A selection of our work across Applied AI projects and Research releases. Many engagements are confidential — what you see here is either public or approved for sharing.
Full voice pipeline (ASR → LLM → TTS) with custom fine-tuned models for Armenian, English, and Russian. On-prem deployment with sub-300ms end-to-end latency.
Conversational inbound agent replacing tier-1 support across 7 languages. Fine-tuned LLM with product knowledge RAG, live escalation routing, and real-time sentiment analysis.
Autonomous outbound calling agent for lead qualification and meeting booking. Handles objections, books calendar slots, logs structured call outcomes to CRM.
Industrial anomaly detection system deployed across 12 production lines. Custom vision model trained on client-specific defect taxonomy. Real-time alerting with <50ms inference.
Multi-camera traffic intelligence system for violation detection, flow analysis, and incident alerting. Deployed at scale across city infrastructure with edge inference.
Multimodal RAG pipeline over complex legal documents — contracts, precedents, filings. Fine-tuned for jurisdiction-specific reasoning and citation accuracy. On-prem Qwen backend.
Text-to-SQL system over a multi-schema financial data warehouse. Finance and operations teams query P&L, costs, and sales in plain English. Handles complex joins and time-series aggregations.
Personalized AI coaching system with long-term memory of goals, performance data, and past conversations. Structured reflection frameworks with behavioral nudges.
Full automation of 400k+ annual forms using visual document understanding. ColQwen2.5-based extraction with 99.3% field accuracy across Armenian and English.
Self-supervised computer vision for histopathology slide analysis. Trained on unlabeled WSI data with contrastive objectives. Reduced pathologist review time significantly.
Large-scale object detection pipeline for utility pole inventory from streetview imagery. Automated infrastructure audit replacing manual field surveys.
Real-time churn probability scoring with next-best-action recommendations. Integrated into CRM with automated retention campaign triggers.
Hierarchical time-series forecasting across 50k SKUs with external signal integration (seasonality, events, promotions). Reduced stockout rate and excess inventory simultaneously.
Multi-label text classification for unstructured claims narratives. Automated triage routing with urgency scoring, reducing manual review queue by 60%.
Joint Embedding Predictive Architecture trained for long-horizon planning in robotic manipulation tasks. Models predict useful latent representations of future states without step-by-step supervision.
Published methodology for adapting SOTA text embeddings to low-resource languages using only 10k noisy translations. Applied to Armenian, Georgian, Uzbek. ATE-1 & ATE-2 rank #1 on ArmBench-TextEmbed.
ColPali-style visual document retrieval embeddings. Held #1 globally on the ViDoRe benchmark. Open on Hugging Face. Research direction has since been superseded by Physical AI focus.