Systems, studies, and products built under real constraints.

NepSAUL: legal question answering for Nepali

Retrieval and grounded generation over 10,320 Supreme Court case laws, comparing lexical and multilingual dense retrieval before trusting an answer.

Best retrieval

91% Precision@1

Generation

85% truthfulness

Recognition

National hackathon winner

Modernizing a financial platform without breaking its contracts

Work across mobile, web, backend, database, deployment, and Linux hosts. Capabilities moved from a legacy Python backend into tested Go services while existing contracts and delivery paths stayed stable.

Chatterbox Nepali: adapting more than a checkpoint

Adapted a multilingual 500M speech model for Nepali, including Devanagari-aware text handling, alignment analysis, Apple Silicon training support, public weights, and app code.

Document intelligence inside identity onboarding

Multimodal extraction, canonical field mapping, date normalization, English-name handling, and tests around model output. AI handled narrow extraction; deterministic code owned the contract.

MEME-Fusion: what vision models miss in Devanagari memes

CLIP and BGE-M3 combined through cross-modal attention and a learned gate. Across eight configurations, explicit cross-modal reasoning improved F1-macro by 5.9 points over text-only baselines.

NEPSE market data, operated in public

Automated collection of prices, dividends, right shares, and daily floorsheets. Scheduled workflows publish versioned files and public API documentation without a permanent application server.

Coverage

365 companies

Adoption

10 stars · 8 forks

Delivery

2 releases

Privacy Aperture: desktop privacy controls with honest limits

Rust and Tauri prototype exploring application-aware overlays, display brightness control, crash recovery, and native macOS APIs.