01
LLM-powered product feature
Chat, search, summarisation, drafting — built into your existing product. Honest about where the model earns its keep.
References: Notion AI · Linear · Intercom FinSee it →Three directions for shipping AI.
“AI feature” is too broad. Here are the three concrete shapes we'd shortlist for your product — LLM in the surface, predictive numbers, or specialty models on your data. Each one is a real engagement, not a slideware.
02
Predictive analytics
Forecasting, churn, demand, anomaly detection. Trained on your data, deployed where your team works.
References: Stripe Radar · Spotify ML · DoorDash ETASee it →03
Specialty / custom ML
Computer vision, NLP pipelines, recommendation systems — domain-specific models trained on your data.
References: Pinterest Visual · TikTok rec · Tesla AutopilotSee it →Direction
01
LLM-powered product feature
01 · LLM in the surface
AI inside the product.
Chat, search, summarisation, drafting — built into your existing app. Grounded in your data, gated by approvals, costed properly. Honest about where the model earns its keep and where it doesn't.
Sample moments · what the user sees
Assist
Inline drafting
Cursor-style autocomplete for the right surfaces. Not every input box — only where it earns the latency.
Trust
Grounded answers
RAG against your docs, citations on every claim. The user can verify what the model said.
Sane
Cost guardrails
Per-user budgets, model fallbacks, cache hit-rate dashboards. The CFO conversation, pre-empted.
Typography
Inter + JetBrains Mono
Palette
Lavender, deep violet, electric purple
Voice
Confident, useful, never magical.
Best for
SaaS adding intelligence to existing UX without an in-house ML team.
Direction
02
Predictive analytics
02 · Numbers that forecast
Numbers that forecast. Not just report.
Forecasting, churn modelling, demand prediction, anomaly detection. Trained on your data, deployed where your team already works — Slack, dashboard, email.
Sample moments · what the user sees
Predict
Forecasting
Demand, churn, revenue — with confidence intervals, not just point estimates.
Watch
Anomaly alerts
Slack pings when something's odd. Not just “number went up.” Why.
Embed
Slack-native
Numbers where the team already lives. No new dashboard tab to forget.
Typography
Inter — data-clean, no decoration
Palette
Amber, deep brown, signal
Voice
Numeric, honest, useful.
Best for
Data-rich businesses making operational decisions: ops, risk, supply chain.
Direction
03
Specialty / custom ML
03 · Specialty models
Models trained on your data.
Computer vision, NLP pipelines, recommendation systems — domain-specific models trained on your data, not just an API call. Edge-deployable, on-device when the latency matters.
Sample moments · what the user sees
Vision
CV pipelines
Detection, classification, OCR — trained, evaluated, monitored.
Match
Recommendation
User-item, content-based, hybrid. Cold-start handled honestly.
Edge
On-device
TensorFlow Lite, ONNX — runs at the edge. Sub-100ms inference.
Typography
Inter + JetBrains Mono
Palette
Mint, deep forest, signal green
Voice
Technical, depth-first, honest about limits.
Best for
Edge / on-device, regulated domains, proprietary data, real-time inference.
Pick a direction. We'll build it.
“AI” means three different engagements with three different teams, costs, and risks. Tell us the problem and we'll match it to the right shape — never the most expensive one.