Omer Abramovich
Product leader & builder
Product judgment. Builder instinct.
Strategy ↔ Systems ↔ Experience

I turn messy problems into products that move the business.

Product leader. 0→1 builder. AI practitioner.I work from strategy and user behavior all the way to prototypes, systems, experiments and measurable outcomes.

Explore my work Let’s talk
ProductAIAutomationGrowthUX
THE WORK, IN ONE PICTURE
From a messy problem to a clear product systemUsers, friction, data and ideas converge into product judgment, followed by building, testing and measuring an outcome. Users Data Friction Ideas Productjudgment Something that works.BUILD → TEST → MEASURE Find the leverage point.
CLARITY IS THE PRODUCT.
01 / What I love

Hard, slightly
messy problems.

The ones where people, business economics, data and technology collide. I enjoy turning something vague into something people can see, test and use.

01

Find the real problem.

Observe behavior and workflows before prescribing technology.

02

Build to learn.

Make flows, interfaces and AI behavior tangible. Then test the assumptions.

03

Prove the value.

A feature, model or agent only matters when it changes an outcome.

02 / Selected work

Six problems.
Six different ways in.

What needed to change. The product decision that mattered. And what it made possible.

↓ 06 STORIES
01 / Agentic financial insights

Less manual analysis.
Faster action on financial performance.

Financial reports show what changed. This agent investigates why—and what to do next. It reviews KPIs, drills into operational data, and turns significant findings into recommended actions and automated escalation tickets—helping teams uncover revenue risks, prioritize performance gaps and reduce manual investigation.

Financial report
Sample data
MetricActualvs. last month
Revenue MTD€4.31M+3.2%
Daily actives91.2K+4.2%
ARPU€46.7−6.1%
Registrations18.4K↑ 27%
Activation 21.3%↓ 9.6 pp
The insight

More signups.
Less valuable traffic.

A new partner channel accounts for 62% of registration growth, but only 17% of its signups activate.

Inside the thinking

Review financial and operational KPIs, prioritize meaningful anomalies, and investigate the data behind them. Turn the finding into a proposed action and an evidence-backed escalation for the right team. This acquisition example illustrates a broader capability—not an acquisition-only tool.

Business KPIs to improve

  • Analyst hours saved
  • Time to detect anomalies
  • Time to identify root causes
  • Time to escalation
  • Issue-resolution time
02 / AI business analystArchitecture

Less BI backlog.
Faster answers for the business.

An agentic AI business analyst that lets product, finance and operations teams ask complex business questions across 113 tables and 15 views—giving the business faster self-serve answers while reducing routine query-writing and data preparation, so BI analysts can focus on deeper investigation.

Ask a complex business question / inspect the answer path
“Which acquisition sources increased registrations in Spain last week but reduced activation quality and net revenue versus the prior four-week average?”
Relevant dataChecked queryExplained finding
CONTEXT

Registration, activation and NGR facts by market, acquisition source and comparison window.

CONTROLS

Read-only access, business rules and query limits.

Retrieve the right joins and business definitions, run checked SQL, and show the evidence behind the answer.

Agentic AIRAGIntent routingGuardrailsEvalsObservability
KPIs to measure

Analyst hours savedTime to verified answerAnswers accepted without rework

Inside the thinking

The problem

Routine business questions create repetitive query-writing and data preparation. Product, finance and operations wait for answers while BI analysts handle work that could be made self-service.

The insight

This is an on-demand investigation tool, not a daily report summary. A useful answer needs the right data, the correct business definitions and a calculation that can be inspected.

The decision

Design a natural-language analyst over 113 tables and 15 views. Route the question, retrieve relevant schema and examples, generate SQL, apply deterministic checks, then execute with read-only permissions and explain the result.

The quality loop

Evaluate against representative questions and expected results. Trace retrieval, query generation and execution; feed analyst corrections back into the evaluation set. Measure hours saved and time to a verified answer alongside correctness.

Give analysts back the routine work. Keep the evidence behind every answer.

03 / Disengagement detectionProof of concept

See likely disengagement before the session ends.

