Asaf WishniaSenior AI Product Manager

I build AI productsfor high-stakes work.

0 to 1 LLM and agentic products across B2C, B2B and B2B2C. Deepest in healthcare, proven in SaaS and subscription commerce.

Let's talk

Former CPO and co-founder at two AI startups
Published AI researcher
Clinical mental health background
Harvard Medical School Executive Education

A finding in the AI reviewer: the sentence, the rules cited and replacement copy. Sample data.

Order check

verdict APPROVED
checked 4 systems

1,570monthly active users

+34%conversion in the three months after the order check launched, shared with marketing work

$4,500saved every month across all the automations I built

~55team hours saved every month across all automations, from run logs and team reports

~60%less time on targeted processes

+20%productivity in a 50-employee B2B SaaS pilot, within three months

90%primary-diagnosis agreement in a pediatric AI voice intake, interim cohorts (κ=.88)

7+startups advised on AI strategy, positioning and go-to-market

Selected work

01 · FDA and FTC review

An AI reviewer that has to prove every finding

It reads a whole health website, cites the rule behind each risk, checks every quote against the live page, and leaves the decision to people.

Agentic pipelineEvalsHuman sign-off

Try the checker

Overview of the review tool: findings by severity, coverage and sign-off. Sample data.

02 · Subscription order check

Every subscription order, checked before it ships

Research into why subscribers left pointed to wrong orders and slow answers. I built agents that check every subscription order across four systems before it ships, and pass people only the ones that need them.

In productionMulti-agentDiscovery to impact

See how it decides

After launch

1.5 daysaverage ticket handling, down from 5

−90%bug rate on subscription orders

−20%churn, once it leveled off

03 · Patient app

A patient app, taken from 0 to 1

A subscription care platform for people living with inflammatory bowel disease. All 28 screens are live on this site, ready to click through.

0 to 1Interactive prototypeSample data

Open the case

04 · Patient data

Identity behind a vault, answers leadership can read

One flat export mixed names with clinical answers. I specified a model that splits them, and a dashboard with no route to a name.

PRDData modelPrivacy by design

See the model

Identity vaultname, contact, key mapAudit-logged

Patientspseudonymous IDNo identity

Submissionsprotocol week, sourceNo identity

Answersone row per answerNo identity

Leadership dashboardde-identified marts onlyRead only

More automation: morning brief, QA and AI adoption

Research and clinical background

Clinical roots.Published AI research.

Research and background

First author on an LLM diagnostic study. Evaluating ChatGPT's Diagnostic Capabilities for Mental Health Disorders, 2024.
90% primary-diagnosis agreement. Pediatric AI voice intake, interim cohorts (κ=.88).
Six years of research and clinical work. With children and adolescents.
M.A. candidate in Clinical Psychology. And Harvard Medical School Executive Education, AI in Health Care.

How I work

Research first. Then I build it.

Discovery

Data shows where users drop. Their words show why.

Strategy

Roadmaps are trade-offs. I make the cuts explicit.

Translation

Clinicians, patients, data. I translate them into KPIs.

Build

Research, spec, PRD. Then I build it hands-on.

Evals

Clinical logic sets the rules. Evals enforce them.

Agents

Agents take the repeat work. People keep the judgment.

Experience

Healthcare is my depth.The range is wider.

Senior AI Product ManagerB2C digital health and subscription e-commerce in the US. A patient platform from 0 to 1, AI agents across operations, and a real-world evidence program scoped with a pharma partner.

Chief Product OfficerAI platform for therapists and their patients: follow-ups, session transcription and summaries. Led the technical and research team and built the MVP.

Co-founder and CPOB2B SaaS analytics on workplace data. Signed design partners and ran a 50-employee pilot: +20% productivity in three months and a 15% drop in retention-risk indicators. Led roadmap, integrations and go-to-market.

AI Optimization and Strategy LeadConversational AI for psychiatric assessment, built with psychiatrists, patients and researchers. Designed the clinical logic, scoring, guardrails and evaluation flows, and ran a pilot with a government ministry.

Advisor to 7+ startupsAI strategy, positioning and go-to-market

AI researcher in psychiatryLLM diagnostic accuracy. Published, and presented in Toronto and Porto.

Let's talk.

Senior or Lead AI Product Manager

AI-first products and AI agents

Vertical or domain PM

Healthcare and other high-trust industries

Head of Product

Early-stage AI startups

Message me on LinkedIn Open Claude skills kit

Asaf Wishnia AIPMWork is shown without company names, with sample data. Employers are on my CV and LinkedIn.

