Agentic AI · Code Generation · Tool Use
Param Popat builds reliable agents that turn intent into executable code.
Tech lead and senior ML engineer working on LLM post-training, compiler-in-the-loop evaluation, structured automation, simulation, and on-device intelligence.
About
Agentic AI, LLM post-training, tool use, code generation, simulation, and production ML systems.
I'm a tech lead and senior machine learning engineer focused on agentic AI systems that turn natural language intent into validated, executable code over typed automation interfaces. My recent work spans LLM post-training, tool-use evaluation, verifier-guided feedback, compiler-in-the-loop repair, and reliability guardrails for multi-turn agents.
I have led production-facing ML efforts across code generation, structured reasoning traces, distributed evaluation, and on-device constraints. The common thread is making model behavior measurable and executable: parsing generated programs, validating them against schemas and compilers, repairing failures, and building evaluation loops that expose real reliability gaps before deployment.
- Architected an on-device agentic harness that coordinates LLM reasoning across 600+ actions and 100+ triggers, improving executable code rates from 78% to 99% through AST parsing, structured traces, compiler feedback, and automated repair.
- Directed post-training and evaluation strategy using 80k mined workflows plus 200k synthetic and human-annotated tool-use trajectories for SFT, RLHF/DPO-style preference optimization, and verifier-backed evaluation.
- Built reliability guardrails for production agents, including constrained decoding, entity resolution, dynamic user disambiguation, and strict validation of non-deterministic model outputs.
- Scaled distributed training and evaluation with Ray-based infrastructure, debugging online/offline metric gaps and increasing LLM judge pass rates from 60% to 88%.
- Previously built photorealistic 3D simulation pipelines, 3D Gaussian Splatting reconstruction for 1000+ indoor environments, closed-loop RL evaluation at 50k+ agent-hours, and on-device sensor models for real-time gesture inference.
Experience
Production ML leadership across agentic code generation, post-training, evaluation, simulation, and AI security.
Tech Lead - Senior Machine Learning Engineer, Agentic AI
Apple Inc., Cupertino, CA
- Architected and led an on-device agentic harness for natural-language-to-code generation over a typed automation DSL, coordinating LLM reasoning across 600+ actions and 100+ triggers.
- Designed a multi-turn execution framework with AST parsing, structured intermediate traces, compiler-in-the-loop feedback, and automated repair, raising executable code rates from 78% to 99%.
- Directed post-training data and evaluation strategy using 80k mined workflows and 200k synthetic plus human-annotated tool-use trajectories for preference optimization and verifier-guided feedback.
- Engineered token-budgeted training rollouts to keep multi-turn reasoning within a 64k context window while preserving production latency constraints.
- Led Ray-based distributed evaluation and training workflows; debugged online/offline evaluation disparities and improved LLM judge pass rates from 60% to 88%.
- Built reliability guardrails with constrained decoding, entity resolution, dynamic disambiguation, and strict validation for structured automation outputs.
Machine Learning Engineer, Simulation & On-Device ML
Apple Inc., Cupertino, CA
- Architected a scalable 3D simulation pipeline moving from NeRFs to 3D Gaussian Splatting for real-time spatial rendering on constrained devices.
- Built distributed GPU training infrastructure to reconstruct 1000+ high-fidelity indoor environments with surface smoothing and dynamic level-of-detail pipelines.
- Scaled closed-loop reinforcement learning evaluation to 50k+ agent-hours for robotic perception and long-horizon planning stress tests.
- Co-developed a transformer-based on-device sensor foundation model using IMU and PPG signals for real-time, calibration-free gesture inference.
AI Research Intern
AI Zwei (Dailight), New York, NY
- Engineered an automated model-compression pipeline for production AutoML, combining knowledge distillation, weight palettization, and post-training quantization.
- Delivered a one-click optimization workflow that doubled inference QPS for enterprise deployments while improving latency stability and robustness.
Project Trainee – AI
Robert Bosch GmbH, Bengaluru, IN
- Led AI safety and red-team research on black-box model extraction via teacher-student distillation with randomized query payloads.
- Invented query-distribution defenses using anomaly detection and lightweight feed-forward models to flag out-of-bound requests.
- Patented model-extraction defense mechanisms later commercialized into Bosch's AI security product line.
Education
Academic foundation in computer science with focus on machine learning and artificial intelligence.
M.S. in Computer Science
Columbia University, New York, NY
GPA 3.98/4.0, TA for courses in Computer Vision and NLP.
B.Tech in Computer Engineering
Nirma University, Ahmedabad, IN
GPA 9.12/10.0, TA for Deep Learning.
Selected Work
High-signal technical areas for recruiters and research teams.
Agentic Code Generation
Natural-language-to-code systems over typed automation interfaces, with multi-turn reasoning, tool retrieval, entity grounding, compiler feedback, and repair loops.
LLM Post-Training & Evaluation
SFT, RLHF/DPO-style preference optimization, verifier design, LLM judge calibration, distributed eval harnesses, and metric debugging for agentic tool-use reliability.
Simulation & Embodied AI
Photorealistic 3D environment reconstruction, NeRF and Gaussian Splatting pipelines, closed-loop reinforcement learning evaluation, and long-horizon planner stress tests.
On-Device ML & Optimization
Model compression, quantization, distillation, token-budgeted rollouts, and real-time inference systems for constrained production environments.
Articles
Thoughts on AI safety, computer vision, and the intersection of technology and society.
Protecting Society from AI‑Generated Misinformation
This Analytics Magazine piece discusses how synthetic content generated by AI can mislead the public and proposes ethical guidelines for detection tools, regulatory oversight and responsible AI deployment.
Read article →Siamese Networks for Visual Tracking
An overview of Siamese neural networks and contrastive loss for real‑time object tracking, demonstrating a fully convolutional Siamese model with a ResNet backbone that achieves 0.88 precision on the OTB dataset.
Read article →The Journey of Mathematics (Layover at Pi)
A personal reflection on why many people dislike formal mathematics and how exploring timeless concepts like π can reveal the discipline's deeper beauty and interconnectedness.
Read article →Publications & Patents
Research contributions in machine learning, computer vision, and AI security.
- Granted European Patent: AI Model-Extraction Defense (EP4007978B1) - query-based defenses using synthetic abstention classes and gain-adaptive lockouts, assigned to Robert Bosch GmbH.
- Animal/Object Identification Using Deep Learning on Raspberry Pi - published in Springer's Smart Innovation, Systems and Technologies series.
Get in Touch
For recruiting, research, or collaboration conversations, email is the best way to reach me.