CASE FILE 001 · FORWARD-DEPLOYED / APPLIED AI
CASE FILE / FORWARD-DEPLOYED + APPLIED AI
I take ambiguous AI products from decision to deployment—product, architecture, implementation, evals, native release, telemetry, and production outcomes. One product created a category; the other reached the App Store in 14 days.
A 22-rendition voice ledger, R0–R13 roast lab, same-board comparisons, and 2,309 green tests make subjective voice behavior inspectable.
Independent per-photo verdicts require a 2-of-3 majority. Malformed assessments fail closed, one-off false alarms can be outvoted, and contract tests pin both sides of the safety bar.
A production voice regression became a controlled same-input evaluation. Model outputs, contracts, accessibility, and smoke paths became release gates.
Built for roles where one engineer must turn an ambiguous model capability into a dependable customer outcome.
EXHIBITS A—B / SHIPPED PRODUCTS
Yapword keeps a generative character inside live gameplay without giving the model control of game truth. That’s My Best creates a social-photo product without biometric identity, account access, or staff photo review.
Continuously interactive AI word game
The App Store’s first and only word game where a generative-AI character reads the live letter grid and responds throughout the entire game.
Multimodal AI friend quiz
Creator-supplied Instagram-grid screenshots become a playable friend quiz through a privacy-preserving 2-of-3 multimodal safety gate—without face matching, Meta access, or staff reviewing private photos.
EXHIBIT A1 / CATEGORY CREATION
Yapword was the first—and remains the only—word game on the Apple App Store where a generative-AI character reads the live letter grid and responds throughout the entire game.
View the App Store listingYapoleon reacts to every guess, creates contextual hints, remembers the player, and critiques the run.
Most AI word games use a model before or after play. Here, the rules engine owns the board and outcome while Yapoleon owns the live voice—opponent, narrator, hint system, relationship memory, and postgame critic.
Letters, placements, attempt, mode, standing, and player history become bounded input.
Validity, score, board truth, completion, standing, and memory remain outside the model.
A board-aware reaction, contextual hint, or postgame critique returns without authority over the outcome.
ARTIFACT / THE YAPOLEON VOICE LEDGER
A chronological production ledger traces every architecture, register, dial, roast-lab, surface, and incident change beside the real lines each version produced. The public exhibit keeps proprietary prompt text redacted while showing the engineering method and its receipts.
A long enumerated denylist flattened cadence and wit. Replacing it with a positive rule—what comic move to make—restored specificity without surrendering the boundary.
Identical game states were replayed across voice renditions and six relationship standings, making memory and register changes visible instead of relying on taste and recollection.
Deterministic software emits facts and state; the model turns them into language. That boundary keeps personality adaptive without allowing the LLM to invent score, memory, or game truth.
SYSTEMS HIDING UNDER THE CHARACTER
The repo shows a character product built as an operating system—not one impressive prompt.
Imperial Regard persists standing, recurring play habits, hint behavior, favorite openers, and notable wins. It emits bounded declarative facts—never a mood or instruction—so memory changes the comic material without scripting the joke.
Every live persona call records model, latency, token usage, request and deployment hashes, fallback state, and a bounded humor sample. Exact and near-duplicate detectors alert when Yapoleon repeats the same comic core inside one game.
A scheduled X pipeline posts only curated, closed-day roasts, revalidates spoiler safety, claims content before posting to prevent duplicates, degrades media failures to text, and chooses at-most-once delivery when the network outcome is ambiguous.
CASE STUDY / PRIVACY-PRESERVING MULTIMODAL SAFETY
Shipped from concept to the App Store in 14 days by replacing facial recognition and Meta access with user-controlled screenshots. When an image is suspected of containing a minor, one model opinion cannot reject or clear it: up to three independent adjudications vote per tile, while unresolved output fails closed. No staff review private photos.
The creator supplies 1–4 screenshots from a grid they already control.
The workflow isolates individual posts and rejects interface chrome, partials, and duplicates.
A clean first verdict proceeds. A suspected tile triggers up to two independent rechecks and requires a 2-of-3 majority.
The system favors objective, visually grounded memories with strong quiz potential.