Case study 03

A patient app,taken from 0 to 1.

A subscription care platform for people living with inflammatory bowel disease: a monthly check-in, a personal plan and a care team. I took it from discovery to this interactive prototype.

28 screens. On a phone it opens full screen. No sign-up.

Role: Senior AI Product Manager, led 0 to 1
Stage: Interactive prototype, sample data
Domain: Digital health, subscription
How it was made: Research, spec and PRD, then code with AI

Every screen

Pick any screen and start there.

The prototype opens on the screen you choose. From there you can move through it as a patient would, or jump to any of the 28 from the screen list.

Decisions

What the screens are arguing for.

Process

How it got here.

Discovery

Listen on both sides

Recurring discovery with clinicians at a US academic hospital
A mixed-methods study of about 70 low-engagement users

Strategy

Cut in writing

Streaks and badges cut: the wrong incentive for ongoing care
The full conversational AI guide parked for a later phase

Spec

Tie it to a number

PRD and spec, with a metric on every active item
KPIs that healthcare and pharma stakeholders can track

Build

Give teams something to click

I built the prototype in a code editor with an AI assistant
I reviewed every change

Presented without company names. Products, prices, people and terms in the prototype are samples.

Next · 04 · Patient dataIdentity behind a vault, answers leadership can readContinue

Asaf Wishnia AIPMWork is shown without company names, with sample data. Employers are on my CV and LinkedIn.

Case study 01

An AI reviewer that hasto prove every finding.

Health marketing lives under FDA and FTC rules, and one sentence can make a supplement read as a drug. I built an agent that reviews a whole website, proves each finding against the live page and leaves the decision to people.

Every view works: overview, findings, sign-off, Train the AI. Sample site.

Role: Product and build, end to end
Runs as: Claude Code plugin, one command
Output: A report the team signs off on
Used by: The team and marketing. Legal counsel has the last word.

Try it

Pick a sentence. Be the reviewer.

These sentences are invented for this demo. The rules and citations are the real ones the tool uses.

Or test your own copy

Here, a small keyword pass stands in for the AI reviewer. Your text stays in your browser.

Choose a sentence on the left to run the review.

How it decides

One page, through the whole pipeline.

People stay in charge

Every "no" teaches it.A person confirms.

Each number on the overview opens the items behind it. Each decline needs a reason, and the AI turns that reason into a lesson the team can check.

Decline needs a reason. The AI proposes a lesson and shows what else it would settle.
A person confirms the lesson. Until then, what it would settle stays on hold.
High-risk findings are guarded. Only a "no claim in context" lesson can settle disease, disclosure, FDA status, endorsement or review findings.
Rewrites lift approvals. Change the sentence and the old sign-off no longer counts.
Real people's words can only be cut. The tool never writes new words into a quote (16 CFR 255.1(b)).
Replacements may not add facts. And text shared across pages gets one fix.

Evidence

How I know it holds.

40+rules from 21 CFR 101.93, 16 CFR 255 and 465, and the FTC Health Products Compliance Guidance

42automated checks on the tool itself

0findings wrongly settled, across two eval rounds on 10 real decline scenarios, graded from a legal and a product view

A finding in the review tool, with the rules cited. Site text redacted. — The finding view. Site text redacted.

Replacement copy with sign-off from product and marketing. Site text redacted. — Replacement copy and two sign-offs. Site text redacted.

A first-pass review tool. Legal counsel signs off.

Next · 02 · Subscription order checkEvery subscription order, checked before it shipsContinue

Asaf Wishnia AIPMWork is shown without company names, with sample data. Employers are on my CV and LinkedIn.

Case study 04

Patient data leadershipcan use, and compliance can sign.

Questionnaire data sat in one flat export with more than a hundred columns. Identity was mixed into clinical answers, labels were cryptic and versions were stacked. Every question became an analyst request. I wrote the PRD and the data model that fix it.

Thirteen sections, four data mockups. Names removed, synthetic data.

Role: PRD and data model, with a clinical expert
Stage: Specified for review. No build shown here.
Domain: Patient-reported outcomes
Audience: Leadership, clinical review, compliance

Step one

Before modeling, classify everything.

Every column gets one of four labels, and compliance signs the list. These columns are samples of the kinds of problem in the real export.

Flat export

Identifier

Moves to the vault

Quasi-identifier

Coarsened

Clinical

Becomes answers

Junk

Resolved before migration

The model

One lock. Three ways in.

