— AI Implementation · Pillar one —
AI that works on Tuesday morning.
Operational AI built for override, not against it. We frame the pain, pick the right place for the LLM, and ship in weeks not quarters.
— 01 / Selected AI work —
Four cases. Real numbers. Real architecture.
US WHOLESALER · 2025 · ANONYMIZED
Vendor onboarding, weeks to hours.
GPT-4o pipeline mapping inconsistent supplier files into a 120-column Acumatica master. Humans approve only the edge cases.GPT-4o · Python 3.11 · Pandas · Pydantic · n8n · PostgreSQL · Acumatica
Read the vendor-AI case → 60–75%effort reduction
per vendor onboarded
Weeks → hours cycle
120-column Acumatica master
80% automation target
284h scoped engagement
RETAIL SHELF · NATIONAL CHAIN · 2025 · ANONYMIZED
Per-store, per-shelf recommendations in seven days.
A national retail chain owner said: "Tell me what to put on each shelf, in each store, in the country." We translated that into a concrete optimisation surface — sell-through, freshness, category interaction — wrapped in an approvable dashboard so store managers stay in charge and override behaviour becomes data.GPT-class LLM · Python · KPI ingestion · approvable dashboard
Read the retail-shelf case → 7 daysvague brief →
working PoC
Per-shelf, per-store
Human-in-loop design
Override-as-data feedback
Generalises to any SKU
NAVON · MULTI-MODAL COMPLIANCE · 2025—NOW
60+ policy categories. Text, images, async video. One LLM call.
A regulated US user community needed real-time compliance scoring across text, images, and async video. We chose not to train a separate classifier. Instead: an LLM-as-policy-engine pattern — master prompt plus DB-stored per-user preferences. Single call returns answer plus structured compliance JSON across 60+ categories.GPT-class LLM · Node.js · PostgreSQL · REST + Swagger · WebSocket
Read the Navon case → 60+policy categories
scored per call
Two delivery teams
One interface contract
Parallel sprint cadence
Lead architect
SKARB.AI · KCG PRODUCT · 2025—NOW
Production GraphRAG family archive. Shipped solo in ten weeks.
Built for ourselves so we know what shipping production AI actually takes. GraphRAG-style hybrid retrieval, voyage-3-large 1024d multilingual embeddings, pgvector + HNSW, smart routing, encrypted memory, eval harness that killed BM25+RRF and parked Cohere reranker.GPT-4.1 / 4.1-mini · voyage-3-large · pgvector · Supabase · React · Capacitor
Read the Skarb.ai case → 10 weeksstart to shipped
web + iOS + Stripe
53 screens · 23 tables
15 edge functions
2,714 tests
EN / RU / PL trilingual
— 02 / How we approach it —
Four principles. Earned the hard way.
01
Frame the pain
Translating a vague brief into a concrete optimisation surface is the hardest part of the job. We do it before we touch a model.
Discovery is its own deliverable
→
02
Pick the right place
Choose where the LLM earns its place — and just as importantly, where to keep it out. Most AI failures start with the wrong location.
Hybrid > pure LLM, every time
→
03
Build for override
Humans stay in charge. That's not a limitation, it's the design. The override is data the system learns from.
The only design that survives reality
→
04
Ship in days
Working PoC in one week. Production engagement in weeks. Quarters are for theatre. We learn from contact with reality, fast.
Velocity is a quality signal
→