Projects

Problem: A 600-seat financial contact center had nine-minute queue times, SLA penalties, and new AI governance requirements that made cloud-heavy automation hard to justify.

Solution: An on-prem voice AI stack with streaming ASR, model cascade, RAG-backed answer grounding, safety controls, and guaranteed human escalation.

Problem: Teams were deploying inference services without a shared operating standard, which meant unstable latency, poor GPU utilization, and no reliable view of cost per request.

Solution: A production benchmarking and observability platform built around profiling, Prometheus metrics, Kubecost attribution, and model-level release decision support.

Problem: Thirty percent of catalog queries returned no results, p95 latency was above 1.5 seconds, and the cloud LLM bill was climbing toward an unsustainable operating model.

Solution: A hybrid retrieval stack with semantic search, BM25 fallback, reranking, and a fine-tuned Mistral-7B served through a cost-aware production path.

Problem: Media teams were losing roughly 30% of ad budget to manipulated Telegram channels, while manual review of 100 channels could take up to 25 analyst hours.

Solution: A hybrid fraud-detection stack combining rules, anomaly scoring, topic-aware thresholds, batch processing, and one-click feedback loops through a Telegram bot.

Problem: Search across more than 10 million products was too slow, too expensive, and too brittle to support commercial discovery at scale.

Solution: A multimodal retrieval and recommendation platform built around CLIP embeddings, FAISS HNSW, TensorRT optimization, reranking, and controlled rollout.