Identity exists in one place. Pick who is asking and see what they can reach.

Asks

Gets

Control

The best argument in the project

The clinician wanted one line per patient.The model needed one row per answer.

Both were right, at different layers. Storage is long, because only that shape survives new questionnaire versions. The clinical view is generated wide, aligned by protocol week.

Sample patient. A missing value stays unknown and is never stored as zero.

Decisions in writing

Positions taken, with the reason.

Reversible pseudonymization, with an audit-logged vaultFollow-ups arrive by phone and email and must attach to the same patient. Irreversible de-identification would cut that link.

Store long, present wideEvery submission maps to its nearest protocol week: baseline, W4, W8, W12, W24.

Kilograms as the one unitThe entered value and unit are kept for audit.

Human-readable question labels"Bowel movements per day" in place of "SCCAI 1". The dictionary carries both.

Current and prior medications are separateConcatenated strings are parsed. What cannot be parsed goes to a review queue and is never silently dropped.

Clinical-trials field retainedUnused today. It keeps a future research collaboration possible.

Follow-up outcomes are first-class dataEach carries its source and protocol week, wherever it was captured.

Dashboard for leadership only, at firstThe permission model is role-shaped from day one, so access can widen later.

What leadership gets

"What's our week-8 remission rate by diagnosis?"

Answered in under a minute, unaided. That is the bar in the PRD. Cohorts under five patients are suppressed.

Response by diagnosis

Invented numbers, to show the view. Not results of any product.

Targets

How I would know it worked.

100%of columns classified, with compliance sign-off

≥99.5%cell-level match between migrated data and the original export

≥80%of leadership using it weekly within 30 days

70%fewer ad-hoc analyst requests from leadership within 60 days

<5 daysto onboard a new questionnaire version

0identity fields returnable from any table a dashboard credential can query

Out of scope, on purpose

Patient-facing assessment redesign
Real-time analytics
Predictive models, before the data is validated
External data sharing

Sequence

Audit Classify every column, ratify the privacy approach
Foundation Schema, vault, question dictionary
Migrate Repeatable pipeline, validated against the original
Dashboard Built only on the validated marts

Next · 05 · Automation and adoptionAgents for the repeat work, and a company that uses themContinue

Asaf Wishnia AIPMWork is shown without company names, with sample data. Employers are on my CV and LinkedIn.

Case study 05

Agents take the repeat work.People keep the judgment.

Inside a subscription health company I built AI agents across Shopify, Zendesk and internal tools, then led adoption across the company and taught the teams to build with AI.

Subscription order check, in depth

Result: $4,500 and ~55 team hours saved every month across all automations, ~60% less time on targeted processes
Tools: Claude, Shopify, Recharge, Zendesk, Monday, Python, Playwright
Pattern: Clear cases pass. Unclear cases go to a person.
Scope: Operations, support, marketing, leadership

Subscription order check

Four systems, three verdicts, every morning.

Each subscription order depends on a clinical program of six protocol tiers. Checking one by hand meant ten steps, one subscriber at a time. I built agents that go order by order, and they ran in production.

ShopifyProduct matches the subscription

RechargeShipping dates line up

Clinical databaseFits the program stage and monthly questionnaires

ZendeskTickets and requests from this customer

VerdictApproved, Review, or Bug: hold, Monday board, CS ticket

See how it decides

Morning brief

Agents feed an orchestratorthat writes a morning brief.

Scheduled agents read the store, the task boards and email. Analysts for commerce, support and marketing report to a master agent, and leadership opens one page.

Today's recommendations

Three decisions worth making before lunch.

Support

Commerce

Marketing

Anomalies

Structure of the real brief. Contents removed.

Also built

Same pattern, other teams.

Clinical QA

End-to-end questionnaire testing

Python and Playwright, across regions over VPN
Checks that each profile gets the right recommendation
HTML and PDF report, one entry script, a team runbook

Customer support

A grounded shipping assistant

Answers cite the team's own knowledge
Asks a clarifying question when it needs one
When the answer isn't known, it says so and routes to a lead
Local prototype, with a daily activity report

AI readiness

Help center cleanup

Medical claims removed from the articles an AI sales agent reads
So the agent starts from compliant text

AI adoption

Adoption happens one role at a time.

I led company-wide adoption and taught the team to build with AI.

