable actions
Consider a note stating that a dog was stubborn and refused to move.
The word stubborn is an interpretation. It does not describe a directly measurable physical event.
EthoPipe's approach is to identify observable motor patterns, using standardized terminology grounded in ethological literature. Depending on the actual observation, a record might describe a behaviour using terms such as balking or posture_freeze.
The objective is to preserve what can be observed while removing unsupported conclusions about the animal's intentions or emotions.
This distinction makes the resulting data more suitable for comparison across observers and studies.
2. Enforcing strict schemas at ingestion
EthoPipe uses Pydantic v2 to validate incoming data. Alice describes a strict configuration using ConfigDict(strict=True, extra="forbid").
Strict validation helps prevent unexpected types and unrecognized fields from silently entering the system.
Incoming observations are expected to include information such as timestamps, standardized behavioural classifications, environmental context, and physiological measurements.
The principle is to reject invalid or incomplete records at the system boundary rather than allow their problems to spread into later processing stages.
3. Applying biological limits
Software validation alone cannot determine whether a value makes biological sense. A record may have the correct data type and still contain an impossible measurement.
To address this, EthoPipe applies physiological boundaries informed by veterinary references, including the Merck Veterinary Manual.
Alice gives canine heart rate as an example, with a configured range of 30 to 250 beats per minute. A reported value of 400 BPM is flagged and isolated rather than accepted as an ordinary observation.
These boundaries are part of the pipeline's validation policy. They are intended to catch suspicious measurements, not replace veterinary interpretation or clinical judgment.
4. Standardizing records for interoperability
After validation, EthoPipe maps records to Darwin Core, an international standard used to facilitate the sharing of biodiversity information.
The project uses the MeasurementOrFact structure and distinguishes HumanObservation from MachineObservation through dwc:basisOfRecord.
This step matters because a dataset becomes more useful when its structure can be understood beyond the system that created it.
By combining controlled terminology, strict validation, biological boundaries, and standardized metadata, EthoPipe aims to make behavioural observations easier to audit and reuse.
Why Determinism Matters More Than Creative AI
Generative AI is often valued for its flexibility. It can interpret ambiguous requests, produce alternative explanations, and generate varied responses to similar prompts.
In scientific data processing, however, that flexibility can become a liability.
Alice's definition of determinism is simple: for an unchanged input state, the pipeline should produce the same structured output.
If an animal shelter processes an intake note on Monday and processes the same note again on Friday, the resulting normalized record should not change simply because the model produced a different interpretation.
This requirement shapes how EthoPipe uses large language models.
The pipeline interfaces with foundation-model APIs, with Alice's responses describing a migration from Gemini 1.5/2.5 Pro to Gemini 3.5 Flash. The model's generative temperature is set to zero, and structured JSON schemas are enforced through the SDK.
The model is not intended to act as an independent behavioural authority. Instead, it functions as a constrained text-normalization component within a system that applies explicit rules and validation.
Temperature zero alone does not guarantee perfect reproducibility across all model versions, API changes, or service configurations. Alice's broader engineering approach therefore places importance on schema enforcement, tests, documented changes, and validation outside the model itself.
The distinction is important: predictable formatting is not the same as scientifically correct interpretation. A structured record can still be wrong, which is why the rest of the pipeline must be able to detect and isolate invalid data.
For Alice, deterministic engineering is a way to reduce uncertainty at the points where software can control it.
Designing for Reliability as an Independent Developer
Maintaining a complex data pipeline as an independent developer creates a different set of constraints from working within a large engineering team.
Alice has responded by imposing strict architectural limits on EthoPipe.
The project maintains a Python-only runtime and a maximum budget of eight runtime packages. Its current dependency budget is reported as eight out of eight. It also avoids external broker daemons such as Redis and Celery, keeping state transitions within a leaner architecture.
These decisions reflect a preference for minimizing operational overhead and limiting the number of moving parts that must be maintained.
Dependencies can introduce maintenance burdens, compatibility problems, and potential supply-chain risks. Background services can also add deployment and debugging complexity. For a solo maintainer, every additional component represents something that must be understood, tested, and kept operational.
The project also emphasizes automated testing. Alice reports 27 passing tests and at least 90% code coverage.
Those figures are useful indicators of the project's current testing discipline, although passing tests and code coverage do not, by themselves, establish scientific validity or guarantee the absence of defects.
The larger goal is to make failures visible and manageable.
Alice identifies three particularly concerning failure modes:
- Anthropomorphic contamination: Human judgments enter the dataset as though they were objective biological observations.
- Silent schema drift: Changes to upstream APIs or incoming data formats cause fields to become misaligned without an obvious error.
- Physiological boundary violations: Biologically implausible measurements are accepted as valid records.
