Skip to content
Portrait of Nikhil Bhagat

Nikhil BhagatAI/ML engineer and NLP researcher, Biratnagar, Nepal

I build AI systems1 that have to be right,2 not just fluent.3

  1. 1In production

    24/7 customer-service agents on WhatsApp, Instagram and web chat for Kribaat, and HisabKitab, a WhatsApp accountant with about 1,600 automated tests in CI.

  2. 2Verified, not guessed

    NyayaSathi checks every citation in five stages. 0 fabricated citations on an adversarial set, Precision@10 0.871.

  3. 3Low-resource languages

    NEPVERSA, a version-aware Nepali statutory retrieval benchmark. Under review at ACM TALLIP.

Six years building production backends, machine learning and generative AI systems for clients in Hong Kong, Japan, Italy, the Philippines and Nepal. My research is on retrieval, evaluation and guardrails that keep large language models truthful in low-resource languages.

Download resumeEmail me

Work

  1. Oct 2024 – now

    Lead Senior AI Engineer (contract)

    Kribaat, Kathmandu

    • Autonomous customer-service agents for real estate agencies and restaurants, in Nepali, Romanized Nepali, English and Chinese. A verification gate checks every price, listing ID and booking code against live tool results before a reply is sent.
    • Fixed a production language-drift incident with per-user reply-language state, a three-locale message catalog and a rewrite guard that preserves figures, dates and negation.
    • Rehearsal Lab: an offline evaluation and reinforcement-learning environment where agents replay seeded bookkeeping scenarios, are scored by a deterministic judge and must beat production before human approval.

  2. Jun 2022 – now

    Senior Backend Engineer (contract)

    CineArtLink, Hong Kong (remote)

    • Lead backend for a professional network of film, media and theatre talent on Django REST Framework.
    • Docker microservices, engagement-based feed ranking, JWT auth, Redis caching and GitHub Actions CI/CD. Planning the move from Heroku to AWS and mentoring junior developers on test-driven development.

  3. Feb 2020 – Dec 2025

    Independent contract engineer

    Clients in Japan, Hong Kong, Italy, the Philippines and Nepal

    • Real-time GPS multiplayer game backend with PostGIS geofencing (Japan). Bilingual RAG assistant for Nepal Financial Reporting Standards (Nepal). Short-video community backend for 10k+ users (Hong Kong). WebSocket reliability fixes for a real-time Kanban board (Philippines).

  4. 2025

    Bachelor of Computer Application

    Tribhuvan University, Mahendra Morang Adarsh Multiple Campus

Research

NEPVERSA: a version-aware Nepali statutory retrieval benchmark derived from amendment footnotes

Bhagat, N. (2026). Preprint, Zenodo. Under review at ACM TALLIP. Mentor: Prof. Achut Raj Kattel, Tribhuvan University.

Laws change, and a legal assistant that answers with last year's text is confidently wrong. NEPVERSA asks retrieval systems what a provision said on a given date: 455 provisions from 5 Acts, 99 amendment operations and 181 point-in-time questions, of which 61.3% expect the system to abstain. Gold answers come from the amendment footnotes themselves, not from an LLM judge.

Recovering that history meant reading scanned Gazette notices and PDFs that extract to corrupted Unicode. I found 7 silent extraction defects, pinned each with mutation-tested regression tests, and shipped a one-command reproduction that re-verifies every published number by SHA-256.

Read the preprintCode and data

Projects

  • NyayaSathi

    Offline legal research assistant for Nepali advocates over 330 Supreme Court judgments and 7 statutes.

    0 fabricated citations on an adversarial set; Recall@5 0.968; runs on an 8 GB laptop

  • HisabKitab

    WhatsApp-native AI accountant for VAT-registered businesses. Reads bills from photos, applies Nepal VAT and TDS in exact integer paisa, saves only after the owner confirms.

    TypeScript microservices; MCP tool servers; about 1,600 tests in CI

  • AgentGuard

    Zero-trust security sidecar for AI coding agents. Inspects MCP and A2A tool calls for prompt injection, tool poisoning and secret leakage.

    Go, MIT licensed; under 5 ms p99 overhead; cosign-signed releases

  • PalikaBook

    Anti-corruption evidence platform for Nepal's 753 local governments. Anonymous, tamper-proof video evidence with a court-ready chain of custody.

    Go, Flutter, Next.js; Merkle-root verification; demo on request

  • SANU

    Nepali-English fine-tune of Qwen2.5-7B with QLoRA, released as GGUF for local inference. Its failures on legal questions led to NEPVERSA.

    QLoRA 4-bit NF4; Ollama and llama.cpp

  • AutoGPT contribution

    Merged pull request replacing a multiprocessing queue with a thread-safe queue, removing per-execution subprocess spawning.

    Merged upstream; PR #11618

AI and ML
LLM agents, tool calling, MCP, RAG, dense and BM25 retrieval, fine-tuning (QLoRA, LoRA), GRPO, evaluation and guardrails, PyTorch, Hugging Face, LangChain
Research
Low-resource and legal NLP, benchmark construction, evaluation design, reproducible pipelines
Backend
Python, Go, TypeScript, Django, FastAPI, Node.js, PostgreSQL, PostGIS, Redis, Kafka, NATS
Infrastructure
Docker, Kubernetes, AWS, Terraform, GitHub Actions, Linux

Resume

First page of Nikhil Bhagat's resume

Two pages: research, experience, projects and skills. Updated October 2026.

Download resume (PDF)

Ask Hauba

Hauba is an assistant trained on my work. Ask it about a project, the stack behind it, or whether I fit your role.

Nikhil Bhagat
Chat with Hauba
Online
Play a small game I made: a hiring manager chasing me across a 3D city