case_file_001 · forward-deployed ai lead
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.
The two exhibits below are the native App Store releases. Further work — including Yapoleon's Court and the AI Boardroom Forecast Audit — is on the main workfolio.
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
A character inside the deterministic game loop.
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
I removed the dependency instead of negotiating with it.
Creator-supplied Instagram-grid screenshots become a playable friend quiz without face matching, Meta API access, or staff reviewing private photos.
01 / one accountable human
AI development agents did not own a decision, a release gate, or a production outcome. Every link in this chain resolved to one person.
Product decision
Architecture call
Implementation
Evals + tests
Native release
Production outcome
exhibit a1 / category creation
Most AI word games use a model before or after play. Here the rules engine owns the board and the outcome while the character 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 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.
renditions logged
controlled roast ladder
live production lines
tests green in captured build
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 letting the model invent score, memory, or game truth.
exhibit b / constraint-driven architecture
Shipped from concept to the App Store in 14 days by replacing facial recognition and Meta access with user-controlled screenshots. The pipeline segments the grid, excludes uncertain imagery, creates the quiz, and bounds what the model's proposed answer key can reach.
The creator supplies 1-4 screenshots from a grid they already control.
The workflow isolates individual posts and rejects interface chrome, partials, and duplicates.
Unsuitable content, apparent minors, and uncertain images are excluded from automated selection.
The system favors objective, visually grounded memories with strong quiz potential.
Selected images become answerable questions, choices, and contextual reactions.
The default path seals the model's proposed key; a paid override lets the creator replace it first. Either way the player payload carries no answers and the seal is immutable — without staff reviewing the photos.
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. Safety stays conservative, and the invariants hold whether or not the creator edits the key.
decision record / model governance
Production AI becomes dependable when authority is explicit. Models generate useful content; they do not own truth, scoring, safety, or spend.
Server-owned and invariant-bounded
The model can propose and perform; code decides the payload, the reveal, and the score.
Deterministic outside the model
Scores, memory, completion, and product rules remain inspectable.
Untrusted content isolated as data
Prompt-injection boundaries and explicit gates protect control flow.
Failure paths designed in advance
Fallback chains, classified retries, budgets, and cached degradation.
Cost is a product constraint
Output caps, rate limits, attempt budgets, and measured generation cost.
Release gates across the stack
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.
SafetyA deterministic boundary that isolates untrusted user content and keeps safety decisions outside the model.
OperationsHealth-driven recovery, safe backups, and local-only networking for production memory infrastructure.
ResearchA dated, source-qualified reference on how answer engines crawl, retrieve, and cite web content.
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.
Indexability is engineered
Server-rendered facts, crawl controls, canonicals, sitemaps, structured data, performance, and clean information architecture.
Citation surfaces are designed
Source-qualified claims, entity clarity, extractable answer blocks, crawler policy, and repeated verification across answer engines.
Coverage without content sludge
Useful taxonomies, service clusters, multilingual pathways, internal-link graphs, and explicit editorial standards.
Telemetry closes the loop
Cross-platform events, funnels, session replay, analytics parity checks, and search signals turn discovery into an operating system.
observability loop
Telemetry 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, analytics parity checks, explicit event contracts, and no-secret / no-PII payload rules keep the numbers decision-grade.
exhibit d / measurement
Every product has one quality property no test can assert — whether the voice is good, whether the scoring is fair, whether it's safe to publish. So I build the grader for that property, and then I take it out of the product. Three products in six weeks; read them in order, because the grader work matures across them.
Is the character's voice good — and what does it cost?
A multi-turn bench that walks whole games through the real prompt builder, sweeping the model's thinking level.
~82% of inference COGS was invisible thinking. Zero thinking preserved the voice. The highest setting burned 8,991–11,663 tokens per game — $80–105 per 1k games — and was not better.
Is the scoring rubric fair, learnable, and un-gameable?
Yapword's harness, forked and re-pointed at a different measurement: 30 demands x weak/mid/strong, plus a fixed-mold probe, against the live judge.
72.2% mid / 0% weak / 96.7% strong — a clean skill gradient. Fixed mold 0% off-axis. 5 of 5 anti-gaming probes pass live. 1,095 calls, 0 errors.
Is it safe to publish when the classifier itself is unreliable?
A fail-closed state machine with quorum voting over an injectable classifier — extracted out of the product entirely.
Shipped as llm-safety-gate: zero-dependency, MIT, every failure path unit-tested without a network.
Ad hoc → reused → extracted as infrastructure.
The first harness was written for one product's problem. The second product didn't get a new one — the first was forked and aimed at a question it was never built for, which is the first evidence the method transfers. By the third, the grader stopped living inside the product at all and shipped as a standalone library someone else can inject their own classifier into.
Code can be graded against a reference build. Voice, fairness, and taste cannot. So the scoring function is validated by its properties instead.
Better input must score better
Strong 96.7% > mid 72.2% > weak 0%. If the ordering breaks, the rubric is broken.
The negative control
One fixed rhetorical mold, applied to all 30 days, must lose on its off-axis days. It wins 0%.
Five adversarial probes, live
Naked flattery, prompt injection, legitimate audacity (the false-positive check), delimiter breakout, grovel-on-economy. All pass.
The model never emits the score
It returns five clamped axis sub-scores; a pure server-side function the model cannot see or reach computes the result.
Distributions, not point values
Hosted inference is not bit-reproducible at low temperature, so every cell runs at least three times and the spread is reported.
The calibration also produced a finding I kept rather than buried: the target was specified as a median win-rate, and the median turned out to be bimodal and degenerate — an offline sweep of hundreds of candidate curves found zero with a median in band. The mean is smooth and stable across independent runs. The document says so, and says why.
exhibit e / distribution surface
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, the publishing boundaries, the escalation decisions, and the consequences — the same authority split that governs every other model in this file.
Products, tools, and commercial sites — each a real distribution surface. 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 stay accountable when decisions had real consequences.
Roughly $400M in total volume as half of a two-person team — buying, selling, leasing, and repositioning.
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.