~/services/ai
AI Systems: From PoC to Production
Enterprise RAG, LLMs, and NLP solutions for real business impact: architecture consulting, rapid PoC development, and production deployment with predictable operating costs.
01 /
Service Capabilities
MODULE 01
Chatbots & Enterprise Assistants
- ✓ Retrieval on enterprise knowledge bases
- ✓ Slack, Web, and messaging integrations
- ✓ Safety guardrails & inference cost control
MODULE 02
RAG & Semantic Search
- ✓ Dense Vector + BM25 Hybrid Search & Re-ranking
- ✓ Document, Image, and Video indexing
- ✓ Retrieval accuracy benchmarking
MODULE 03
Fine-tuning & Model Evaluation
- ✓ LoRA / SFT on proprietary data
- ✓ Quantitative baseline comparison benchmarks
- ✓ Inference speed optimization & self-hosting
02 /
Case Studies
NexusAI: Enterprise Multimodal RAG & Inverted HyDE
Intelligent Knowledge Retrieval Across Enterprise Documents
Enterprise document repositories (PDFs, Office files, images, videos) often suffer from vocabulary mismatch between user queries and raw documents. NexusAI introduces Inverted HyDE (pre-generating hypothetical queries during offline indexing to map retrieval into Query-to-Query matching, eliminating runtime LLM latency) combined with hierarchical document chunking and grounded generation with precise source citations.
Inverted HyDE
Query-to-Query matching eliminates runtime latency
Multimodal
Processes PDF, Office, Image & Video files
100% Grounded
Direct source citations in generated responses
SYSTEM ARCHITECTURE
Document Parsing & Hierarchy
Preserve hierarchical structures (pdf, images, videos, office files)
Inverted HyDE Indexing
Generate hypothetical queries offline (Query-to-Query matching)
Sub-second Retrieval
Dense Vector + BM25 + Cross-Encoder Re-rank
Grounded Answer Generation
Grounded responses with accurate citations, zero hallucination
LinguistAI: LLM + speech
Automated 4-Skill English Mock Exams & Grading
- VẤN ĐỀ
- 4-skill exam preparation is bottlenecked by evaluation: speaking and writing require human teachers, causing students to wait days for feedback.
- CÁCH LÀM
- Generate reading tasks using LLMs aligned with exam formats, evaluate speaking using Whisper plus fluency & pronunciation scoring, and grade writing using LLMs with standardized rubrics.
- KẾT QUẢ
- Students receive instant, detailed criteria feedback right after submission, reducing teacher workload to auditing flagged edge cases.
GPT-4ClaudeWhisperReact
4 Skillsauto-graded with instant feedback
KiKi (Zalo): speech
Text Normalization for Vietnamese Automatic Speech Recognition
- VẤN ĐỀ
- In-car virtual assistants misrecognized non-verbatim spoken expressions: numbers, dates, units, street names, and proper nouns.
- CÁCH LÀM
- Built text normalization pipelines for training data and ASR outputs, expanded pronunciation dictionaries for proper nouns, and constructed category-specific evaluation benchmarks.
- KẾT QUẢ
- Increased ASR accuracy by 12% with significant reduction in numerical and location errors. Simultaneously optimized Kubernetes inference services, reducing service errors by 15%.
PyTorchKaldiPythonKubernetes
+12%ASR accuracy boost
03 /
Process & Engagement
$ ./engagement
01
Architecture Advisory
Assess business domain requirements, model selection, and stack budgeting.
02
2–4 Week PoC
Functional prototype on real enterprise data with evaluation metrics.
03
Production Deployment
Cloud deployment, cost/quality observability, and full documentation handoff.