AI that works in operations, not just pilots. We build systems, software, and automation your team runs after we leave.
Execution friction
According to McKinsey, 70% of AI pilots never reach production. Gartner projects that by 2026, over 80% of enterprises will have deployed generative AI — up from fewer than 5% in early 2023. The bottleneck is operations, adoption, and sequencing — not model quality. Our 30-day delivery blueprint is designed to break through these barriers.
Common early signals
Support, reporting, and handoffs keep senior people in low-leverage tasks.
Impact: Margin gets squeezed while teams still feel understaffed.
Signal: Cycle time keeps increasing even after adding headcount.
Tools that ignore mixed data, approval gates, and Khmer-English handoffs fail after pilot.
Impact: Initial excitement fades once real teams own the workflow.
Signal: Usage spikes in week one, then plateaus or dies.
Without a ranked roadmap, teams cannot defend where budget and engineering should go first.
Impact: High-potential initiatives are delayed while low-impact work gets funded.
Signal: Priorities get reopened every sprint without commitment.
Strategic ecosystem
We select AI providers based on production reliability, multilingual capability, and cost efficiency for Southeast Asian operations. OpenAI and Anthropic power reasoning-heavy tasks. Meta Llama and Mistral enable self-hosted deployments where data sovereignty or cost control is critical — common requirements for Cambodian financial and government clients. Our infrastructure runs on Vercel and Cloudflare for edge performance across ASEAN, with PostgreSQL and Supabase for data persistence. This stack handles real ASEAN conditions: variable connectivity, multilingual content including Khmer, and compliance across jurisdictions.
AI Providers
Model and inference ecosystem
Technology
Core product and infrastructure stack
Partners
Delivery and integration partner layer
Build tracks
Four build tracks. One accountable partner. Systems your team owns from day one. As Harvard Business Review notes, organizations succeeding with AI focus on workflow integration, not model sophistication.
Track preview
Operational AI systems teams run daily.
Result focus
Reduce repetitive operational load while improving response speed and consistency.
Delivery mode
Fast discovery + iterative rollout, with measurable milestones every sprint.
Typical components
Trusted by teams at
8 clients
Delivery system
A focused sequence from diagnosis to deployment, built for execution speed. Each phase applies across our four build tracks.
30-day build loop
4 phasesDays 1-5
Map friction, blockers, and highest-leverage opportunities.
Structured discovery interviews across operations, product, and engineering. Each workflow is scored on automation potential, data readiness, and business impact using a weighted matrix.
Output: Opportunity map with ranked wins.
Auto loop
What happens in the call
00-05 min
Align on business goals and constraints
05-15 min
Identify high-friction workflows to fix first
15-25 min
Map architecture and realistic rollout path
25-30 min
Commit next steps, owners, and timeline
You leave with
AI Angkor Intelligence is a Phnom Penh-based consultancy for Cambodia and ASEAN teams. The World Bank projects Southeast Asia's digital economy will exceed $600 billion by 2030. We deliver systems that help teams capture that opportunity — in real workflows, not slide decks.
Nicolas Delrieu
Founder & Lead Consultant
12+ years across hospitality tech and applied AI. Former technology lead at NagaWorld Group. Active contributor to Khmer NLP open-source projects on Hugging Face. Focused on production systems teams can adopt and maintain independently.
Delivery results
3 examplesB2B services team — Cambodia
Challenge: Lead qualification and follow-up averaged 48-hour response time. The sales team manually triaged inbound leads from three channels, with no prioritization logic.
Shift: Deployed an AI triage agent using Anthropic Claude with retrieval-augmented generation (RAG) over the company CRM. The agent drafts contextual responses, scores lead quality, and routes to the right rep — with human approval before send.
Response time cut from 48 hours to under 4 hours — 12x improvement. Lead-to-meeting conversion increased 34%.
Operations-heavy SME — ASEAN
Challenge: Weekly reporting consumed 6+ hours of manual spreadsheet compilation. Four department heads submitted data in different formats, requiring manual reconciliation before executive review.
Shift: Built an automated data ingestion pipeline connecting accounting, inventory, and HR systems. A summarization layer generates executive-ready reports with anomaly highlighting using OpenAI GPT-4.
Reporting reduced from 6 hours to 15 minutes — 96% time savings. Anomaly detection caught a $12K billing discrepancy in the first month.
Customer support function — Phnom Penh
Challenge: Senior staff spent 2+ hours daily answering repeated Tier-1 questions across Khmer and English. Knowledge was trapped in individual inboxes and undocumented.
Shift: Deployed a bilingual (Khmer-English) knowledge assistant built on a Qdrant vector store with 1,200+ indexed FAQ entries. The system handles Khmer script queries natively with custom tokenization.
73% of Tier-1 tickets handled by AI, freeing 12+ senior hours per week. Customer satisfaction scores improved 18%.
Khmer NLP
The challenge
Khmer uses 33 consonants, 23+ dependent vowels, and has no whitespace between words. Standard tokenizers like BPE and WordPiece — trained primarily on Latin and CJK scripts — fail to segment Khmer text correctly. Off-the-shelf LLMs tokenize Khmer at 3-5x the cost of English, producing more tokens per word and degrading context window efficiency.
Training data bias compounds the problem: less than 0.01% of Common Crawl is Khmer. Hallucination rates spike when models encounter low-resource languages, and Khmer-English code-switching — common in Cambodian business communication — breaks context continuity in most commercial LLMs.
33
Khmer consonants
<0.01%
Khmer in Common Crawl
3-5x
Token cost vs English
Our approach
We build custom Khmer tokenization using ICU segmentation with domain-specific dictionaries for finance, hospitality, and government. Our fine-tuned embeddings for Khmer semantic search are hosted on Hugging Face, enabling open-source Khmer NLP research and production deployment.
RAG systems use Qdrant vector stores optimized for Khmer-English bilingual retrieval, handling code-switching natively. Our Khmer OCR pipeline processes invoices, contracts, and government documents — reducing error rates from 34% with generic OCR to under 5% with domain-tuned models.
Capabilities
FAQ
Direct answers so teams can decide quickly and move to implementation without ambiguity.
Still unsure?
Bring your current workflow. We will map a realistic first deployment path in one call.
Answer bank
7 itemsNo. Most engagements are designed for focused teams that need practical wins without enterprise overhead.
No. We start from your current data reality, then improve quality in stages while shipping usable outcomes.
Yes. Delivery is built for cross-functional teams across operations, product, and engineering.
You leave with a ranked opportunity map, an execution blueprint, and clear next actions your team can start immediately.
Most initial AI deployments follow our 30-day delivery blueprint: 5 days diagnosis, 7 days design, 10 days deploy, 8 days scale. Complex integrations may extend to 60-90 days.
We serve hospitality, finance, NGOs, retail, logistics, and professional services teams across Cambodia and ASEAN. Our custom AI systems are built for operational context, not theoretical models.
Yes. We have specific experience with Khmer NLP, Khmer OCR, and multilingual AI systems combining Khmer, English, and French for Cambodian business contexts.
Focused answers for teams planning real deployment.
Start here
30 minutes. You leave with a ranked opportunity map and a clear first deployment path.
Not ready for a call? Take the free 3-min AI Audit →
