2026 - NOW
Bespoke Labs
Member of Technical Staff · ML Research Engineering
Synthetic worlds, reward design, and hardened coding-agent environments for post-training.
CURRENTLY BUILDING AT BESPOKE LABS + MIT CSAIL
> whoami
ROMIR
PATEL
ML researcher & engineer.
I build training worlds for agents that need to actually reason.
01
SYSTEM PROFILE
Research, but make
it ship.
I work on the systems around capable AI: agent training environments, rigorous evaluations, synthetic data, and formal verification. I like research that stays interesting after it meets a real codebase.
- 5,000+
- AGENT ENVIRONMENTS
- 100K+
- MODEL INTERACTIONS
- 30%+
- FEWER INVALID PROOFS
02
QUEST LOG
SELECTED
EXPERIENCE
/// COMPLETED + IN PROGRESS
+ 5K+ ENVIRONMENTS
2025 - NOW
MIT CSAIL
Machine Learning Researcher · Post-Training
Multimodal autoformalization and incremental verification for mathematical reasoning systems.
2026
AgentHub
Machine Learning Research Engineer
LLM evaluation and closed-loop synthetic data systems for database reasoning.
2024 - 2025
TextArena.ai
Machine Learning Researcher · Post-Training
Open agent environments for strategic reasoning, negotiation, and self-play.
03
RESEARCH FILES
SAVE SLOTS
FILE_001 · NEURIPS 2025
Autoformalization via cycle consistency + incremental type-checking
Reliable theorem formalization from LaTeX and diagrams with verification in the loop.
STATUS: PUBLISHEDFILE_002 · AUTOMATICA
Physics-informed neural networks with model predictive control
Neural-symbolic methods for solver selection and scientific machine learning.
STATUS: SUBMITTED04
LOADOUT
TOOLS EQUIPPED
AGENTIC AIREINFORCEMENT LEARNINGLLM EVALUATIONFORMAL METHODSMULTIMODAL MLDISTRIBUTED TRAININGPYTHON + TYPESCRIPTSQL + CUDA
EDUCATION Carnegie Mellon University · B.S. Computer Science & AI