AI/ML engineer with a software-engineering foundation. I work across the full path from model to product: training and fine-tuning deep learning models, building LLM-backed systems with retrieval and structured outputs, and deploying them as real services rather than leaving them in notebooks.
What I focus on
- LLM systems — RAG pipelines with hybrid retrieval and reranking, LoRA fine-tuning, prompt design for structured JSON outputs, hallucination guardrails, and evaluation that goes beyond eyeballing answers.
- Deep learning — CNNs and transformers in PyTorch, transfer learning, and computer vision including 3D perception on point clouds.
- Classical ML — gradient boosting, embedding models, and the unglamorous baselines that tell you whether a neural network is even worth training.
- Production engineering — FastAPI, WebSockets, Postgres, Docker, and React frontends, so the model reaches a user.
I also work on decision-making under uncertainty — Markov Decision Processes and Bayesian methods — in a research context.
LLM and retrieval systems
| Project | What it is |
|---|---|
| Speak Fill | Voice-dictated form filling. The browser streams 16 kHz PCM over a WebSocket to a FastAPI backend running faster-whisper in-process, reconstructing an interim/final transcript contract through polled re-transcription with bounded chunk commits. An LLM maps free-form speech onto an arbitrary form schema as {field_key: value} JSON, with an explicit null contract for low-confidence values and anti-cross-fill constraints. Runs against GPT-4o-mini or a local Qwen2.5-7B-Instruct behind one provider flag. |
| Spotify Song Recommendation | Word2Vec trained on playlists-as-sentences — 157K playlist sequences, a 2.7M-track vocabulary, recommendations by cosine similarity in embedding space. Postgres trigram search, live metadata enrichment, React 19 + Tailwind frontend. |
| Plant Identifier & AI Care Assistant | A fine-tuned EfficientNet identifies 50 houseplant species; an LLM generates location- and season-aware care advice with Pydantic-structured outputs, hallucination guardrails, and response caching. Deployed on Hugging Face Spaces. |
Modeling and data
| Project | What it is |
|---|---|
| Higgs Boson Detection | Deep feedforward network vs XGBoost on the UCI HIGGS dataset (28 features, 11M events). XGBoost led on AUC (0.81 vs 0.74) and calibration (Brier 0.18 vs 0.26); the network won on recall (0.77 vs 0.32) and F1 (0.76 vs 0.48) — a clean illustration of the precision/recall trade-off between model families. |
| LinkedIn Scraper | Python client covering 45 endpoints of the Fresh LinkedIn Scraper API — jobs, profiles, companies, posts, ads, groups — with input validation on every category. |
Deep learning & LLMs
Engineering
Background
Education
- M.Sc. Computer Engineering in Artificial Intelligence — University of Padova. Focus on computer vision, 3D perception, robotics, and decision-making models. Thesis on an MDP-based automated experiment-design framework.
- B.Sc. Computer Engineering — Persian Gulf University.
Academic projects
- 3D point cloud matching with PointNet — a modified PointNet architecture producing 3D descriptors for point-cloud alignment, with improved robustness in noisy environments.
- Food recognition and segmentation — a CV pipeline detecting and segmenting food items, comparing remaining against original portions.
Earlier career
Two and a half years as a senior mobile developer — React Native, Kotlin, Flutter — shipping to real users. It's why the ML work here tends to come with a working frontend attached.
Certifications
Deep Learning Specialization and Machine Learning Specialization, DeepLearning.AI.
Languages
English (professional working) · Persian (native) · Italian (basic)
Open to conversations about LLM systems, deep learning, and applied ML.