What is failing?
Choice, not content availability.
07 / 09 · Consumer Recommendation & Growth Product
Tonight
The same person can be a different viewer tonight.
Streaming solved access. It did not fully solve choosing. Tonight asks what fits this session—time, company, mood, energy and appetite for novelty—then combines that with long-term taste and catalog truth to offer a deliberately small shortlist.
CONSUMER PRODUCT + EXPERIMENTATION
How do you optimize satisfaction without turning personalization into repetitive engagement?
This case is deliberately less agentic. It proves recommendation objectives, consumer behavior, experiment design and long-term retention trade-offs.
01 / THE PRODUCT PROBLEM
Start with the decision the user cannot make well today.
Streaming solved access. It did not fully solve choosing. Tonight asks what fits this session—time, company, mood, energy and appetite for novelty—then combines that with long-term taste and catalog truth to offer a deliberately small shortlist.
The portfolio test: a client should understand the human problem before seeing a model, agent or framework.
02 / HOW RIKA THINKS
From ambiguity to a decision system.
The five-step reasoning pattern stays consistent; the actual product logic is specific to this problem.
What changes session to session?
Time, company, mood, energy, novelty and device/context.
What can optimization damage?
Overfitting to history, popularity bias, filter bubbles, maturity/rights mistakes.
What should rank?
Persistent taste + current intent + catalog eligibility + exploration.
What should the metric teach us?
Did the viewer find something worth starting and feel satisfied afterward?
03 / TRANSFERABLE PROBLEMS
The pattern travels. The constraints change the solution.
This is how I would reuse the thinking without copying the product.
Food delivery
Historical taste helps, but tonight's occasion and group change the recommendation.
Music
Long-term taste differs from workout, focus, commute or party intent.
Ecommerce
A user profile is not the same as the current shopping mission.
Travel discovery
Known preferences combine with time/location constraints in the current moment.
04 / USERS + BUYER
Who receives value, who operates it, and who pays?
AI products often fail when one persona is treated as ‘the user.’
Intent-rich, time-poor viewer
Has multiple services or a large catalog but wants to start something quickly.
Shared-household / co-viewer session
Two or more preferences must be reconciled without endless compromise browsing.
Streaming / FAST / niche OTT product leader
Wants higher perceived catalog value, faster play decisions and retention without abusing watch-time optimization.
JOBS TO BE DONE
Functional, emotional and social value.
Turn a session context into a short, eligible, high-confidence shortlist.
Avoid wasting scarce leisure time browsing or regretting the choice.
Choose something that works for everyone in the room without negotiation fatigue.
CUSTOMER JOURNEY
The question changes as evidence arrives.
Arrive
Viewer opens the service.
Default recommender can still work.
State intent
Time, company, mood, novelty.
Intent can be explicit or later inferred.
Rank
Candidate pools are generated and reranked.
Rights/maturity/diversity gates apply.
Play
Viewer chooses one of three.
Explanation answers ‘why tonight?’
Satisfy & return
Outcome signal records completion/abandonment/satisfaction.
Next session updates, but one bad session should not rewrite identity.
05 / PRODUCT DECISION RECORDS
The product is the sum of choices and trade-offs.
Each decision includes an implicit reversal test: better evidence can change the choice.
Session over profile
Long-term taste is useful context, not a complete description of what the viewer wants now.
Three confident choices over one hundred plausible choices
The product optimizes decision quality and time-to-play, not catalog exposure.
Satisfied Session Rate over raw watch time
More minutes are not always more value; completion, abandonment and lightweight satisfaction matter.
Exploration is a product policy
One exploration slot protects novelty and catalog breadth instead of letting the ranker collapse into familiar popularity.
06 / AI SYSTEM
The architecture separates AI judgment from exact rules, evidence, tools and human authority.
Every connector has a defined job. Animated dashed lines represent active evaluation/learning loops rather than decorative motion.
Primary decision/data flowContext/evidenceHuman/consequential pathContinuous evaluation loop
HOW I WOULD BUILD IT
Point by point, from state to operation.
Session encoder
Translate explicit or inferred context into a structured session vector.
Candidate generation
Blend collaborative filtering, content similarity, continue-watching, editorial and exploration pools.
Feature layer
Combine long-term taste, session intent, title features, recency, completion history and eligibility.
Ranker
Score expected session fit, not generic engagement.
Policy re-ranker
Apply rights, maturity, regional availability, diversity and repetition constraints.
Exploration controller
Reserve controlled novelty so the system learns without flooding the user.
Explanation layer
Generate ‘why tonight?’ only from ranking features and catalog metadata.
Experiment + feedback
Play, abandonment, satisfaction and return behavior update objectives and A/B decisions.
07 / EVALUATION ARCHITECTURE
How do I know the AI deserves to ship?
Component eval, system eval and product outcome are deliberately separated.
NDCG/Recall by candidate pool
NDCG/Recall by candidate pool, session-intent match, cold-start/new-title coverage.
Rights/maturity validity
Rights/maturity validity, repetition cap, catalog-source truth.
Genre/source concentration
Genre/source concentration, long-tail exposure and exploration acceptance.
Control vs explicit-session vs inferred-se
Control vs explicit-session vs inferred-session treatments.
Time-to-first-play
Time-to-first-play, browse abandonment, Satisfied Session Rate, early abandonment and repeat-session return.
Retention and catalog diversity must not d
Retention and catalog diversity must not degrade while short-term starts increase.
Closed loop: trace → classify failure → add/refresh eval case → change source/retrieval/model/prompt/rule/tool → regression suite → controlled release → monitor outcomes/overrides → repeat.