Selected images become answerable questions, choices, and contextual reactions.
The creator confirms the answer key, seals the quiz, and shares it—without staff reviewing the photos.
EXHIBIT B1 / THE THREE-VERDICT GATE
This is visual content classification, not identity recognition. The system never determines who a person is. It asks only whether a supplied tile appears to contain someone under 18, then applies a conservative per-tile voting policy.
If every tile clears, publishing continues after one call. Nothing is re-run merely to burn latency and money.
A second agreeing flag reaches the majority immediately. A clean disagreement sends the disputed tile to the deciding vote.
Two clean verdicts outvote a one-off false alarm. Two flags exclude the tile. Each photo receives its own tally.
majority required
The automation does not need to know who anyone is. The operator does not need to see the user’s photos. The platform does not need to grant account access. A per-photo majority handles ambiguous model behavior, and factual authority stays with the creator.
AI development agents multiplied throughput. They did not own a decision, a release gate, or a production consequence.
DECISION RECORD / MODEL GOVERNANCE
Production AI becomes dependable when authority is explicit. Models can generate useful content; they do not own truth, scoring, safety, or spend.
Inspect the public engineeringThe model may propose; it does not become the authority.
Scores, memory, completion, and product rules remain inspectable.
Prompt-injection boundaries and explicit gates protect control flow.
Fallback chains, classified retries, budgets, and cached degradation.
Output caps, rate limits, attempt budgets, and measured generation cost.
Contracts, invariants, accessibility, smoke paths, and release gates.
PUBLIC RECORD / SOURCE AVAILABLE
Focused systems extracted from production lessons and published in the open.
Fallback chains, retry classification, budgets, and graceful degradation for production model calls.
02A deterministic boundary that isolates untrusted user content and keeps safety decisions outside the model.
03A competitive AI game with server-owned scoring, bounded model outputs, explicit safety gates, and hard cost ceilings.
05Production memory infrastructure, contributed upstream.
EXHIBIT C / SEARCH DISTRIBUTION
I treat technical SEO and generative-engine optimization as product infrastructure: make the facts crawlable, the entities legible, the answers citable, and the feedback measurable.
Server-rendered facts, crawl controls, canonicals, sitemaps, structured data, performance, and clean information architecture.
Source-qualified claims, entity clarity, extractable answer blocks, crawler policy, and repeated verification across answer engines.
Useful taxonomies, county and service clusters, multilingual pathways, internal-link graphs, and explicit editorial standards.
Cross-platform events, funnels, session replay, GA4 parity checks, search signals, and AI-visibility research turn discovery into an operating system.
OBSERVABILITY LOOP / POSTHOG
PostHog is the cross-platform operating record for product behavior—not a pageview counter. Anonymous session, game, prompt, deployment, and relationship identifiers connect AI behavior to completion, abandonment, replay, sharing, registration, and retention.
Platform-tagged events make game starts, submissions, completions, hints, shares, registration, and purchases comparable across surfaces.
Session evidence exposed rage clicks on a keyboard that looked live while validation was blocking it; the UI state was then bound to the real validation state.
Production allowlists, GA4 parity checks, explicit event contracts, and no-secret/no-PII payload rules keep the numbers decision-grade.
@YapoleonGreater extends Yapword into a governed, near-autonomous social-publishing system. It generates publication-ready responses in character within owner-defined rules; I retain the voice, publishing boundaries, escalation decisions, and consequences.
Inspect the X accountPublic domains traced to the live and repository record. Private systems and retired experiments are intentionally excluded.
PRIOR RECORD / OPERATOR JUDGMENT
More than a decade in commercial real estate taught me to sell complex assets, advise executives, repair operational data, and remain accountable when decisions had real consequences.
Nearly $500M in total volume as part of a two-person team.
Roughly 1,900 mapping errors identified and corrected.
Commercial judgment and technical execution in the same room.
NEXT FILE / THE RIGHT HARD PROBLEM
If you need someone who can turn an ambiguous AI opportunity into a shipped, measured product—without losing the last mile between model and customer—let’s talk.
alex@midnightdev.dev