I build practical AI systems that solve real problems — AI integration, backend engineering and automation, turned into reliable production software. Systems that run efficiently, scale sensibly, and produce measurable results.
My work sits where research rigour meets systems ownership: retrieval and grounded generation, agent workflows, speech and low-resource NLP, and the unglamorous backend plumbing that keeps any of it alive once real users arrive.
Architecture, implementation and the pager after launch — one person, all three.
|
Ingestion, model routing, evaluation sets, caching, deployment, monitoring. Cost per run is measured, not guessed. |
Agents, workflows, document intelligence. Deterministic code owns contracts and permissions; humans approve expensive actions. |
Corpus ingestion, retrieval baselines, grounded generation. Citations that survive all the way to the user. |
|
Low-resource NLP, ASR and TTS adaptation, dataset design, multilingual comparison. Evaluation limits stated, not hidden. |
Go, Python, TypeScript, PostgreSQL, Docker, CI/CD, Linux. The layer AI features actually ship on. |
Nepali legal QA, multimodal hate detection, sentiment, profanity. 4 published papers — IEEE, ACL Anthology, LREC workshop. |
|
Retrieval plus grounded generation over 10,320 Supreme Court case laws. Lexical against multilingual dense retrieval, measured end to end.
|
Production fintech — details sanitized. Legacy Python capabilities migrated to tested Go services without breaking existing contracts. Mobile, web, backend, database, deploy.
|
|
Prices, dividends, right shares and daily floorsheets collected on a schedule and published as versioned open data.
|
500M multilingual speech model adapted for Nepali: Devanagari handling, alignment analysis, Apple Silicon training path. Public weights. |
|
CLIP and BGE-M3 joined through cross-modal attention and a learned gate; eight configurations ablated.
|
Merged, regulated workflow. Multimodal extraction with prompt design, canonical field mapping, date normalization and deterministic code validation. |
📝 NECPrepOpen-source NEC licence exam prep — mock tests, flashcards, local progress.
|
Native macOS privacy controls in Rust and Tauri, shipped with a tested-versus-planned capability matrix.
|
| Paper | Venue | Year |
|---|---|---|
| MEME-Fusion: Multimodal Ablation Study of Hate Detection and Sentiment Analysis on Nepali Memes | CHiPSAL @ LREC | 2026 |
| Retrieval Augmented Generation Framework for the Nepali Legal Domain Question Answering | IEEE ICNLP | 2026 |
| Evaluating Sentence Embedding Models for Nepali Sentiment Analysis | KUSET 18(2) | 2024 |
| Profanity and Offensiveness Detection in Nepali Using Bi-Directional LSTM | ICON — 21st Intl. Conf. on NLP | 2024 |
| Phase | What you get | |
|---|---|---|
| 1 | Audit — 1–2 weeks, fixed scope | Opportunity map, cost analysis, system design, delivery plan |
| 2 | Build — multi-week engineering | Implementation across AI, application, integration and infrastructure |
| 3 | Handover | Documentation, team training, cost exposure, explicit support boundaries |
Bring me the place where work, cost, or reliability is breaking down.
Website · Repositories · LinkedIn · Hugging Face · Kaggle · waglesameer5@gmail.com