I'm an AI research engineer at AutonomyAI, where I build AI coding agents that ship features into large existing codebases. Before that I spent two years at ANZ building LLM and RAG systems for the bank, and worked on deep learning for medical imaging at IIT Bombay.
I write about AI agents: what works and what breaks.
Writing
- Sep 5, 2026A safety model will grade whichever text you show it last
A forged question-and-document pair, pasted into the text a safety guard is checking, makes it grade a harmless decoy while the real request passes through untouched. The same trick clears Shieldstral, gpt-oss-safeguard, and Llama Guard.
- Jun 28, 2026Mistral OCR 4: The Frontier of Document OCR
An independent benchmark of Mistral OCR 4 against frontier vision models and the open parsing stack on olmOCR-Bench. It lands in the top accuracy tier and is the only system here that also returns grounded, confidence-scored output an agent can cite.
- Nov 26, 2025 · leaddev.comWhy AI Economics Are Fundamentally Broken
AI's variable costs never decrease with scale, which breaks the classic software growth playbook.
In the press
Recent places I've been quoted, cited, or interviewed.
Projects
Experience
- AI Research Engineer, AutonomyAIOct 2025 - Present
- Building agent harnesses: orchestration loops, tool design, context management, guardrails
- Multi-agent orchestration: planner and subagent loops, task decomposition, parallel execution
- Context engineering for long-horizon tasks: compaction, memory, retrieval
- Codebase ingestion and retrieval so agents can work in large production codebases
- AI Engineer, ANZJun 2023 - Sep 2025
- Built a dozen production agent systems for engineering workflows across the bank
- Fine-tuned Gemini models on internal documentation for domain-specific answers
- Ran RAG systems in production across internal applications
- Brought LLM API costs down with caching and query optimization
- Won the Global Generative AI Hackathon
- ML Research Assistant, MeDAL Lab, IIT BombayMar 2022 - Jun 2023
- Trained nucleus segmentation models across datasets with incompatible label sets
- Published the work at BIOSTEC 2024
- Built deep learning models for automated histopathology classification
Research
Combining Datasets with Different Label Sets for Improved Nucleus Segmentation and Classification, published at BIOSTEC 2024 in Rome, with collaborators from IIT Bombay and Tata Memorial Centre.
Public histopathology datasets label cell nuclei with incompatible class sets, so models are usually trained on one dataset at a time. The paper proposes a class-hierarchy method for training a single model across them, improving both segmentation and classification.
Subscribe
New essays on AI engineering, by email. No spam. My writing has reached 400k+ readers so far.