Atul Agarwal

Control Systems & Software Engineer

Control systems by training, then eight years shipping: three in silicon, five in production software. Most useful where those layers meet, on code that has to be correct and on time.

Get in Touch →

About

I am a generalist engineer. Control theory, chip design and production software, in that order, each learned from zero.

I studied electrical engineering at IIT Jodhpur, then a master's in control & automation at IIT Delhi. Two papers came out of it: sliding-mode control of a converter feeding a destabilising load, validated in hardware-in-the-loop, and system identification, recovering a system's dynamics from input–output data without assuming the model.

None of that prepared me for MediaTek. Digital design is its own discipline and I learned it on the job. Three years of timing closure, RTL verification and physical design, shipping across multiple production SoC tapeouts.

In 2021 I taught myself software engineering from scratch, again with nothing carrying over, and moved into DeFi as a smart contract engineer. I went on to write the core contracts for Hubble Exchange, a perpetuals DEX that cleared $150M+ in volume, and then co-founded Moar Market, a lending protocol on Aptos that peaked at $7.5M TVL, building it in Move along with the backends, liquidation engine and risk monitoring it needed to stay up. The code is public, adversarial and holds user funds directly. A single bug is not a ticket in a backlog, it is money gone in minutes. Four firms have audited my contracts, with no loss-of-funds findings.

Three fields, none of them handed to me by the last one, and I got good at each. What I want now is the opposite problem: robotics is the first thing I have worked on that needs all three at once. Most recently I built a real-time C++ and ROS 2 motion control stack. It runs at 200 Hz on a PREEMPT_RT kernel, does computed torque with online payload estimation, and goes from a MuJoCo model to a real 5-DOF arm. The estimator that works in simulation recovers almost nothing on hardware. The writeup covers the four reasons why and what replaced it.

Control theory, hardware and production software are usually three different people. I am looking for work where they have to be one. Based in India, open to international relocation.

Robotics & Control

Payload-Adaptive Motion Control on the SO-ARM101
Real-time C++ / ROS 2 control stack, sim to hardware
2026

One control loop, written once, driving two plants behind a single PlantInterface, a MuJoCo model and a real 5-DOF arm over a Feetech servo bus. Runs at 200 Hz on a SCHED_FIFO thread under a PREEMPT_RT kernel. That is the runner today, and the tracking figures below are not real-time measurements. Controllers (PD, computed torque, and adaptive computed torque with online payload estimation) never learn which side of the plant boundary they are on.

  • Computed torque vs naive PD (simulation, 200 g known payload): settled arm RMS 0.0216 → 0.0028 rad (−86.9%), end-effector miss 0.0300 → 0.00018 m (−99.4%), gated by a blocking rigid-body equivalence test (worst disagreement 1.33e-15 N·m on gravity, 8.66e-13 N·m on the nonlinear bias) before any controller runs
  • Online payload RLS (simulation): recovers 102.4% of the RMS gap between the empty-model controller and the perfect-model bound; identifies a 200 g payload as 202.3 g
  • The sim-to-real finding: that same estimator returns 0.8 g for a 90 g payload and 9.4 g for a 180 g one on hardware. Four coupled reasons, chief among them that commanded torque is a position overlay rather than plant torque, and that acceleration from a 12-bit encoder carries ~13× more noise than signal inside the regressor
  • The replacement, closed loop: identify mass statically from motor current at rest, calibrate the instrument, freeze, then track, cutting arm RMS a further 31% at 90 g and 45% at 180 g against computed torque with no payload model, with the payload identified as 82.8 g and 176.0 g against true masses of 90 g and 180 g
  • Real-time characterisation: wakeup jitter p99 6.18 µs and p99.9 11.34 µs on a 5000 µs period, against a 2.14 ms serial round-trip. The scheduler is ~189× cheaper than the bus, which is what locates the real bottleneck
  • Instrumented for honesty: current logs report stale reads, truncated commands, bridge saturations and missed deadlines in its log header, so a result cannot silently rest on fabricated samples
C++17 ROS 2 Jazzy PREEMPT_RT Pinocchio MuJoCo Computed Torque RLS Estimation Serial Bus / Feetech
Read the full writeup Source & data

Earlier Experience

Three years designing silicon at MediaTek, then five building production financial systems. Different domains, same demands. Hard constraints, externally audited correctness, and systems that fail expensively when the reasoning is wrong.

Moar Market
Co-Founder & CTO
2024 – 2026

A leveraged lending protocol on Aptos, built from collateral vaults, variable-rate interest models, oracle integrations (Pyth, Switchboard), and a liquidation engine. Peaked at $7.5M TVL with $60M cumulative borrowed.

