Hi, I'm Irene

Eliza built decision support systems for therapists by helping them to monitor their quality of care and patient progress. Our closest competitors all required that therapy sessions be recorded, and were focused primarily on producing higher-fidelity session notes for patient records and insurance billing. They were in the market, had secured funding, and had existing customer bases - which looked like proof that recording was viable and necessary.

Design partners at Eliza had shared ethical and practical concerns with recording. This feedback was echoed by investors.

Instead of seeing recording sessions as an immutable constraint, I took it as a challenge. I designed a set of features that engaged directly with the therapist in a series of structured journaling prompts. These prompts were designed for the therapist to introspect on the care they had offered.

Through these prompts I collected information on therapist evaluations of patient progress, management of transference and countertransference, and other aspects of the therapeutic relationship that reflected how treatment was being provided. I paired the therapist's perspective with Likert-rated feedback from patients on their own well-being, providing a ground truth for symptomatology and global functioning.

Together these features allowed Eliza to:

• Provide feedback, calibration, and coaching to therapists around challenges in providing care.
• Provide data on individual progress and well-being at the patient and client roster level in the context of clinical decision-making frameworks.
• Generate sufficiently rigorous and detailed progress notes for each session in clinically appropriate language, with special care to preserving the privacy and well-being of the patient. This outcome was especially important — generating timely, reliable progress notes saved clinics tens of thousands of dollars each year.
• Preserve the autonomy and agency of the therapist, provide an appropriate level of oversight, generate cost savings for clinics, while maintaining high-trust relationships between therapists and their clinical managers, and respecting the therapeutic framework and patient privacy.

Twitter's internal platform team was responsible for providing all of the tools and services used by the engineering teams who owned feature production. The organization had evolved a development ecosystem that was struggling under the weight of a chaotic operational structure, fragmented tooling, and ongoing massive organizational change. While I had operationalized quarterly research that directly informed the internal product roadmap - this work was slow relative to on-the-ground challenges. As I was carrying out long-horizon research I often observed behaviors and heard stories that collectively indicated issues that needed more immediate attention in order to facilitate work in progress. Broken internal communication structures prevented most managers from having a complete picture, even when they were working to resolve smaller facets of these issues. Most managers had a stable set of collaborators, and predictable circles of information, but I had a broad range, and could move through the organization with relative ease - both in breadth and organizational depth.

• I began writing quick-turn memos summarizing the nature of the issues that I was learning about on the ground and surfacing them to senior staff in platform services.
• This often prompted follow up questions - essentially "here's what we need to know in order to triage and figure out a solution."
• I carried out time-boxed direct interventions. I'd budget four hours to run down and analyze secondary data (slack messages, internal tool usage, commits, whatever I could find where it existed) and carry out quick turn interviews with ICs who were likely to have insight on any given problem.
• Findings were summarized with recommendations, notes on other known issues or constraints and delivered to senior staff, usually within 48 hours and never more than 72 hours.
• With practice this turned into a dual track research process - I carried out and oversaw long-time horizon research projects, but built a system to deliver ad-hoc quick turn research in parallel.
• In addition to reducing execution risk in day to day production, this research unearthed issues with budgeting and spending, misalignment with organizational objectives, and legal vulnerabilities, and allowed leadership to address these issues early.

I took the Backcountry Climbing Class with the Washington Alpine Club in 2016. We learned to rappel (surprisingly easy), to self arrest in a fall on ice (surprisingly hard), how to build pulley systems and use friction knots to get out of sticky situations. Interspersed with all of this were lectures, where we covered things like basic risk assessment (when it comes to weather big changes mean big problems), and what really happens in a serious field incident. I developed an interest in the way that climbing teams were making decisions in the field, fueled by my own experiences as a climber and, student, and by extensive reading of the North American Accidents in Mountaineering.

• Training included technical skills and offered frameworks for interpreting and decisioning based on field data, but offered no training on how to make decisions under duress.
• Humans are unreliable decision makers at the best of times. Add to that being tired, hungry, potentially injured, and emotionally shocked as the result of an initial critical field incident and decision making capacity gets worse.
• Decision making and emotional regulation under duress is a skill, and it can be taught.
• I identified the lack of this training as a critical skills gap in the training of backcountry athletes.
• I identified and recruited experts, built a coalition of supporters, and designed a training program to be taught by a psychologist who had worked with athletes and first responders, oriented to teaching critical decision making skills - including challenging biases, assumed leadership, and returning to and maintaining an "emotionally online" cognitive state - intervening in default fight or flight responses.
• I advocated for, and secured funding for that training program and the first class was offered Q1 2017 - the first class of it's kind offered by any North American climbing organization.

What really matters is whether someone can navigate their life better, and that comes down to symptoms. I found it really surprising that practicing psychologists are beholden to this ill-fitting framework because of how deeply entrenched it is in insurance billing, while at the same time insurance is struggling to evolve towards value based care - where this framework is actively counterproductive. Of course, the field of psychology is way ahead of me. The Research Domain Criteria (RDoC) initiative was launched by NIMH specifically because diagnostic categories were failing to predict treatment response or biological mechanism. Boschloo et al. (2015) constructed a network of 120 symptoms across 12 DSM-IV diagnoses using data from 34,000 patients and demonstrated that the strategy of summing symptoms into diagnostic categories loses clinically meaningful information. Individual symptoms have different predictors, different relationships to each other, and different responses to treatment.

The problem I wanted to work on is how to build a bridge between the way that our current diagnostic infrastructure functions today based on the DSM, and the apparent conflict between what practitioners, academic researchers, and public health agencies all recognize as superior for understanding and effectively treating mental illness.

In order to build this bridge, I started with the DSM-V. Even if its categorical organization is flawed, the symptom patterns in it are not arbitrary. Researchers and clinicians actually observed these co-occurrences. It's grounded in empirical research and domain expertise that make it a usable proxy dataset for a world in which I didn't have longitudinal data on thousands of patients. Also it's the structure payers and clinicians are currently trapped inside of.

• I decomposed individual DSM-V diagnosis into its symptoms, etiological conditions, associated features, and comorbidities to derive a flat, consistent data structure.
• In parallel I built a symptom dictionary that provided canonical definitions for symptoms across different chapters regardless of how the relevant author committee had worded it.
• The goal was to invert the organizing principle that could be tied back to the current operating framework. Instead of "here's a diagnosis, your patient may have all of or just some of its symptoms", the entry point becomes "here is a set of symptoms, where do they appear across this clinical landscape, what do they reliably co-occur with, and at what developmental stages do they tend to emerge?"
• The approach is deliberately agnostic to which analytical method gets applied. The structured dataset is useful for network analysis (as in Boschloo et al), or for building interpretable treatment planning tools, or for grounding LLM-based clinical decision support in symptoms and the networks they form.
• A symptom-primary data structure makes it possible to build tools that change the clinical question at every stage.