R&D Landmark 01Applied NLP · Salesforce Apex + Agentforce
An AI that reads every buyer comment
Prospects comment on floor plans in any language — on a long-text field that the platform's no-code automation structurally cannot process. We documented each platform wall, then engineered around them.
- Custom Apex asynchronous pipeline invokes a version-controlled LLM prompt template on every new or edited comment.
- Classifies sentiment (5 classes), detects the language, and writes a one-line English rationale for the sales team.
- Verified on French, English, Spanish, German, Italian — and Québécois slang.
- Self-healing: failed classifications retry on the next edit; results land in under a minute.
R&D Landmark 02Conversational AI × Web
An agent that shows, not just tells
Chat AIs answer in text. Ours operates the website: ask "where's the nearest grocery store?" and the map moves.
- Proprietary command protocol (structured envelopes + a postMessage bridge) lets the agent reposition a full-screen map, filter listings and navigate pages.
- Real driving routes and drive times computed live; nearest options ranked, starred and framed with the project pin.
- Navigation-aware: every message silently carries page context, so the agent knows what the visitor is looking at.
- Solved LLM fabrication with deterministic envelope generation — the model can never corrupt a command.
R&D Landmark 03System Architecture
A proprietary multi-agent architecture
- One router + 9 specialized subagents (search, project knowledge, objections, booking, email nurturing…), each backed by deterministic Apex services.
- Grounded by design: data is fetched by code, never invented by the model — the LLM presents, it doesn't improvise.
- Knowledge library and objection playbooks live in CRM records the team edits directly — behaviour changes without touching code.
- The same agent serves web chat and automated email nurturing from a single brain.
R&D Landmark 04Performance Engineering
66% faster, by replacing the default engine
The platform's default retrieval-and-generation chain produced 15–20 second answers — a conversation killer. We re-engineered the response path with our own methods.
- Fact-only payloads: 6,000 → 1,400 characters through the LLM; media and descriptions joined client-side.
- Deterministic pre-reasoning execution — actions run before the model speaks, not through it.
- Redundant inner LLM passes skipped — one generation step instead of two.
- Response-time instrumentation built into the chat — the ⏱ 5.3 s chip in the capture is a real production measurement.
R&D Landmark 05Real-time Analytics
One month of manual reporting → real time
The platform's native analytics couldn't express the analysis we needed; the manual study took a month to produce and was stale on arrival.
- Proprietary dashboard app on the Cloudflare edge, querying the Salesforce semantic layer live — ANSI SQL over Data Cloud, plus direct CRM queries.
- Scientific-grade Plotly visualizations: demand-vs-price bubbles, 3D scatter, correlation heatmaps, selection-driver analysis.
- 2,641 leads, unit & option selections and VIP deposits — refreshed continuously, cached at the edge.
R&D Landmark 06AI Analyst
Ask the data anything — it writes its own SQL
An AI analyst embedded in the dashboard: plain-language questions, in English or French, answered live from Salesforce data.
Writes & runs SQL liveA Claude-powered tool loop generates SQL against the semantic layer — read-only, guard-railed, row-capped.
Auditable, not a black boxEvery answer shows the exact SQL it ran, so any analyst can verify the result.
Beyond pre-built chartsHandles questions no dashboard anticipated — "which units get the most interest, and are they priced above similar ones?"
R&D Landmark 09Geospatial Intelligence
Neighborhood intelligence, built in-house
Interactive maps that replaced web-agency quotes — now a reusable data pipeline.
- Python pipeline: geocoding with bounds validation, brand-logo enrichment, photo matching, bilingual builds.
- 329 curated points of interest across two markets (Châteauguay QC · West Palm Beach FL) — each verified, including catching closed locations other sources still list.
- Walk-time rings, distance-sorted essentials, smart clustering, light/dark, FR/EN.
- The same map is the canvas the AI agent drives (Landmark 02).
R&D Landmark 10Immersive Web
Marketing sites engineered like products
- Real-time WebGL shader backdrops — generative "ink & silk" motion tuned to each brand, in light and dark.
- An Apple-style scroll-driven video technique: AVIF frame sequences painted to canvas — roughly half the payload of standard formats, engineered around browser decoding quirks.
- Modern-CSS R&D: view transitions, scroll-driven animations, container queries — all with graceful fallbacks.