Show

Interactive training deck

In Hebrew, for non-technical teams
Real internal workflows as the worked examples

Teach

Self-paced Academy

Bilingual, 10 chapters
Hands-on simulators, 50 questions
80% pass mark, certificate

Embed

Role playbooks

Daily use cases for seven teams
Patient-data safety built into the training

Scale

Shared skills

Reusable agents and skills
An open skills kit on GitHub

Next · AboutClinical roots. Published AI research.Continue

Asaf Wishnia AIPMWork is shown without company names, with sample data. Employers are on my CV and LinkedIn.

Case study 02

Every subscription order,checked before it ships.

Subscribers were sliding back to one-time orders. My research traced it to wrong orders and slow answers from support. I built agents that check every subscription order across four systems each morning, weigh the evidence, and pass people only the orders that need them.

1.5 daysaverage ticket handling time, down from 5

−90%bug rate on subscription orders

−20%churn in the three months after it leveled off

+34%conversion over three months, shared with marketing work

Role: Ran the discovery, designed and built the agents
Stage: In production, every morning at 7:00
Systems: Shopify, Recharge, internal clinical database, Zendesk, Monday
Output: Approved, Review or Bug, for every order

Why I built it

Subscribers were sliding backto one-time orders.

Here a subscription follows a clinical protocol, so every lapse cost twice: the customer left the program, and the treatment data lost its continuity. Most of it happened in the first weeks.

44%Month 124%Month 232%Later

68%of cancellations came within the first two months

1.1 monthsmedian time from joining to the first attempt to cancel

Share of cancellations by tenure, from exit surveys and support email. Estimates from the tracking dashboard.

Discovery

I read every channel wherecustomers talked to us.

Calls, chats, tickets, Reddit and exit surveys, classified by meaning rather than by keyword and tracked every two weeks. I wanted the reason behind each cancellation, in the customer's own terms.

Support callsHuman agents, transcribed
AI support callsTranscripts
Chat and emailCustomer and support threads
TicketsGathered by a second automation I built
RedditA crawler on the relevant communities
Exit surveysVoluntary, from cancelling customers
User researchSessions with subscribers

ClassifyBy meaning, not keywordsSemantic analysis maps each conversation to a reason.

TrackA trend dashboard, every two weeksReasons over time, by tenure and by channel.

26%of stated reasons were the subscription and service experience: wrong or missing items, signing up by accident, no way to adjust the order. The largest group.

31%of stated reasons were in the company's own hands: the subscription experience, support and delivery.

1 in 3answers marked "Other" turned out to be fixable problems once they were read for meaning.

The team's side

The same problem, seen from support.

The Head of CS brought me what the team was living with. Requests to change or fix an order came in faster than the team could reach and prioritize them. Customers felt it as slow answers, and trust eroded.

Percentages are of cancellations with a stated reason, from exit surveys and support email. Counts are withheld.

The decision

Catch the wrong orderbefore it ships.

Wrong orders created tickets, and the tickets waited in a queue the team could not keep up with. Fixing orders after a complaint kept both problems alive. So I moved the check upstream, to every subscription order, every morning.

Before: ten manual steps for each subscriber

1Pick today's subscriberSchedule
2Filter by subscription tagStore
3Copy the customer's emailStore
4Open the patient recordClinical record
5Read the plan and its dateClinical record
6Verify the expected productsStore
7Does it match? If not, flag the gapDecision
8Preview the next orderClinical record
9Adjust the cartStore
10Wait a day, verify againWait

How it runs

Every morning at 7:00.

The cutoff is set against worldwide Shopify orders. From there the agents go order by order, and people sign off on the result.

7:00 cutoffToday's subscription orders, worldwide

CollectOne agent per system: Shopify, Recharge, the clinical database, Zendesk

DebateAn orchestrator weighs the evidence against written rules

VerdictApproved, Review or Bug

Sign-offSupport checks against tickets. Clinical sign-off where it is needed

The debate

Every source makes its case.The rules decide.

One agent per system pulls the evidence on the order, and an orchestrator weighs it. The questionnaire and the phase plan in the clinical database are the north star. A specific request from the customer, by ticket or by phone, can change the answer within written rules that several teams agreed on. Most conflicts end there. Whatever does not becomes Review, and a person decides.

Demo. No action in real systems.

ShopifyWhat is in the order

RechargeWhen it ships

Verdict

Clinical databaseThe north star: the protocol

ZendeskTickets and calls from the customer

Sample orders, invented to show each path. The real weights and rule set stay internal.

Verdicts

Three outcomes, each with a next step.

