Biratnagar, Koshi Zone, Nepal • [email protected]
I build machine-learning systems from first principles through deployment, with a software engineering background and a habit of testing my own results honestly (baselines, leakage-safe validation, error analysis).
My work spans PyTorch, Transformer architectures, tokenization, ML inference, data pipelines, LLM applications, FastAPI backends, and Docker-based deployment. I like understanding how ML systems work underneath high-level abstractions and turning those implementations into tested, usable software.
🎓 Education: BCA, Tribhuvan University. Final-semester examinations completed August 2026
🔬 Experience: ML Internship at FlyRank AI, published capstone on data leakage and search intelligence
📚 Currently: Preparing for graduate studies in Computer Science / Data Science, and continuing to build open-source
ML projects
- 🧠 Machine Learning & Deep Learning — PyTorch, TensorFlow, Scikit-Learn, Transformer architectures
- 🤖 LLMs & NLP — tokenization, BPE, RAG, AI agents, Hugging Face
- ⚡ ML Inference — ONNX Runtime, INT8 quantization, CPU inference, memory optimization
- 📊 Data Processing & Pipelines — Apache Spark, Delta Lake, DuckDB, Airflow, dbt, PostgreSQL
- 🔧 Backend, MLOps & Deployment — FastAPI, Docker, Kubernetes, CI/CD, automated testing, PyPI packaging
- 🌐 Applications — React, Next.js, Streamlit, Metabase, Power BI
| Category | Technologies |
|---|---|
| Languages | Python, TypeScript, SQL, C#, Dart |
| Machine Learning & AI | PyTorch, TensorFlow, Scikit-Learn, SciPy, Statsmodels, NumPy, Pandas, OpenCV, Hugging Face |
| LLM / Inference | Transformers, ONNX Runtime, FAISS, FlashAttention, BPE Tokenization |
| Data & BI | Apache Spark, Delta Lake, DuckDB, PostgreSQL, BigQuery, Airflow, dbt, Metabase, Power BI |
| Backend & MLOps | FastAPI, Docker, Kubernetes, GitHub Actions, Pytest, Pydantic, MLflow, Prefect, Prometheus |
| Frontend & Mobile | React, Next.js, TailwindCSS, Streamlit, Flutter |
Decision-support ML system for SEO prioritization (FlyRank capstone research paper).
- Engineered feature pipelines across a ~79 million row production warehouse using DuckDB, and trained the evaluation-first model on a focused 22,006-row analytical slice.
- Identified a client-overlap leakage trap: a naive split gave an inflated 0.940 precision. A strict client-grouped holdout split revealed the honest 0.640 baseline.
- Audited model limitations, documenting single-seed variance and absolute opportunity size constraints (zero-click boundary cases).
- Published the methodology, leakage audit, and results as a research paper.
Transformer implementation, training, optimization, and lightweight inference.
- Scaled the architecture to roughly 28M parameters and trained on TinyStories with gradient accumulation to fit ~6 GB VRAM.
- Exported to ONNX with INT8 dynamic quantization for lightweight inference through FastAPI.
Medallion lakehouse ETL/ELT with idempotent incremental processing, quality quarantine, and governed serving.
- Docker-first Apache Spark + Delta Lake pipeline through Landing, Bronze, Silver, and Gold layers.
- Declarative YAML quality engine that routes invalid records to quarantine tables with no silent drops.
- Airflow with a lightweight SQLite profile for testing and a PostgreSQL + LocalExecutor profile for backfills.
- Gold data served by FastAPI + DuckDB with RBAC, monitored with Prometheus and Grafana.
Horizontally scalable, latency-optimized model serving on Kubernetes.
- ONNX Runtime with 8-bit dynamic quantization to cut container memory and CPU latency.
- Locust load tests verifying exactly 0 dropped requests during live RollingUpdates.
Containerized inference API and native SDK with zero boilerplate.
- Serves
.joblib/.pklmodels fetched over HTTP at startup via environment variables. - Published on PyPI with strict input validation.
Decision-support analytics stack: governed KPIs, reconciled data, and rigorous A/B testing.
- Containerized PostgreSQL + dbt + Metabase pipeline with uniqueness, non-null, and referential tests.
