Arpan Ghoshal

MS in Computing, Entrepreneurship and Innovation, Courant Institute of Mathematical Sciences & Stern School of Business

Completed the two-year program in one year; research assistant in deep learning and applied AI.

BTech in Computer Science and Engineering, Data Science Specialization

Independent AI Engineering & Consulting, Europe & United States, Remote

• Engaged by 3 startup clients across Europe and the US to upgrade production AI systems, working directly with founders from scoping through architecture, delivery and handover.
• Audited and re-architected existing LLM and RAG pipelines, reducing inference and infrastructure costs while improving answer accuracy and system efficiency.
• Built evaluation suites (golden datasets, LLM-as-judge scoring, regression tests) to benchmark each client’s AI features, then used the results to ship substantial feature improvements.
• Delivered AI agents, MCP servers and workflow automation on top of the upgraded stacks.
• Retained by Operating Equity Partners at the founder’s request to maintain and extend the AI systems built for their ventures.

Operating Equity Partners (Vivid Labs Holdings, Inc.), San Carlos, CA, Remote

• Led an international team of 12 across design, product and engineering, running the company end to end and partnering directly with Halsey Minor (co-founder of CNET, Salesforce, Google Voice and NBCi).
• Built both 0-to-1 AI products end to end, a trusted news network that fights misinformation and an AI avatar legal assistant, owning the architecture, LLM pipelines, retrieval, backend and integrations.
• Secured continued investor funding for the ventures by demoing the working products built.
• Built the agent reasoning layer for legal and news-verification workflows: multi-step investigation, evidence retrieval, tool-augmented reasoning and typed structured outputs.
• Designed the eval and observability stack: golden test sets, LLM-as-judge and human-review scoring for factuality and hallucination, output guardrails, and LangSmith/Langfuse tracing of quality, latency and cost per request, used to catch regressions before every release.
• Fine-tuned open-source LLMs for legal and news use cases and shipped them as product features.

Gisual Inc., Philadelphia, PA

• Built the first commercial machine learning model for power outage detection, achieving 96% accuracy and owning the pipeline from geospatial data acquisition and feature engineering through deployment for utility clients.
• Designed a custom outage-detection architecture combining mathematical modeling, deep learning, and ensemble methods to capture relationships across geospatial, weather, grid, infrastructure, and device data sources.
• Built forecasting models to predict outage patterns and demand trends, improving proactive resource allocation for utility clients.
• Designed and shipped a GenAI-powered outage detection system that analyzed social media signals in real time to identify outage events ahead of conventional monitoring systems.
• Deployed solutions across US and international markets with monitoring dashboards, automated retraining workflows, and foundations for a fiber outage detection model.

Hitch Works (acquired by ServiceNow), Seattle, WA, Remote

• Architected an AWS-based machine learning pipeline using SageMaker, Lambda, and S3 for 24/7 web crawling, ingestion, and large-scale text preprocessing.
• Developed a BERT-based semantic matching model for skills-to-context mapping, improving matching accuracy by 170% over baseline for the workforce intelligence product used by enterprise talent teams.

Brainentech Neuroscience, Bangalore, India

• Built and deployed a real-time AI-powered audio analysis API serving 150k+ users, training NLP models to extract 20+ engagement metrics.
• Applied Canonical Correlation Analysis to EEG and audio stimulus data for the company’s neuroscience product.

NayaOne, London, UK, Remote

• Built the company’s first data infrastructure and data APIs from scratch; later adopted by major fintech clients on the platform.

RoBERTa-based emotion detection model; top trending emotion model on Hugging Face (2020–2022) with 15M+ downloads and 100+ citations; published at IEEE CONECCT 2022.

Execution-safety layer for AI-agent actions: binds a human approval to the exact action, runs each effect at most once across hosts, treats an unknown outcome as ambiguous rather than failed, and leaves a hash-chained receipt. Official LangChain middleware integration; MCP gateway, LangGraph and OpenAI Agents adapters. Python, SQLite/Postgres. v0.12 on PyPI.

Hosted control room for every AI agent a company runs, bought or built (Claude Code, Cursor, ChatGPT, LangChain, anything on MCP): records every action, applies allow, ask or block policy before it runs, and flags drift.

The same boundary for money: refunds, vendor payments and payouts an agent attempts go through, wait for a person to approve that exact amount, or are refused, with per-agent budgets and split-payment detection.

LLM tool that flags contract risks, hidden fees, and one-sided clauses in plain English; PandaDoc’s CEO and CTO proposed integrating it into their platform.

Certifies writing as human. Records how the words arrive in the browser (typing, edits, pauses, pastes) and gives the writer a public certificate link instead of an AI detector’s guess.

Privacy-first wellbeing companion for Mac, iPhone and iPad. Runs on-device AI, including speech-to-text; end-to-end encrypted sync with device-only keys, no accounts. Launching soon.

Semi-supervised model predicting frame 27 of a clip from its first 13, via contrastive learning on unlabeled video.