08 / FEATURE PRD
A concrete example of how strategy becomes engineering work.
The PRD is intentionally feature-level and includes AI behavior, deterministic controls, telemetry and non-goals.
PRD example — Session intent shortlist
Problem
Viewers often browse across a large catalog even when they can describe what fits the current session.
Outcome
Capture minimal session intent and return three eligible choices with transparent fit within seconds.
User stories
- As a viewer, I can say I have 35 minutes, low energy and want something familiar-but-not-repetitive.
- As a shared session, I can indicate who is watching so maturity and taste constraints are respected.
Functional + AI requirements
- Capture time available, company, energy/mood and novelty preference with an optional skip path.
- Generate candidates from at least three pools so one algorithm cannot dominate.
- Apply rights, maturity and regional eligibility before presentation.
- Return exactly three primary choices plus one optional exploration choice.
- ‘Why this?’ must cite session-fit factors, not invent a narrative.
Acceptance criteria
- No ineligible title can be shown.
- All shown titles fit the selected time window within tolerance.
- A skip-intent path still uses the baseline recommender.
- Exploration choice cannot displace all high-confidence options.
- Treatment telemetry supports session-level experiment analysis.
Telemetry
session_context_opened · intent_submitted · shortlist_rendered · why_opened · play_started · abandoned_5m · session_satisfied · exploration_accepted
Non-goals
- Replacing the core catalog recommender with an LLM
- Optimizing only watch time
- Assuming every user wants to answer a questionnaire
Engineering handoff
Typed state/schema · API/tool contracts · error states · permissions · eval fixtures · analytics events · rollout/rollback.
09 / BUILD VS BUY + RELIABILITY
Do not custom-build commodity infrastructure—and do not pretend every API always works.
BUILD DIFFERENTIATION
Session objective, ranking policy and satisfaction loop
Use standard infrastructure for storage, models, tracing and connectors where it does not create strategic advantage.
BAD-DAY MODE
Fallback to baseline recommender
Timeouts, stale data, partial results, rate limits and provider failures are explicit product states—not invisible backend details.
10 / WORKING PRODUCT JOURNEY
Use the product from input to changed state.
Every click represents a user/product decision and explains why the information is needed.
TONIGHT · SESSION-AWARE DISCOVERY
Your profile is not enough.
The same person may want a comfort comedy on Tuesday and a demanding documentary on Saturday.
Three choices, each for a different reason.
BEST SESSION FIT
Harbor
31 min · calm · familiar-but-not-repetitive.
EXPLORE
The Signal
One exploration slot.
SAFE CHOICE
Afterlight
Matches historical preference.
Why Harbor fits tonight.
TIME FIT
31 / 35 min
No unfinished ending risk.
SESSION FIT
Low effort
Matches current energy.
TASTE FIT
High
Uses long-term preference as context.
NOVELTY
Moderate
Avoids exact repetition.
Playback started. Now measure satisfaction, not just minutes.
Was this worth the session?
A lightweight signal helps distinguish accidental playback from a genuinely satisfying recommendation.
Was this case study useful?
A quick signal helps me understand what is useful to recruiters, founders and product teams.
11 / PRODUCT LIFECYCLE
Autonomy and market scope are earned in stages.
Objective proof
Validate decision friction and session-context signal value.
Explicit session layer
Three-choice shortlist on top of existing recommender.
Experiment system
A/B test intent capture, shortlist and explanation.
Inferred context
Reduce friction using time/device/co-viewing signals with user control.
Platform layer
Offer session-intent ranking as a configurable OTT capability.
GTM + ADOPTION
A credible route into the market.
Streaming/FAST/niche OTT teams trying to improve perceived catalog value and retention.
Services with broad catalogs and visible browse abandonment.
Start as an optional ‘What fits tonight?’ surface, not a homepage replacement.
Session layer → personalization SDK → cross-service discovery only if rights/business model supports it.
KILL / PIVOT CRITERION
What would make me stop?
Stop or radically simplify if explicit session capture adds friction without improving satisfaction/retention, or if a standard recommender using passive context achieves the same value with less user effort.
12 / TOOL + MODEL DECISIONS
Tools are mapped to responsibility—not displayed as decoration.
Where the source material does not prove an implementation, the portfolio says proposed/candidate rather than “built with.”
Recommender servicecore ML candidate/ranking
Feature storesession + profile features
LLM intent encoderstructured context
Experiment platformA/B + guardrails
Catalog/rights APIdeterministic eligibility
Analytics / Amplitude-stylebehavioral telemetry
13 / SECONDARY RESEARCH
Evidence establishes the problem environment. It does not magically validate the solution.
Deloitte Digital Media Monitor 2026
Deloitte reports 41% of surveyed consumers had cancelled a paid SVOD service in the prior six months; retention remains a material industry problem. This does not prove decision friction is the cause.
Open source ↗Deloitte Digital Media Trends 2026
Deloitte reports flat streaming spend and strong price sensitivity while media companies seek more personalized engagement. Product implication: recommendation quality must create perceived value, not just more inventory.
Open source ↗Portfolio research hypothesis
The specific hypothesis—session intent reduces decision effort and increases satisfied sessions—still requires primary research and controlled experimentation. It is not treated as proven by churn data.
Internal/source-project basisEVIDENCE PACK
Download Tonight Product Evidence.xlsx ↗The working assumptions, PRD, eval suite and roadmap are inspectable.
No spreadsheet preview is embedded. Open the workbook only if you want the detail.