Her approach combines strict Pydantic validation, hardcoded biological thresholds informed by veterinary references, immutable audit records, and quarantine mechanisms for malformed payloads.
Rather than silently accepting questionable data, the system is designed to make the problem visible and preserve the record for manual review.
This is a broader engineering lesson from EthoPipe: reliable systems should not merely process valid inputs efficiently. They should also fail clearly when an input cannot be trusted.
Where AI Can Help Animal Welfare—and Where It Must Stop
Alice sees several promising applications for AI in animal welfare, particularly when the technology is used to organize information that would otherwise be difficult to analyze at scale.
One example is extracting structured information from shelter records. Frontline staff and veterinary professionals may document observations in short, informal notes rather than complete extensive database forms. A carefully constrained language model could help transform those notes into standardized records for subsequent review.
Computer vision offers another opportunity. Markerless pose-estimation frameworks such as DeepLabCut and SLEAP can track body landmarks across video footage. Those coordinates can support calculations involving movement, gait asymmetry, and other kinematic characteristics.
Bioacoustic analysis may also help identify patterns in kennel audio, including vocalizations that warrant closer attention from animal-care staff.
These applications share a common principle: AI can help identify, organize, or measure information, but its outputs must remain subject to appropriate validation.
Alice is equally clear about where automated systems should not be granted independent authority.
Behavioural assessments and life-or-death decisions
Animal shelter decisions can be affected by historical bias, inconsistent records, and breed-related stereotypes. An automated system that labels an animal aggressive or unadoptable based on poorly calibrated notes could amplify those problems.
For that reason, Alice argues against allowing an unmonitored algorithm to determine an animal's fate.
Physical interventions
She also rejects connecting autonomous AI agents directly to aversive training devices, electric fences, or physical containment systems.
An algorithm should not independently trigger interventions that affect an animal's bodily integrity or physical freedom.
Veterinary diagnosis and treatment
AI-generated veterinary recommendations require particular caution. Alice argues that generative systems should not prescribe medication dosages or direct veterinary interventions without deterministic validation boundaries and approval from a qualified veterinary professional.
These limits are consistent with the architecture behind EthoPipe. The more consequential the decision, the less appropriate it is to rely on an unconstrained model output.
AI can assist with data processing, but assistance should not be confused with accountability.
Building With AI Without Giving Up Engineering Judgment
Alice's use of AI extends beyond EthoPipe's data pipeline. She also uses AI tools as part of her own development workflow.
As a self-taught developer working across several disciplines, she has found that AI can help organize complex ideas, interpret technical documentation, and explore architectural alternatives.
She describes using AI as an active technical sparring partner: challenging assumptions, discussing design choices, and translating complicated material into more manageable mental models.
Her workflow also involves experimenting in controlled environments, building command-line tools, and examining error logs to understand how systems behave when something breaks.
This approach reflects the same mindset that informs her data pipeline. Rather than treating an AI-generated answer as authoritative, she uses it as an input to a process of investigation.
That distinction is particularly relevant for developers working with rapidly changing AI tools. A generated solution may be useful, but it still needs to be understood, tested, and evaluated against the system's actual requirements.
Alice also speaks candidly about the personal difficulty of navigating technology work without a conventional software-engineering background. She describes ongoing imposter syndrome, uncertainty about employability, and the challenge of keeping pace with changing dependencies and model versions.
Her experience suggests that technical progress is not always a straight line from formal education to employment. It can involve combining knowledge from different disciplines, learning through experimentation, and building systems while still questioning whether one's skills are sufficient.
The engineering challenge is to turn that learning into software that remains maintainable when the tools and assumptions around it change.
Learning Cloud Architecture Without Prematurely Moving to the Cloud
Alice's interest in infrastructure also developed through formal learning opportunities.
As a fellow in the 2026 Google Cloud Launchpad Program with Mentor Me Collective, she explored concepts including VPC networking, identity and access management, containerization, Pub/Sub message queues, and Cloud Run architectures.
She also applied concepts from a full-stack development sprint at EBAC to her work on EthoPipe.
Rather than building disconnected exercises for every new technology, she adopted what she calls the "One-App Rule": channeling coursework concepts into the architecture of one ongoing project.
This approach allows new ideas to be evaluated in the context of an actual engineering problem.
Cloud concepts helped her think about component boundaries, stateless interfaces, and future scaling requirements. However, EthoPipe itself has not yet been deployed to cloud infrastructure.
That distinction reflects an intentional choice. Learning how to design for cloud environments does not mean every application needs an immediate cloud deployment.
For a project still undergoing architectural hardening, adding infrastructure prematurely could introduce complexity before the core data-processing workflow is sufficiently stable.
