All 31 projects.
- AI & NLPCase study
LLM post-editing for test-time domain adaptation
Keep the translation model frozen, let it draft, and have an LLM edit the draft for a specific domain. No retraining, and most of the fine-tuning gain.
87–92% gain recovered vs full fine-tuning
- Python
- Hugging Face
- NMT
- LLMs
- AI & NLPCase study
Clinical translation for English–Nepali
A domain-adapted medical translation pipeline comparing fine-tuned mBART and NLLB-200, with multi-metric evaluation and error profiling.
70k+ entries parallel corpus
- Python
- PyTorch
- Hugging Face
- mBART
- Data EngineeringCase study
Near real-time lakehouse
A medallion lakehouse with schema enforcement and quality gates, built so analytics and ML read from fresh, trustworthy tables.
- Python
- SQL
- PySpark
- Delta Lake
- AI & NLPCase study
RAG analytics assistant
An assistant that answers analytics questions from grounded retrieval, with routing and caching to control latency and cost.
- Python
- Embeddings
- Vector search
- Caching
- MLOps & CloudCase study
ML experimentation and regression framework
Config-driven experiments with versioned data, tracked metrics, and regression tests so model changes can't quietly get worse.
- Python
- Config-driven runs
- Metric tracking
- Drift checks
- Machine Learning GitHub
Neural combinatorial optimization with RL
Reproduced Google Brain's pointer-network + REINFORCE approach to the travelling salesman problem in PyTorch 2 on an A100, and tested whether a critic or moving-average baseline trains better.
1.6% from near-optimal with sampling; the critic halved the greedy gap (4.6% vs 9.0%).
- PyTorch
- Reinforcement learning
- Pointer networks
- A100
With Saina Zardosht
- AI & NLP GitHub
FFN ablation study in transformer LLMs
Trained a small GPT on Tiny Shakespeare and compared feed-forward designs (none, GELU 1x/4x, ReLU 4x, SwiGLU) on loss, perplexity, parameters, and convergence.
FFN width mattered more than activation; SwiGLU was the most parameter-efficient.
- PyTorch
- Transformers
- GPT
- Ablation design
- Data Engineering GitHub
Medallion architecture for capital markets
Bronze, silver, and gold layers for trading data: raw ingestion, cleaned tables, derived positions, and PnL views.
- Python
- DuckDB
- Parquet
- SQL
- Data Engineering GitHub
AutoTradeX: real-time market ingestion
Phase one of a multi-agent trading system: live market data streamed over WebSockets into Kafka, served through FastAPI.
- Kafka
- WebSockets
- FastAPI
- Python