- A/B analysis with
scipyandstatsmodels: Chi-Square SRM gate, 2-proportion Z-tests, p-values, and 95% CIs (28.7% relative conversion lift in the demo data).
| Project | What it is | Highlights |
|---|---|---|
| StatTest-Pro | A/B testing framework | Sample-size planning, SRM halt at p < 0.01, published on PyPI, Pytest + GitHub Actions |
| LexiByte | BPE tokenizer from scratch | GPT-style Unicode-aware regex pre-tokenization, pip install lexibyte |
| Aegis Omnisearch Agent | Lightweight RAG agent | Custom ReAct loop with Gemini, local FAISS retrieval, INT8 ONNX CPU inference |
| ZeroProp Engine | Neural net in pure NumPy | Dense layers, ReLU, Softmax Cross-Entropy, live React + Canvas training dashboard over WebSockets |
| FlowTrace | DAG-orchestrated ML pipeline | Prefect DAG, Pandera gates that halt training on bad data |
| TabTrace | Reproducible tabular ML | Feature-justification decorator registry, 5-fold stratified CV, LogReg vs. Random Forest |
| Customer Churn Risk Intelligence | MLOps churn pipeline | MLflow tracking and registry, FastAPI inference, Skops instead of pickle |
| Multimodal Phishing Detection | URL + linguistic fusion | ISCX URL features and PhiUSIIL text features with soft-voting fusion |
| OverfitLab | Overfitting diagnosis | Dropout (p=0.5) and L2 weight decay in PyTorch |
| LunarLander-v2 Agent | RL agent | Trained with PPO |
| DataMart-Flex | Star-schema data mart | 100k+ records, Kimball model, DuckDB, Power BI DAX measures |
| CohortLTV-Engine | Retention and LTV analytics | 1M+ rows in ~174 ms with DuckDB window functions and CTEs |
| ExecPulse-BI | Sales BI dashboard | 50k+ transactions to a star schema, SQLite via SQLAlchemy |
| InsightStory-EDA | Automated data storytelling | 9 DuckDB SQL queries, auto-generated 5-slide .pptx / .pdf deck |
| DataCleanse-Lite | Multi-source ETL | 100k messy records in ~2.24 s, Pydantic contracts, ~21% quarantined |
AI / ML Engineering Intern — FlyRank AI | Jul 2026 – Sep 2026
- Built a CTR Opportunity Scoring model to prioritize SEO metadata reviews, engineering pipelines on ~79M Parquet rows with DuckDB and training on a focused 22k-row slice.
- Identified and mitigated client-overlap leakage with grouped validation, and published a research paper.
Data Science & ML Apprentice — Skill Shikshya | Apr 2026 – Jun 2026
- Covered ML mathematics, classical ML, and deep learning; built and containerized FastAPI ML applications with Docker.
Software Engineering Internships & Contracts (alongside BCA)
- Walkers Hive, Full-Stack Intern (Oct – Dec 2025): designed the AcademiaOS MVP with FastAPI, Celery, and Next.js.
- Nextwave Technology, Software Engineer (Apr – Jul 2025): features, fixes, UI revamp, and Play Store launch of the Academia mobile app.
- Walkers Hive, Software Engineer (Nov 2024 – Apr 2025): e-commerce admin panel with React, MUI, and Redux-Saga.
- CodSoft, Android Development Intern (Dec 2023 – Jan 2024): Flutter apps with Firebase Auth and BLoC.
Bachelor of Computer Application (BCA) — Nihareeka College of Management and Information Technology, Tribhuvan University, Nepal Final-semester examinations completed August 2026
- FlyRank AI — Machine Learning Internship Certificate
- Skill Shikshya — Data Science & ML Diploma
- Anthropic Academy: Claude Code in Action · Building with the Claude API · MCP Advanced Topics · Claude on Amazon Bedrock · Claude on Google Vertex AI
- Kaggle: Pandas, Feature Engineering, Intro to ML, Intermediate ML
- Email: [email protected]
- LinkedIn: linkedin.com/in/bibek-dhakal-771ba5334
- Portfolio: bibek-dhakal-fr.vercel.app