Work Experience

Built an image-to-product retrieval system with a dual-encoder model combining CLIP-based visual embeddings and textual product representations, reaching 92% top-10 accuracy and enabling visual search from uploaded or shared images.
Trained a Query Understanding model on user query logs and click data to infer intent and reformulate low-confidence queries, reducing zero-click searches by 4%.
Improved search relevance through multi-metric fine-tuning and category-consistent retrieval, lifting conversion rate by 4%; ran A/B experiments routing each query between semantic and Elasticsearch pipelines, increasing add-to-cart by 3%.
Extended image search to video input with an FFmpeg transcoding, keyframe extraction and scene detection pipeline, improving retrieval relevance by 15%.
Built an agentic LLM quality-control and ticket-resolution system with Pydantic AI and LangChain, using an LLM-as-judge harness and Langfuse tracing to score and autonomously answer support tickets.
Established LLMOps practices — prompt evaluation harnesses, Langfuse observability, and token-cost dashboards in Grafana — cutting LLM inference costs by 25%.

Built LSTM and Prophet price-forecasting models achieving 12-18% MAPE across major coins such as BTC and ETH.
Developed a recommendation system with LightGBM and behavioral KMeans user clustering, increasing simulated ROI by 9.3%.
Designed portfolio optimization combining Modern Portfolio Theory and Deep Q-Learning to maximize Sharpe ratio across 20+ cryptocurrencies.
Integrated Monte Carlo simulations for profit expectation; the prototype outperformed an equal-weighted baseline by 15% in backtesting.

Co-built an end-to-end MLOps pipeline to train, version, and deploy models using Airflow, Spark, MLflow, Katib, Feast, TensorFlow Serving, FastAPI, Kafka Streams and GitLab CI/CD, reducing training and deployment time by 70%.
Built a recommendation system to infer speed for streets lacking sufficient data, expanding coverage from 1M to 3M shared streets.
Launched an ETA system across 5+ cities in Iran and Iraq, improving R² by 20%, and reduced ETA MAPE by 5% with street-speed forecasting models.
Implemented an HMM map-matching algorithm to align driver GPS probes to streets and compute per-driver speed.
Developed a Golang microservice to benchmark model accuracy in real time with Prometheus and Grafana, speeding up QA by 80%.

Integrated vector database solutions for efficient similarity search to surface related items in Snapp Shop, increasing conversion rate by 5%.
Fine-tuned and deployed a pre-trained OCR model (EasyOCR) to read ID cards in the driver-signup flow, cutting signup time from days to hours.
Optimized a transformer model with ONNX to boost inference speed by 10% and decouple training from serving.
Engineered a sentiment analysis service using SVM to analyze over 10,000 tweets daily for real-time insight into public sentiment.
Mentored over 5 new joiners and launched a structured mentorship program and a new interview pipeline.

Fine-tuned and deployed LLM models on GPU to translate text between the support team and foreign customers.
Developed a CRF-based NER model powering an address search engine, increasing successful searches by 15%.
Built and deployed a stacked LSTM model to forecast stock and cryptocurrency values, achieving 87% prediction accuracy.
Designed a type-ahead search system for stock lookup using prefix matching over a custom trie, reducing stock search time by 30%.
Developed a BERT-based chatbot to answer stock inquiries, driving a 20% increase in user engagement.

Deployed neural-network time-series forecasting on Raspberry Pi 4 and decision-tree classification on ARM Cortex-M52.
Led the design and implementation of an event detection service to alert on a patient's abnormal behavior, achieving a 0.95 F1 score.
Employed TensorFlow Lite to reduce memory usage by 50%, enabling on-device inference on mobile phones.
Launched a Kafka pipeline to ingest sensor data via Protobuf into a data lake, capable of handling 20k+ messages per second.
Implemented Kalman filtering to enhance GPS positioning by 10%.