Alice's current direction emphasizes a lean local application while preparing its architecture for future requirements.
The goal is to understand the available tools without allowing infrastructure decisions to distract from the underlying scientific problem.
The Hardest Lesson: Owning the System
For Alice, one of the most difficult parts of developing EthoPipe has been learning to take full ownership of the system.
Model deprecations, changing dependencies, and architectural complexity have forced her to reconsider implementation choices while preserving the project's reliability.
Migrating the extraction pipeline to a newer model, keeping the runtime dependency budget limited, and maintaining the existing test suite have all contributed to her understanding of what sustainable engineering requires.
Her conclusion is that engineering competence cannot be measured simply by how much syntax someone remembers or whether their educational background follows a conventional route.
It also depends on whether they can understand a system's constraints, make deliberate trade-offs, test their decisions, and respond to change without allowing the architecture to collapse under its own complexity.
That lesson extends beyond independent development. Every software system eventually encounters changing requirements, aging dependencies, and unexpected failures. The difference lies in whether the system has been designed to expose those problems and recover from them.
Alice's approach is to document the process rather than hide the uncertainty.
The Transparency Project records architectural milestones, test failures, and design decisions in public project logs. Its documentation and scientific lineage are part of the effort to make the work auditable.
The project has also established a permanent Zenodo DOI, 10.5281/zenodo.21211371, and maintains a documented scientific identity through persistent identifiers.
EthoPipe is currently in Phase 4, focused on journal publication preparation, Zenodo minting, and architectural hardening. Alice reports that the project has received open-source documentation sponsorship from GitBook.
She remains candid about its uncertain future. She does not know whether EthoPipe will ultimately work as intended or gain wider adoption in the research community.
But she believes that documenting the work is valuable regardless of the outcome. Even if the project does not achieve its intended reach, a transparent record of the experiments, decisions, and limitations can still contribute to open science.
What Comes Next for EthoPipe
Alice's current development priorities focus on strengthening the data layer and preparing the project for more demanding workloads.
Her plans include:
- Localized storage adapters: Developing dedicated adapters to persist normalized records into a query-optimized Kimball Star Schema.
- Data-layer hardening: Improving Supabase PostgreSQL real-time replication endpoints and binding direct Google AI Studio response schemas at the ingestion gateway.
- Stress testing: Building automated end-to-end tests for scenarios involving multiple dogs.
- Scientific publication: Preparing the project's
paper.md and paper.bib files for formal submission to the Journal of Open Source Software.
Longer term, she wants to automate Darwin Core Archive exports, connect Zenodo citations with international biodiversity repositories, and deploy a public telemetry portal for The Transparency Project.
These goals remain part of the project's ongoing development rather than completed outcomes.
The intended direction is clear: make it easier to collect, validate, preserve, and share structured evidence about animal behaviour.
If the infrastructure works as intended, standardized data could help researchers compare observations more consistently, evaluate interventions more objectively, and support animal welfare discussions with stronger empirical evidence.
Achieving that goal will require more than AI alone. It will depend on sound data standards, careful validation, reproducible methods, and appropriate human oversight.
Data Over Dogma
EthoPipe brings together several strands of Alice Severi Gonçalves's experience: the rule-based thinking of legal training, the structure of language, the experimental mindset of early computer exploration, and years of practical involvement with dogs.
Its central idea is that software can help separate observable evidence from subjective interpretation, provided the system is designed to enforce clear boundaries and expose uncertainty rather than conceal it.
The project is still developing, and its wider scientific impact remains uncertain. Yet its engineering priorities offer a useful perspective on responsible AI development: use models where they can help, constrain them where consistency matters, and never confuse a well-formed output with a verified fact.
For developers working on scientific, healthcare, or other consequential data systems, that distinction is especially important. Reliable automation is not simply about getting a model to produce an answer. It is about building the surrounding infrastructure that determines whether the answer can be trusted, reproduced, reviewed, and corrected.
Alice's work is a reminder that some of the most important technology projects begin with an unglamorous problem: information is being collected, but the way it is recorded makes it difficult to trust.
Her response is to build the structure that the data is missing.
Data over dogma.
About Alice Severi Gonçalves
Alice Severi Gonçalves is an independent data scientist and computational ethologist based in São Paulo, Brazil. Her work focuses on canine behaviour, quantitative animal welfare science, deterministic data processing, and reproducible research infrastructure.
She is developing EthoPipe through The Transparency Project, an open-science initiative that aims to transform subjective canine behavioural records into standardized, auditable datasets. Her technical interests include Python, data engineering, AI-assisted text normalization, strict schema validation, biodiversity data standards, and cloud architecture.
Developer Stories is a CoderLegion interview series highlighting the people, ideas, projects, and engineering decisions behind software development.
