experience

Work experience

Manana Labs

Engineer, Mar 2026 to Present

  1. Built a research agent for a neuroscience lab at Stanford. It plans an analysis, generates hypotheses, resolves each hypothesis, runs code in a sandbox, spawns parallel agents, reviews its own results, draws figures, and writes a report, like a real researcher. Now working on making it self-improving. Built on LangGraph and PydanticAI.
  2. Designed the architecture of an eval framework, in collaboration with the lab, to measure the performance of different stock agent harnesses (Claude Code, Codex, OpenCode) on neuroscience research tasks, against metrics like task score, consistency across repeated runs, cost, and wall time. Also built a Docker sandbox to run each agent without contamination.
  3. Built the supply-side energy data pipeline for Net Zero India, a collaboration with Princeton University and IIT Delhi to model how India can reach net zero emissions. Researchers define scenarios for energy assets, their demand and supply, and the pipeline runs them on an energy system model to find the lowest cost way to meet that demand.
  4. Researching and building agentic information retrieval. Pushing the limits of different retrieval mechanisms to understand which is a good solution for a given goal (latency, recall, precision, etc.). Approaches under research: BM25, agentic ripgrep, semantic RAG, and a fine-tuned SLM.
  5. Built a carefully curated eval suite for the retrieval work above. It is designed for debugging, not ranking: each group of cases isolates one way retrieval can fail, so when a case fails, it points at a cause instead of a number.
  6. Built an agent for a media monitoring team that queries different databases and knowledge sources, asks questions like a data scientist, and prepares a final report. Built using the Claude Agent SDK.

Niural AI

Software Engineer, Aug 2025 to Mar 2026

  1. Architected backend APIs for US payroll, automating year-end tax document generation across workflows moving millions of dollars, inside a large existing enterprise codebase.
  2. Built serverless services on AWS Lambda and DynamoDB, provisioned through CloudFormation and ECS, with Cognito, CloudWatch, S3, and ECR wired in for auth and observability.
  3. Shipped LLM-backed APIs that automate complex payroll logic, with the correctness checks needed for a domain where a wrong answer is a financial error.

MLExperts

Machine Learning Engineer, Mar 2023 to Aug 2025

  1. Led a document information extraction system on self-hosted LLMs with custom parser pipelines, reaching 90%+ accuracy on an internal benchmark built for the task.
  2. Fine-tuned EfficientNet and CNN architectures for screen and document understanding under real latency constraints.
  3. Owned end-to-end production rollouts for multiple ML features, and established the evaluation practice behind them: offline benchmarks, pre-production validation gates, and structured failure analysis on live traffic.
  4. Productionised unlearn-diff, an open-source PyPI library for machine unlearning in diffusion models, for a lab at Michigan State University.
  5. Built high-throughput backends with FastAPI, Django, Celery, and Redis for concurrent async workloads and horizontal scaling; ran CI/CD in GitHub Actions and CircleCI with automated container vulnerability scanning.

Hyperce LLC

Python Developer Intern, Nov 2022 to Feb 2023

  1. Built an image-processing proof of concept for the fashion industry, including a real-time AR try-on prototype in OpenCV and NumPy, and a published Hugging Face contribution.

Freelance

Python Developer, Dec 2019 -

  1. Upwork: exploratory data analysis, an embedding based RAG system, a LangChain agent for automated blog posting etc.
  2. Fiverr: Agentic chatbot, automations, AWS Deployment, etc.

Personal projects

  1. Aftersight (Python, OpenTelemetry, PyPI). Observability infrastructure for self-improving agents. Writes every agent run into the repository as plain files, so a coding agent can read its own history with the tools it already has (rg, jq) and find out why a run failed.
  2. ImagePrism (Python, ONNX Runtime, PyPI). Compare two images across several kinds of similarity in one call: pixel-level duplicate, same concept, same object, similar style, same person. Each is scored as its own dimension, so you pick the ones your problem needs. Runs on CPU, no PyTorch, no GPU, no API keys.
  3. NL2SQL (Python, FastAPI). Database agnostic natural language to SQL, built for databases with large schemas. Works in two steps: first find the relevant schemas and tables, then write the query using only those tables.
  4. Law Baje (Python, LangGraph, OpenAI). RAG system for Nepali law using hierarchical semantic search and hybrid retrieval, over 80% accuracy on document referencing. Cut hallucination rates against the base model through structured grounding and evaluation metrics written for the legal domain.

Open source contributions

  1. Unlearn Diff (PyTorch, diffusion models). Library for machine unlearning in diffusion models, with metrics quantifying forgetting quality against retained model utility.
  2. Agno, formerly Phidata. Resolved function-calling inconsistencies in the core multi-agent framework, improving reliability of the agentic execution layer.
  3. Hugging Face (huggingface.js). Added the task summary for mask generation to the Hugging Face tasks docs.