Applied AI Engineer. I build production AI systems — LLM pipelines, agents, retrieval, evaluation — and the infrastructure that keeps them trustworthy once real users depend on them.
Two things I've built are public and used by other people: CTRLRun, an execution-safety layer for AI-agent actions, and EmoRoBERTa, an emotion-detection model with 15M+ downloads on Hugging Face.
github.com/CTRLRun/ctrlrun · ctrlrun.dev · Apache-2.0 · creator and maintainer
An agent that can send money, delete infrastructure, or email a customer needs more than a good prompt. CTRLRun is the layer that makes those actions safe to execute:
- Approval binding — a human approval is bound to the exact action it was granted for. Change the amount, the recipient or the tool, and the approval no longer matches.
- At-most-once execution — each logical effect executes once, across processes and across hosts, on SQLite or Postgres.
- Ambiguity is not failure — an unknown outcome is recorded as
AMBIGUOUS, never silently as failed. The library never maps an unrecognised exception to "it didn't happen", because that is exactly how an agent ends up paying an invoice twice. - Hash-chained receipts — portable JSON, tamper-evident, showing which policy decided each action.
- Guarantees as tests —
ctrlrun verifyruns the guarantee catalogue against your policy in your CI, so the claims in the README are checked rather than asserted.
Ships an MCP gateway, LangGraph and OpenAI Agents SDK adapters, and OpenTelemetry export. v0.6.1 on PyPI, 4,400+ tests.
huggingface.co/arpanghoshal/EmoRoBERTa
Context-based emotion detection over 28 emotion classes. Top trending emotion model on Hugging Face 2020–2022; 15M+ downloads, 100+ citations.
| Project | What it is |
|---|---|
| ContractReviewer | 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. |
| QualyDM | Conversation-first outbound platform where an LLM runs the DM exchange end to end. 10k+ DMs/month, 40% reply rate, 20% booked calls. |
| NoteToDoctor | Voice-first health tracker. An LLM pipeline turns a 30-second spoken check-in into 12 structured health streams. |
| GenAI Review-Insights Platform | Mines product reviews across the web into survey-style insights. Adopted by multiple companies. |
| Clinical Notes Embeddings | Fine-tuned BioBERT for domain-adapted clinical-note embeddings, improving retrieval over EHR data. |
| Video Frame Prediction | NYU Deep Learning (Yann LeCun). Semi-supervised model predicting frame 27 of a clip from its first 13, via contrastive learning on unlabelled video. |
Europe & United States, remote · Mar 2026 – present
- Design and deliver production GenAI systems — RAG pipelines, AI agents and agent harnesses, workflow automation, LLM backend services — for early-stage startups across Europe and the US, working directly with founders to scope requirements and owning architecture, implementation, deployment and handover.
- Retained by Operating Equity Partners after transitioning out of the Director role: I continue to own, maintain and extend the production AI systems built for their portfolio ventures — LLM pipelines, the agent reasoning layer, and the evaluation/observability stack.
- Translate ambiguous business requirements into technical roadmaps, and communicate the trade-offs of LLM system design to both technical and non-technical stakeholders.
San Carlos, CA, remote · Feb 2025 – Feb 2026 (retained since Mar 2026)
- Owned product and technical strategy across multiple early-stage AI ventures, partnering directly with Halsey Minor — co-founder of CNET, Salesforce, Google Voice and NBCi — to incubate and launch new companies.
- Led two 0-to-1 AI product launches: a trusted news network aimed at reducing misinformation, and an AI avatar legal assistant for consumer-facing legal support.
- Designed and implemented the end-to-end AI architecture for both: LLM pipelines, retrieval systems, backend services, data infrastructure and the integrations that made the core product experiences possible.
- Built the agent reasoning layer for the legal and news-verification workflows — multi-step investigation, evidence retrieval, tool-augmented reasoning, typed structured outputs, and explicit exception handling for ambiguous queries.
- Built the LLM evaluation and observability layer: automated eval pipelines, output guardrails, and tracing (LangSmith / Langfuse) to monitor quality, latency and cost in production.
- Fine-tuned and adapted open-source language models for specialised use cases, turning emerging capabilities into shipped product features. Python, Flask, Pydantic, vector DBs, LangChain, AI avatar systems.
Philadelphia, PA · Sep 2023 – Feb 2025
- Built the company's first commercial machine learning model for power-outage detection, reaching 96% accuracy and owning the pipeline end to end — geospatial data acquisition, feature engineering, training, deployment for utility clients.
- Designed a custom outage-detection architecture combining mathematical modelling, deep learning and ensemble methods to capture relationships across geospatial, weather, grid, infrastructure and device data.
- Built forecasting models for outage patterns and demand trends, improving proactive resource allocation for utility clients.
- Designed and shipped a GenAI-powered detection system that analysed social media signals in real time to identify outage events ahead of conventional monitoring.
- Deployed across US and international markets with monitoring dashboards, automated retraining workflows, and the foundations for a fibre outage detection model.
Seattle, WA, remote · Sep 2021 – Jun 2022
- Architected an AWS machine learning pipeline on 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.
Bangalore, India · Jan 2021 – Aug 2021
- Built and deployed a real-time AI 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.
London, UK, remote · Jun 2020 – Aug 2020
- Built the company's first data infrastructure and data APIs from scratch; later adopted by major fintech clients on the platform.
New York University — MS in Computing, Entrepreneurship and Innovation, Courant Institute of Mathematical Sciences & Stern School of Business (2022–2023). Completed the two-year programme in one year; research assistant in deep learning and applied AI; worked with founders building startups.
PES University, Bangalore — BTech in Computer Science and Engineering, Data Science specialisation (2017–2021).
AI/ML — LLMs · RAG · AI agents · MCP · human-in-the-loop approval workflows · LLM evaluation & guardrails · fine-tuning (LoRA/QLoRA) · structured outputs (Pydantic) · OpenAI & Anthropic APIs · OpenAI Agents SDK · LangChain / LangGraph · Hugging Face Transformers · PyTorch · TensorFlow · scikit-learn · XGBoost · NLP · audio ML
Data — PostgreSQL · SQLite · Redis · MongoDB · DynamoDB · Elasticsearch · vector DBs (Pinecone, ChromaDB) · pandas · NumPy · ETL pipelines · geospatial data
Backend & cloud — Python · FastAPI · Flask · REST · WebSockets · WebRTC · Docker · Kubernetes · Terraform · AWS (SageMaker, Lambda, S3, EC2, ECS/EKS, RDS, Step Functions) · GitHub Actions
MLOps & observability — MLflow · Weights & Biases · LangSmith · Langfuse · OpenTelemetry · model monitoring · automated retraining · pytest
Certifications — AWS Certified Solutions Architect · NASA Space Apps Challenge mentor and judge
- An Enhanced Context-Based Emotion Detection Model Using RoBERTa — IEEE CONECCT 2022, Bangalore, pp. 1–6.
- CATAQ: Concise Answer to Any Question — ICSES 2021, pp. 1–7.
Relocating to the Netherlands on an Orientation Year (Zoekjaar) permit — authorized to work in the NL with no employer sponsorship or TWV required. Open to Applied AI / ML engineering roles.