Ali Miraftab, Ph.D.

• Developed retrieval and ranking systems based on generative AI and transformer-based recommendation models; designed and implemented a RAG architecture with an additional contextual-intelligence layer to improve product recommendations, ranking quality, and relevance.
• Built parallel contextual recommendation architectures generating sales recommendations and talking points for Sales Reps and Customer Success Engineers, grounded in account history, product usage, and support signals.
• For the contextual-intelligence layer, applied topic modeling, taxonomy extraction, and semantic modeling with LLMs to enrich understanding, improve retrieval precision, and lift recommendation quality and downstream conversion.
• Architected and shipped an end-to-end LLM-powered Product Insights Platform that turns unstructured and structured customer signals (call notes, tickets, reviews, telemetry) into queryable, source-attributed findings.
• Pipeline: unstructured + structured sources → chunking → Vector Search embeddings → structured extraction with Llama 3.3 70B + JSON schema → queryable Delta table.
• Designed a domain-specific taxonomy mapping system (62 tools · 5 product solutions · 36 workflows · 11 outcome categories) with alias resolution and few-shot prompting — >85% tagging accuracy. Every finding is traced to a verbatim source quote for full auditability.
• Built a read-only MCP (Model Context Protocol) server exposing the insights table via SQL and natural-language-to-SQL translation — enabling ad-hoc analysis from Cursor, Claude, and Slack. Cut strategic-question turnaround from ~1 week of manual synthesis to minutes of querying.
• Implemented validation: confidence scoring, duplicate detection, source traceability, version control for extraction runs, and incremental processing.
• Deploy, fine-tune, and optimize production LLMs for sales and customer-success automation; apply quantization, kernel fusion, and dynamic batching to reduce latency and GPU spend.

• Led the end-to-end design of a personalized product recommendation and semantic-search platform for Dell.com and other sales channels; consumer-scale ML over structured and unstructured supply-chain data.
• Architected a two-stage retrieval-and-rank stack: Sentence-BERT bi-encoder for candidate generation and a cross-encoder re-ranker for relevance, with Word2Vec / GloVe and graph embeddings (GraphSAGE, GNN) over the part / SKU / product hierarchy.
• Designed an LLM-based entity-resolution pipeline for part / SKU / product deduplication across heterogeneous supply-chain records:
  • Retrieval: SBERT + HNSW / FAISS ANN search (O(log n) candidates at ~95% recall).
  • Ranking: gradient-boosted classifier over multi-feature similarity (cosine, Jaccard, token overlap, edit distance, length ratio, entity-type match).
  • Clustering: hierarchical agglomerative with complete linkage and confidence scoring; canonical-name selection per cluster.
  • Production: incremental batching, streaming-friendly DBSCAN fallback, and a typed Python EntityResolver API.
• Fine-tuned Llama and GPT models with LoRA / QLoRA / PEFT; built offline evaluation harnesses and online experimentation patterns; defined serving blueprints for downstream teams.
• Owned scope, roadmap, and milestone definitions for the Dell engagement; led ML engineers and scientists; partnered with Dell leadership on requirements and rollout.

• Built a multimodal semantic-search system on CLIP for automated quote / image generation; joint image + text embeddings powering personalized, conversational discovery.
• Applied instruction fine-tuning, LoRA / QLoRA PEFT, and structured prompt engineering for domain adaptation; designed an evaluation harness aligned with qualitative product goals.

Flights — Ranking, Recommendation, Personalization

• Designed and shipped Expedia's first personalized flight recommendation model on iOS SERP, connecting users to the most relevant flights under high-throughput / low-latency online serving constraints. +0.7% conversion rate. Algorithms: LambdaRank (LightGBM) and Two-Tower retrieval-ranking deep model (TensorFlow).
• Built the Web Flights LambdaRank LTR ranker to lift revenue. +1.2% GP/order.
• Predictive cache-fill DNN (TensorFlow) — 2× cache hit-rate, reducing downstream pricing-call volume and infra cost.
• CTR predictor gating flight pricing calls to reduce downstream traffic.
• Price forecasting and alerts using XGBoost — recommend the best time to book.
• Hero-image selection driven by a contextual bandit for SERP relevance.

Vacation Rentals & Lodging — Ranking and CV

• Neural ranker in PyTorch for Vacation Rentals SERP.
• Urgency-messaging XGBoost model predicting property / destination occupancy for personalization.
• Computer-vision stack feeding ranking and marketplace health: image aesthetic scoring (ResNet-152 / VGG-16), YOLO on-demand amenity detection, PyTorch scene classification (ResNet-152), fraud detection combining NLP + EAST text detector. Outputs powered image ordering, hero-image MAB, LTR features, marketplace-health analytics, and CVR boosting in email and ads.

• Fusing IoT contents with geosocial networks for anomalous-behavior detection in smart communities — won a $100K Cisco grant.
• Designed intelligent workload-aware schedulers for multi-cloud environments — resulted in U.S. Patent 10,452,451 B2.
• Published a comprehensive open-source TensorFlow tutorial covering computer vision and NLP for the UTSA ML6973 course.
• Side projects: EEG emotion classification, facial-image emotion detection, sentiment analysis on movie reviews, music classification with CNN + RNN, vehicle make/model recognition, marker detection.

• Designed WCDMA and LTE RAN systems in collaboration with Ericsson and Huawei; GSM / GPRS / EDGE network expansion and NSN license balancing.

• Designed end-to-end telecommunication systems (fiber, wireless, LAN, CCTV) for oil and gas sites.

• Collaborated with HITEC Poland on radiation-therapy solutions for cancer treatment.