Approved

Ready to ship

The sources agree. Support checks it against the tickets, and it ships. First orders also get a clinical sign-off.

Review

A person decides

The fallback whenever no rule settles the case. The order is held, and the agents hand it over with the evidence. They never guess.

Bug

Stopped

Puts the order on hold in Shopify
Adds the issue to the Monday board, in full detail
Opens a ticket for CS

Bug · External

The cause is outside the system

Support did not update an order, or the required first questionnaire was not filled in.

Bug · Internal

The cause is technical

A failure in the clinical database, in cookies, in Shopify or in an API, or in the handoff logic between platforms.

The clinical check

An order can look rightand still be wrong.

A questionnaire maps each patient to one of six protocol tiers, and each plan moves in phases. The internal database holds both. These are the rules an order has to satisfy.

The newest signal leads. Whichever is more recent, the assessment or the last order. A tie goes to the assessment.
One tier per step. A patient moves at most one tier from the last order, whatever the assessment says.
A tier change restarts the plan. The patient goes back to phase 1 of the new plan.
Progress is supply delivered. A patient who skipped orders has not advanced, whatever the calendar says.
Orders one and two ship base products. The patient's own plan starts at order three.
Six months without an order resets the patient. They go back to base products until the relationship resumes.
Changes take about a day to land. So nothing gets flagged twice in the same morning.
The first questionnaire is required. No product ships without it. Later questionnaires are optional: when one is missing, the phase plan is the fallback and a person confirms.
Nothing is guessed. Missing or unreadable data is never filled in by the model.

Reliability and access

When it is unsure, it failstoward a person.

Review and Bug orders stay on hold until someone on the team acts. Access is split by role, and each system is reached the way security agreed to.

Reliability

Review is the fallback. When no rule settles a case, a person decides. The agents never fill a gap with a guess.
A daily alert on the bug rate. It fires when Bugs pass 10% of the day's orders. I started it at 20% and lowered it after debugging, to sit closer to what customers feel.
Safety guardrails escalate. A child under 8 on an adult protocol, or a pregnancy contraindication, puts the order on hold and raises a double flag in Monday and in Zendesk, with an active alert to support, the clinical team lead and me.
People sign off where it matters. Support checks orders against the tickets. Clinical support and the Head of CS give final sign-off on first orders, Review cases, unusual clinical requests, and any Bug that could have led to a risky order.

Access and security

Built on Claude, with access by role. Full control through Claude Code and an internal interface on Cloudflare, only for me, my product team, the Head of CS and the COO.
Daily work in one shared screen. Support works in an artifact in the company's Claude workspace: customers, problem orders, final approval and releasing holds, with permissions set per role.
Each system through the right door. Shopify through its CLI. The clinical database through an API I built with the engineering team. Recharge and Zendesk, which we don't control, through Claude in Chrome, because we didn't allow crawlers on third-party systems.
Only the data a case needs. Records are linked by email and username inside internal systems. Monday items and tickets carry an order or ticket number, and a customer email only when a case can't be solved without it. The processing was approved by the CISO after many conversations with the CISO and Legal.
Built end to end. Discovery, design, integrations, agents and deployment.

What broke once

A three-day drop, traced to one state.

California

Segmented by region. One state stood out.

Orders dropped sharply and steadily for three days, and the funnel showed less engagement with questionnaires. Logins looked normal: people could enter the interface but could not act in it.

Segment first. Before I look for an internal bug, I split the drop by subgroup in our logs and monitoring. It was only California.
The cause. A change in how cookies could be used there blocked actions in our main interface.
The fix, on two fronts. With my product team I changed the cookie consent flow and explained it on the first landing page, and marketing sent California customers a short explanation.

Looking back

If I built it again, I would start with these.

The alert at 10% from day one. Not 20%, lowered later.
A set of orders labelled by people before launch. So accuracy is measured from the first run.
Narrower guardrails from the start. The ones I tightened while it was already running.

What changed

Faster answers, fewer wrong orders,and the churn turned.

I tracked the same dashboards before and after launch, over months.

Ticket handling time1.5 days

Before

5 days

After

1.5 days

Average time to full resolution, 70% faster.

Bug rate−90%

Before

100

After

Share of subscription orders flagged as Bug. Before launch = 100.

Starting on a subscription23.2%

Before

1.3%

After

23.2%

New customers who start directly on a subscription, in the months before and after launch.

Conversion+34%

Over the three months after launch. Marketing changes ran in the same period, so I don't credit this to the order check alone.

