AdrianAnchuela

Problem

Military operations require instant, traceable answers grounded in massive corpora of regulations, technical manuals and historical reports — with zero tolerance for hallucinations.

Architecture & approach

• RAG architecture with hierarchical chunking — small chunks for retrieval, larger context windows for the LLM to reason over.
• Hybrid search combining dense vector retrieval (HNSW index) with BM25 keyword search, fused via Reciprocal Rank Fusion to capture both semantic intent and exact identifiers.
• Enterprise vector database deployed on OpenShift for full data sovereignty and on-prem security compliance.
• Continuous evaluation pipeline measuring Faithfulness, Context Recall and Answer Relevancy with RAGAS-style metrics to keep responses grounded in source material.

Problem

Pure quantitative models miss narrative-driven market moves. The system fuses structured time-series with unstructured text sentiment to capture both signals.

Architecture & approach

• Custom sentence-transformer embeddings to vectorise news headlines and social posts at scale — the same embedding model used for indexing and inference.
• Sentiment-aware feature engineering: aggregating embedding-space sentiment scores into time-series features alongside OHLCV price data.
• Sequence model (LSTM + attention) trained to learn the joint distribution of price action and narrative sentiment for trend forecasting.
• LangChain pipeline orchestrating news ingestion, embedding generation, sentiment scoring and prediction in a continuous loop.

Problem

AI and analytics pipelines are bottlenecked by data movement. Petabyte-scale audiovisual archives and energy telemetry need to move between systems faster than TCP can deliver.

Architecture & approach

• FASP protocol implementation bypassing TCP limitations — predictable throughput at line rate regardless of distance or latency.
• Container-native deployment on OpenShift with horizontal scaling and zero-downtime rollouts across hybrid cloud environments.
• Secure pipelines feeding downstream AI workloads — model training, semantic indexing of audiovisual archives, real-time energy telemetry analytics.
• Custom monitoring integrations surfacing transfer metrics into existing enterprise observability stacks.

Problem

Generic LLMs hallucinate on insurance policies and telecom contracts. The system needs grounded, brand-aligned answers across complex multi-turn dialogues.

Architecture & approach

• RAG pipeline over policy documents and product catalogues — semantic chunking by topic boundary rather than fixed token windows, preserving meaning across retrieval.
• Hybrid retrieval combining dense embeddings with keyword search to handle both abstract concepts and exact policy codes or SKUs.
• DPO (Direct Preference Optimization) fine-tuning on curated preference pairs to align the model with brand tone — without training a separate reward model.
• Continuous evaluation loop using Faithfulness and Answer Relevancy metrics to catch regressions before production deployment.
• Conversation memory layer maintaining intent across multi-turn flows for natural, contextual exchanges.

Problem

Engineers lose hours daily to repetitive form-filling, status updates and context switching across internal tools. Standard scripts break on edge cases — LLMs handle them.

Architecture & approach

• Browser-native extension intercepting page context and surfacing intelligent actions in real time.
• LLM-powered task classification routing user actions to the right internal system without manual lookup.
• Lightweight on-device parsing for sensitive data — no PII ever leaves the browser unless explicitly approved.
• Workflow orchestration layer chaining multi-step actions across disconnected enterprise tools.