Emanuel Otel

Working on Spirax Sarco's MiM platform, an enterprise AI assistant for internal knowledge retrieval, support, and document-based question answering across large technical document corpora.

• Led LLM observability for the MiM platform, implementing Langfuse trace instrumentation and helping the engineering team use traces for debugging, QA, latency analysis, and root-cause investigation.
• Built an end-to-end evaluation system for MiM, including golden datasets, LLM-as-judge scoring, question-level observability reports, KPI dashboards, and benchmark exports for PowerBI/Azure SQL.
• Designed MiM vs Microsoft 365 Copilot benchmarks using isolated per-question conversations, comparing answer quality, grounding, latency, failures, and user-facing behavior across STS and WM business units.
• Generated a traceable synthetic QA dataset from ~5,800 PDFs and ~60k pages, producing ~30,000 QA pairs with lineage from business unit -> category -> document -> page -> evidence block.
• Developed corpus processing and OCR pipelines with resumable page-level registries, Azure-hosted Mistral Document AI extraction, text-quality profiling, and recovery workflows for failed or oversized PDFs.
• Optimized full-corpus ingestion from roughly one week to ~4 hours, enabling faster iteration on retrieval, evaluation, and product-quality experiments.
• Ran retrieval benchmark sweeps across chunking strategies, embedding models, FAISS dense search, BM25, hybrid RRF, MMR, and hierarchical retrieval; reported retrieval quality, context precision, context waste, and token-cost trade-offs.
• Built internal QA and benchmark inspection UIs so product owners, SMEs, and engineers could review generated questions, source evidence, MiM/Copilot answers, judge scores, and failure rationales.
• Investigated bounded agentic retrieval for MiM workflows, showing stronger source coverage and answer quality than single-retrieval baselines for multi-document sales and technical support scenarios.

Led AI initiatives across two major clients:

Founding Engineer working directly with the CEO and CTO. Built and launched two AI products from scratch:

• ChatFusion - LLM-powered chatbot platform processing thousands of messages daily. Built the MVP, led deployment, and onboarded major clients, including one of the largest addiction treatment providers in the US.
• VoiceFusion - Speech-to-speech voice AI platform integrating STT, TTS, and LLMs, scaled to thousands of calls per month. Built the PoC and MVP that became the foundation for production voice AI deployments used by US lending companies.
• Built LLM agents with RAG, data extraction, call forwarding, function calling, and multi-step reasoning.
• Increased chatbot lead generation by 28%, improved engagement by 15% through A/B testing, and reduced LLM error rates by 23% via LLMOps, prompt evaluation, and curated fine-tuning workflows.
• Built internal tools for automation, data scraping, and prompt templating to accelerate development cycles.
• Worked with clients, Sales, Product, and Engineering to ship quickly, deliver measurable results, and close feedback loops.

• Implemented deep reinforcement learning models (VAE + LSTM + A3C) for visual navigation in mobile robots.
• Achieved up to 4x performance improvements through model and pipeline optimization.
• Improved image encoding for VAE-based architectures by up to 1% in data reconstruction quality.

• Built large-scale airport data scrapers with BeautifulSoup and Playwright for 300+ airports, automating ingestion into SQL pipelines.
• Delivered internal APIs and dashboards powering aviation pricing insights and decision tools.

• Developed control system software for industrial automation using DeltaV.
• Designed operator interfaces and ran SATs for international clients to validate safety-critical systems.