/ showcase
BEHIND THE LOGINS
Most of what this lab ships lives where you can't see it — staff consoles, customer dashboards, paid tiers. This is the public tour: 18 production systems, what they do, and the engineering that makes them hold up.
Everything here runs in production today. No demos, no vaporware — ship or it doesn't exist.
VOICE & TELEPHONY
An AI co-pilot listening to live shop calls, whispering to the advisor — not the customer.
Inbound calls stream to a dedicated audio service that buffers each speaker separately and transcribes in five-second chunks — so "customer said X, advisor said Y" is exact, never overlap garbage.
Every fifteen seconds a coaching model reads the fresh transcript and decides whether it has something worth saying. Most ticks stay silent by design — a signal-to-noise gate keeps the advisor from learning to ignore the bubbles.
When a nudge does land, it carries evidence: how many similar calls back the suggestion and what it did to booking rates. After the call, sentiment and topics are extracted and the call is attributed to the work it created.
A tool-calling AI agent wired into a live shop system — it checks availability, quotes real services, and books the work.
Most "AI receptionists" hand you a transcript. This one carries tools: availability checks, appointment booking, customer lookup, rewards balances, diagnostic-code and vehicle-spec reference — all routed through the same integration layer the shop runs on.
Safety lives inside the persona, caught conversationally instead of by a silent validator: impossible vehicle combos get questioned, out-of-market makes get declined, and pricing never drops below the floor the shop actually charges.
GENERATIVE PIPELINES
Upload a handful of photos; get a custom-trained model, composited portraits, and a talking-head video.
Too few training photos? The pipeline synthesizes additional angle variants to bootstrap model quality before fine-tuning begins. Training, generation, background synthesis, and compositing each run on the provider best at that job.
The hard part is reliability: results arrive by webhook or by poll, whichever wins, and a unique-index ingest means the answer lands exactly once — under load, after restarts, every time. Aspect ratios survive the whole chain.
Every model trained on a real face sits behind a consent record, tracked as a biometric asset with explicit retention.
Preset-driven AI video — shot specs in, production b-roll out, provenance on every clip.
Presets describe the shot like a DP would: camera motion, lens, duration. A scene description merges into the preset and the request routes to whichever video model fits the brief.
Each clip records exactly which preset and parameters produced it — so a good result is reproducible, not a slot-machine pull.
AGENTS & MEMORY
A personal agent built on a deliberate split: a small curated memory, and a large on-demand recall.
Memory is consolidated hourly — salience-scored, capped to a small working set, fed into every decision the agent makes. Recall is the opposite: vector-indexed, huge, and queried only when needed. Most agent stacks conflate the two; the split is why this one neither forgets what matters nor drowns in what doesn't.
An autonomous executor researches and drafts toward project milestones under hard gates — bounded cycles, lifetime budget caps, and a fabrication guard that halts and asks instead of inventing context.
A daily self-review reads its own decisions and tool calls and files structured findings. Observation only — it proposes nothing it can't evidence.
Semantic + keyword retrieval fused by reciprocal rank, with privilege filters applied twice — at query and at result.
Dense vectors find what keywords miss; keywords anchor what embeddings blur. Reciprocal-rank fusion merges both lists so neither mode dominates.
In privilege-sensitive corpora, the filter isn't a polite suggestion: scoping is enforced on the way in and re-checked on the way out, so a retrieval can never leak across matter boundaries.
Reads high-stakes documents, extracts facts with citations back to source, and scans the whole corpus for statements that disagree.
Every extracted fact points back to the exact place it came from — claims you can audit, not summaries you have to trust.
Contradiction detection runs as a semantic pair scan across the full corpus: statements that conflict get surfaced side-by-side with their sources, even when they're phrased nothing alike.
It frames considerations, never decisions. The human remains the thinker — that's doctrine, not disclaimer.
SHOP INTELLIGENCE
Seven analytics products, one architecture: AI runs on a schedule, writes to the database, and the UI just reads.
No "Analyze with AI" buttons, no spinners waiting on a model. Intelligence is computed on cron rhythm by pure functions, materialized as snapshots, and read instantly. A broken job degrades one card, never the dashboard.
The seven products: estimate intelligence, service recommendations, no-show prediction, decline recovery, customer lifecycle staging, comeback detection, and cohort retention — each pluggable as a pure function, a schedule, a collection, and a read-only page.
And none of them ever auto-contacts a customer. Outreach is a separate, explicit, human-approved act.
Seven weighted factors predict which appointments will ghost — and the system grades its own predictions.
New customers don't pin to zero: smoothing keeps the math honest when history is thin. Every prediction's outcome is recorded, so the model's calibration is measured against reality, not vibes.
Work a customer declined doesn't vanish — it gets scored, drafted, and queued for a human to send.
Three decoupled phases: urgency scoring, evidence-backed message drafting (knowledge-base first, model fallback), and an eight-state queue. The drafting never touches the sending — a person approves every outbound message.
Retention cohorts anchored to a customer's first real visit — with lifetime value broken out per lead source.
Cohorts start the month a customer first converts, excluding prospects who never did — so retention curves measure real customers, and every marketing channel answers for the lifetime value it actually produced.
Replacing a business's operating software without a flag day: every write goes to both systems until the new one has earned primary.
Provider interfaces with three implementations — legacy, native, and no-op — wired per tenant from config. Application code asks the orchestrator; it doesn't know or care which system is primary this month.
Cross-system foreign keys and idempotent upserts give every record a bidirectional audit trail, so the migration can be verified row by row instead of trusted.
Walk-around photos, line-by-line customer approval, and authorizations that can prove they weren't altered.
Customers approve work item by item through signed, expiring links. Each authorization carries a cryptographic hash of exactly what was approved, by whom, for how much scope — change any of it and the seal breaks.
DATA PRODUCTS
Draw your audience on a map, mail it, and know what every single piece earned.
Audience selection is genuinely geographic: lasso customer pins or pick carrier routes in the browser, with polygon math running client-side. Locked audiences freeze as versioned snapshots — the list you approved is the list that mails.
Attribution closes the loop three ways: QR scans, tracked phone numbers, and address-matched visits, rolled into per-piece profit and loss. Automatic holdout groups measure lift instead of assuming it.
CREATOR TOOLS
A full canvas editor in the browser — with AI background removal, outpainting, and natural-language edits.
Layered composition, rulers, effects, stamps, fonts — plus mask-driven canvas extension and "describe the change" edits powered by image models. Built for production graphics, not demos.
A multi-track video editor that lives in the browser and saves to the same platform everything else runs on.
Tracks, clips, and media management with projects stored as structured data — no desktop round-trip for the cuts that feed the content pipeline.
PLATFORM
Every site in this portfolio — this one included — is a single application serving thirteen isolated tenants.
Each brand gets its own isolated database, its own theme with full light/dark palettes, its own feature flags, and pages composed from JSON blocks — building a new tenant means editing config, not forking code. Bespoke components remain an explicit opt-out for the sites that earn one.
Host-header resolution handles custom domains and subdomains through the same path, and a nine-role permission matrix gates every staff surface.
The floating hashtags that probably brought you here — a 200-tag map of everything on this page.
A registry of two hundred hashtags spans every project in the portfolio. A handful at a time drift through the hero on jeffsperandeo.com, weight-biased toward the flagships, never covering the figure, pausing under your cursor. Public work links straight to its own domain; everything behind a login links here.
WANT ONE OF THESE RUNNING FOR YOU?
Every system above was built for a real operation with real stakes. If your problem looks like one of these — or doesn't look like anything on this page — email Jeff or start at Geneva Lakes Media.