ChurnBefore launch, drawn for shapeMeasured after launch

Leveled off within 6 weeks −20% in 3 months

BeforeLaunch+6 weeks+4.5 months

Churn index, where 100 is the level at which it leveled off.

Lapsed subscribers back20%

Month 1

Month 2

13%

Month 3

20%

Lapsed subscribers who renewed, by month after launch.

The manual check moved to the agents in stages. Support now starts the morning from the verdict list, and Review and Bug get the team's time.
About ten people worked from it. Support, clinical, marketing, leadership and me.
Trust came back. Within about three months, customers saw close, accurate service and clinical results, and more of them chose the subscription from the start.

How to read these numbers

What each one measures.

Handling time. Average time to full resolution of a ticket. Resolved means closed, with two days and no new request after it. Repeat contacts on the same order were tracked for the following week.
Bug rate. Share of subscription orders flagged as Bug, external and internal together, against the level before launch.
Churn. Subscription cancellations, shown as an index where 100 is the level at which the rise stopped.
Starting on a subscription. Share of new customers whose first order is a subscription.
Lapsed subscribers back. Share of lapsed subscribers who renewed, by month after launch.
Conversion. Change over the three months after launch. Marketing work ran in the same period, so the credit is shared.

All figures come from the tracking dashboards and cover subscription orders and their customers. The $4,500 and ~55 team hours a month on the home page are the total across all the automations I built, this one included.

What comes next

Stop the wrong order at creation.

The agents check orders once they exist. The next step I specified moves the check to the moment an order is created: store rules for the purchase sequence, and a backend hook for the clinical rules, with coded reasons in place of free-text notes. Every rule ends in a hold, never a cancellation, and safety rules fail closed. It was written up for approval and a shadow-mode run.

100%of the test matrix must pass: every rule, by positive, negative and edge case, by region

≥95%precision against manual holds during at least a week of shadow mode

0safety misses allowed before a rule can enforce

Gates and targets from the specification.

Next · 03 · Patient appA patient app, taken from 0 to 1Continue

Asaf Wishnia AIPMWork is shown without company names, with sample data. Employers are on my CV and LinkedIn.

About

Clinical roots.Published AI research.Product leadership.

I started in research and clinical work with children and adolescents, moved into AI research in psychiatry, and now build AI products. That path is why I care about what happens when a model is wrong.

Product leadership

B2C, B2B and B2B2C.

A consumer subscription health brand, a B2B SaaS startup I co-founded, and AI platforms that reach patients through therapists and hospitals.

Asaf Wishnia at ARC Summit 2024 — ARC Summit, 2024

Chief Product OfficerAI platform for therapists and their patients: follow-ups, session transcription and summaries. Led the technical and research team and built the MVP.

Advisor to 7+ startupsAI strategy, positioning and go-to-market

AI researcher in psychiatryLLM diagnostic accuracy. Published, and presented in Toronto and Porto.

Research

"We need to remember that AI is not magic."Interviewed in Traumatech Nation by Maayan Hoffman. Chapter 4: AI for Mental Health.

Evaluating ChatGPT's Diagnostic Capabilities for Mental Health DisordersFirst author. Journal article, 2024.doi:10.31579/2835-7957/113

Evaluating the Diagnostic Accuracy of ChatGPT in Pediatric Psychiatry: A Comparative StudyCo-author. Conference abstract, 2025.doi:10.1016/j.biopsych.2025.02.594

Background and education

Six years of research and clinical work with children and adolescents, four of them as a clinical counselor in a closed child psychiatry ward
M.A. candidate in Clinical Psychology
Graduate studies in Information Science
Harvard Medical School Executive Education, AI in Health Care
Undergraduate degree in behavioral sciences, GPA 97.5, President's Honors
Third-place poster at a national child and adolescent psychiatry conference
Presented in Toronto and Porto
Taught research design and statistics to psychology students
Hebrew, native. English, full professional

Beyond product

Together Against Bullying: Hand in Hand

A near-wordless short for a UN international day against bullying at school. Written, directed and produced by me, with AI video tools and Claude.

A parent and child walking to school hand in hand

Let's talk.

Senior or Lead AI Product Manager

AI-first products and AI agents

Vertical or domain PM

Healthcare and other high-trust industries

Head of Product

Early-stage AI startups

Message me on LinkedIn Open Claude skills kit

Asaf Wishnia AIPMWork is shown without company names, with sample data. Employers are on my CV and LinkedIn.