Used an LLM to help uncover and interpret candidate warning signals in session behavior. Tested those signals against 5.1 million events from 4,544 customers—revealing a possible 1–3-minute early-warning window. The LLM helped find what to investigate; measured behavior drove the score.

FROM BEHAVIOR TO EARLY WARNING
01 Explore session behavior
5.1Mevents studied
4,544customers
Actions & timingPatterns before exit
02 / LLM-assisted discoveryWhat changes before someone leaves?

Surface patterns. Interpret them. Propose signals to test.

Longer pausesActivity shifts
Test candidate signals against observed sessions
03 / Behavior-based scoreEarlier visibility into session exit.
1–3 minpotential lead time
LLM-assisted discovery. Data-tested signals.
LLM-assisted discoveryBehavioral analyticsFeature engineering
Next validation

Signal reliabilityFalse-alert rateCustomer experience

Inside the thinking

The problem

Session exits were visible after the customer had already gone. The question was whether measurable changes in behavior appeared early enough to give teams useful visibility.

Where the LLM helped

The LLM was used to help surface and interpret candidate signals in the behavioral analysis. Its role was discovery: identifying promising patterns to investigate, rather than treating a generated explanation as proof of a reliable prediction.

What the data tested

The study covered 5.1 million events from 4,544 customers. Four observable signals were combined into a rolling score. At the tested threshold, roughly 50% of exit-zone events were detected, while about 15% of active play also triggered. These are study observations, not a production guarantee.

The value to validate

The findings suggested a possible 1–3-minute lead time. Further validation would test reliability and customer value separately, with attention to false alerts, safety and experience.

Use the LLM to find promising signals. Use data to test whether they hold.

04 / AI campaign managerConcept & architecture

Less time planning CRM retention campaigns.
Better returns on retention spend.

An agentic AI campaign-manager concept for CRM retention campaigns. It helps teams decide who to target, what message or offer to send, how much budget to allocate and what return to expect—inside a 3-step planning and approval workflow designed to save planning time and improve spending decisions.

Agentic workflowsCampaign planningML + LLMUpliftBudget guardrails
KPIs to measure

Planning hours per campaignIncremental campaign ROI7-day retention

From performance signal to reviewed campaign
Performance changeRelevant audienceRecommendation
01 / BASICSSet the goalObjective
+ audience
02 / CONFIGUREBuild the planCampaign
+ budget
03 / REVIEWCheck & approveExpected return
+ controls
Approved campaignMeasured outcome
↳ Feed results into the next recommendation ↲
Expected ROIBudget limitsManager control
Inside the thinking

The problem

CRM teams move between performance reports, audience selection, campaign planning and budget checks. Manual handoffs consume time and make it harder to assess which actions deserve investment.

The insight

The goal is not more campaigns. It is a relevant recommendation, a clear case for spending and a way to measure the incremental result.

The decision

Design a three-step workflow for campaign basics, advanced configuration and review. Use predictive models for quantitative estimates, an LLM for interpretation and composition, and deterministic rules for budgets, eligibility and responsible-use constraints.

The intended value

Reduce manual planning and improve targeting and budget decisions. Evaluate planning hours per campaign, incremental ROI and 7-day retention before treating the concept’s expected benefits as realized outcomes.

A campaign is only worth launching when the expected value justifies the spend.

05 / Workflow automationLaunched product

14 days to 2 days.
Mortgage processing, accelerated.

Automated the repetitive processing after human approval. In one banking workflow, a person still made the approval decision. Once approved, software robots took over the back-office execution—moving customer data across banking systems and sending confirmation automatically. That turned a long, repetitive process into a much faster one.

FROM APPROVED CASE TO AUTOMATED MORTGAGE PROCESSING
01 Capture the post-approval work. Discover repetition.
Mortgage queue
Approved case
Customer file ••••
StatusApproved
Bank systems
System A / B / C
Customer data
UpdateProcess
Captured sequence
Open approved caseUpdate systemsSend confirmation

Repeated across mortgage-processing sessions. Stable work to automate.

