I put language models on top of structured data — and then measure whether they actually work. Most of what I build ships with a benchmark, an ablation, and a written account of where it fails.

• 🏦  AI Engineer intern
• 🔍  Focus: RAG & knowledge graphs, information retrieval, evaluation of LLM systems
• 🛠  Also ship: FastAPI services, Dockerised deployments, PyTorch training pipelines
• 📍  Sofia, Bulgaria — open to AI/ML engineering roles

Languages  

ML & AI  

Backend & Data  

An LLM document-processing service that actually runs as a service.

PDF invoices in, structured expense reports out. Vision-model OCR with a pdfplumber fallback, an LLM enrichment stage that categorises line items and corrects OCR noise, Pydantic-validated persistence to PostgreSQL, and a generated PDF report. Exposed both as a REST API and a web UI; the whole stack comes up with one docker compose command.

FastAPI · PostgreSQL · SQLAlchemy · Pydantic v2 · Groq · Docker Compose · Streamlit

Can a small local LLM query an ontology reliably? Measured, not guessed.

A local Qwen2.5-Coder model translates plain-language questions into SPARQL against an OWL2-RL geographic ontology in GraphDB. The interesting part is the evaluation harness: 22 questions across 7 reasoning categories, 10 runs each, scored against hand-written reference queries executed live against the triplestore — no hardcoded expected values.

Result: 82% exact-match at 3B, 93% at 7B, with every failure mode traced to a root cause (dropped FILTERs, abandoned MINUS, entity-name abbreviation, reasoner limitations under the open-world assumption).

📄 Full methodology & benchmark report

Ollama · GraphDB · OWL2-RL · SPARQL · owlready2 · Python

A controlled architecture comparison, written up as a paper.

Four models — 2 architectures × 2 tokenisation schemes — trained on ~38K MIDI files with parameter counts matched within ±10% and an artist-stratified split so no artist appears in both train and test. Includes a custom 532-token event tokeniser alongside a REMI+ wrapper.

Finding: raw perplexity isn't comparable across tokenisers, so evaluation uses bits per second of musical time. Measuring loss per position bucket shows the Transformer's advantage over the LSTM grows ~8.5× across the 512-token context window — long-range dependency access, quantified.

📄 Paper (PDF) · 🤗 Trained checkpoints

PyTorch · MidiTok · flash attention · Kaggle / Lightning AI

Find the song from lyrics you misheard.

Lyrics are converted to IPA phonemes and indexed as bigrams, then queries pass through a four-stage retrieval pipeline — inverted-index lookup → Jaccard filtering → TF-IDF reranking → weighted Levenshtein distance using a phoneme substitution-cost matrix — with query-length-adaptive thresholds and optional one-hop query expansion. Benchmarked against SoundEx and Metaphone baselines on Accuracy@1, Recall@K and MRR.

Built with @Bifrost19 as MSc coursework in Information Retrieval.

Python · eng_to_ipa · custom TF-IDF · pandas

Do POS and NER priors help graph ranking or matrix decomposition more?

TextRank (PageRank power iteration) and LSA (TF-IDF + SVD), both implemented from scratch in NumPy and both extended with sentence priors derived from POS and NER tags, blended by a tunable α. Evaluated with ROUGE-1/2/L across a grid over α and summary length, on a morphologically rich language with Bulgarian-specific stemming and stopword handling.

NumPy · CoNLL-U Plus · BulStem · ROUGE

Unsupervised clustering turned into rules a human can read.

Fuzzy C-Means with random restarts discovers latent archetypes in sensory beer profiles; adaptive Gaussian membership functions then convert cluster centroids into linguistic IF–THEN rules. Product t-norm activation is computed in log-space to avoid underflow in high dimensions, and the rule-based approximation is validated against the original FCM model (Pearson r > 0.90) rather than assumed faithful.

Fuzzy C-Means · PCA · NumPy · Matplotlib

Fundamentals, implemented for speed rather than for the textbook.

IDA* with Manhattan distance for the N-Puzzle, MinConflicts for N-Queens, a genetic algorithm with order crossover for TSP, minimax with alpha-beta pruning, plus K-Means, Naive Bayes and decision trees — no libraries. Optimised with O(1) conflict frequency arrays, delta-distance mutation updates and cache-friendly memory layout.

Benchmark: 10,000-queens MinConflicts solved in 0.34s; 15-Puzzle solved by IDA* effectively instantly.

C++17 · -O3 · std::mt19937

MSc — Information Retrieval & Knowledge Discovery · Sofia University "St. Kliment Ohridski", FMI Deep learning with PyTorch, recommender systems, knowledge bases & ontologies, NLP, knowledge discovery from data.

BSc — Computer Science · Sofia University "St. Kliment Ohridski", FMI Algorithms & data structures, object-oriented programming, databases, operating systems, computer networks, software engineering, discrete mathematics & probability.