02 Build & test the mortgage workflow
03 / Software executes after approvalMortgage processing
Example
Read approved case
Update systems A / B / C
Send confirmation
Mortgage request processed automatically.
Edge cases routed to staff for review.
People approve the decision. Software executes the repetitive processing.
One deployed automation example
14 → 2 daysmortgage turnaround
80 FTEequivalent repetitive work freed
Inside the thinking

Capture & discover

Task mining captured the real post-approval work carried out by back-office staff. That made the repetitive sequence visible: open the approved mortgage case, move customer data across banking systems, and complete the clerical steps needed to finish the request.

Prioritize the process

The approval decision stayed with a person. The automation target was the stable, repetitive processing that happened after approval—exactly the kind of high-volume, cross-system work worth automating.

Build & execute

Once the workflow was defined, software robots executed the post-approval steps across the required banking systems and sent confirmation when complete. In this deployment, 8 robots handled four back-office processes and about 550 daily transactions.

Where value is realized

In the mortgage workflow, turnaround time dropped from 14 days to 2 days while freeing work equivalent to 80 employees. More broadly, the launched RPA product saved 100,000+ work-hours and generated $15M in new business.

Discover the work. Keep the human decision. Let software do the repetitive processing.

06 / Customer journey & UXLaunched products

A better customer journey.
A 50% uplift in lifetime value.

Launched 2 consumer products across web and native apps and led a full UX overhaul—connecting activation, engagement and retention improvements with commercial performance, and contributing to a 50% increase in LTV.

Customer experience → commercial value
+50%

Lifetime value
increase contributed to

2consumer products
Web + nativeone connected lifecycle
ActivateEngageRetainGrow value
GrowthUXLifecycleExperimentation

Product and UX changes tied to a measurable business outcome.

Inside the thinking

The problem

A fragmented customer experience needed a joined-up product direction, rather than isolated improvements to individual screens.

The insight

Acquisition, activation, engagement, retention and monetization form one lifecycle. Design the experience and its economics together.

The decision

Launch two consumer products across web and native apps and lead a full UX/UI overhaul. Use funnel analytics and experimentation to identify friction and prioritize improvements throughout the lifecycle.

The outcome

The combined product and UX transformation contributed to a 50% increase in customer lifetime value. Connect improvements in activation and retention to the commercial result, rather than treating visual redesign as the end goal.

Better journeys should create better experiences—and better business results.

03 / How I think

Technology is
a choice.
An outcome
is the point.

A few principles that connect the product decision to the system underneath it.

Judgment before novelty.
01

Model ≠ product.

The product is the decision, action or experience the model enables. Start there, then choose the technology.

02

Probabilistic intelligence.
Deterministic boundaries.

Let AI reason where ambiguity exists. Enforce what must always be true outside the model’s discretion.

03

Build before debating.

A working flow or interactive prototype turns assumptions into something we can test. Learn what matters before scaling the investment.

04

Eval is regression
testing for AI.

Prompts and models are product behavior. Evaluate representative cases, inspect failures and feed what you learn back into the next iteration.

04 / I still build

I don’t stop
at the PRD.

From user flows and rapid mockups to working prototypes, agentic workflows and 0→1 products.

Build

Replit · Claude Code · OpenAI Codex

Design

Figma · Claude Design · UI/UX · interactive prototypes

AI systems

RAG · routing · context engineering · agentic workflows · structured outputs

Quality

Evals · observability · guardrails · human-in-the-loop

Connect

Langflow · APIs · MCP · workflow automation

I don’t claim to replace engineering or design.
I remove distance between an idea and evidence.

THE WORKBENCH / IDEA → EVIDENCE
A FLOW YOU CAN SEE AND QUESTION
Make the next
step obvious.
WORKSPACEOne task.
Next step →
01 / Relevant context
02 / Clear action
03 / Visible feedback
Design: make the interaction testable.
05 / A little about me

Product has always
been my way of
understanding systems.

I started close to the code, moved into automation and product, built companies and platforms, and found myself close to the code again.

Now, AI lets me test ideas at a speed I always wanted as a product leader. What hasn’t changed is the part I enjoy most: finding the leverage point.

15+years across product
& engineering
3patents
2innovation awards
06 / Let’s talk

Have a hard problem?
I’m interested.

Full-time product & AI leadership.
Selective fractional and advisory work, too.

Explore a leadership role Talk fractional & advisory
LinkedIn Request CV Back to top