  • Designed and shipped the full protocol in Move: collateral vaults, variable-rate interest models, oracle integrations (Pyth, Switchboard), and liquidation engine. Peaked at $7.5M TVL with $60M cumulative borrowed
  • Built an internal AI agent with full protocol context to debug failed transactions and explain PnL drops or gains
  • Defined cryptoeconomic risk parameters, collateral factors, and liquidation incentive curves for multi-asset markets. Publicly highlighted by the Aptos Foundation at the $5M TVL milestone
  • Coordinated end-to-end security audit with MoveBit (no loss-of-funds findings); used LLM agents as first-pass PR reviewers and for pre-audit security analysis
  • Built backend distributed systems in TypeScript: indexer pipeline, liquidator bot, risk monitoring, and REST API with retry logic, state management, and failure handling for production reliability
Move / Aptos TypeScript Leveraged Lending Liquidation Engine Security Audits
@MoarMarket
Deploy Trade
Side Project
2026
  • Built an MCP server that exposes DeFi execution strategies as tools callable by LLM agents, enabling agents to plan and run onchain trading workflows end-to-end
  • Underlying execution layer: event-driven grid trading bot on Hyperliquid perpetual futures (TypeScript, RabbitMQ, PostgreSQL, WebSocket-based order lifecycle management)
  • Designed for production reliability: state persistence, retry logic, and order reconciliation across bot crashes and exchange disconnects; observability and tracing instrumented across critical paths
MCP Server TypeScript Hyperliquid RabbitMQ PostgreSQL
GitHub
Hubble Exchange
Protocol Engineer
2021 – 2024

Perpetual futures DEX on Avalanche with $150M+ in cumulative volume. Built a decentralized limit order book (DLOB) on Hubblenet, a custom Avalanche Subnet, in V2, after proving the model with a Curve-style vAMM in V1. Coordinated three external audits (Dedaub, Sherlock, Code4rena) with no critical findings.

V2: Decentralized Limit Order Book (DLOB) on Avalanche Subnet
  • Owned all Solidity smart contracts for the order-book DEX on Hubblenet, a custom Avalanche Subnet (10M+ transactions) where validators doubled as order matchers
  • Wrote custom precompiles in Go for Hubblenet's Subnet EVM (a geth/coreth-based client), pushing order book matching and margin logic directly into the execution layer for performance at scale
  • Built a LayerZero cross-chain asset bridge (lock-mint) that secured multi-million-dollar transfers with no incidents over 2+ years
  • Built liquidator bot, order-matching service, and monitoring infrastructure; coordinated Sherlock and Code4rena audits end-to-end
V1: Virtual AMM Perpetuals
  • Implemented Curve-style vAMM using CurveCrypto invariant, porting complex math (Newton's method iterative solver) to Solidity for on-chain price computation
  • Designed margin engine, funding rate mechanism, and multi-collateral vault system; coordinated Dedaub audit end-to-end
Solidity Avalanche Subnets DLOB LayerZero Bridge CurveCrypto vAMM Security Audits
@HubbleExchange
DefiDollar
Smart Contract Engineer
2021

Developed Solidity smart contracts for DUSD, a stablecoin index on Ethereum (later expanded to BSC), aggregating yield-bearing stablecoins into a diversified, risk-adjusted token. Product suite peaked at ~$45M+ TVL, with audits by PeckShield and Quantstamp. First production Solidity role after transitioning from hardware engineering.

Solidity Ethereum / BSC Yield Aggregation
@defidollar
MediaTek
Senior Hardware Engineer
2017 – 2020

Digital design for production SoCs at one of the world's largest semiconductor companies, across multiple chip tapeouts.

  • Ran the full block-level place-and-route flow end to end, from floorplanning, placement, CTS and routing through timing, power integrity, LEC, LVS and DRC signoff closure
  • Closed a 2.5M-instance multi-voltage hierarchical block on 16nm, owning congestion and timing analysis and routing convergence. Earlier closed a 2.1M-instance block on 28nm
  • Wrote a utility that cut PrimeTime/ICC2 timing miscorrelation by setting endpoint-based clock uncertainty on critical paths, reducing spurious violations chased between the two tools
VLSI / RTL SoC Design Physical Design Timing Closure

Publications

Peer-reviewed control-systems research, first author on both.

  • Nonparametric Analysis of Nonlinear Distortions for Biomolecular Systems
    Atul Agarwal, Abhishek Dey, Rishi Relan, Shaunak Sen. IFAC-PapersOnLine 51(1), 313–318 (2018). IIT Delhi · DTU Compute · doi:10.1016/j.ifacol.2018.05.035

    Frequency-domain system identification. Using broadband random-phase multisine excitation to detect and quantify nonlinear distortion in biomolecular circuits without assuming a model structure, so a modeller can decide whether a linear model is still valid at a given operating point before committing to expensive nonlinear identification. The same question, asked of a robot arm, is what the payload-estimation work below is doing.
  • Sliding Mode Control of Bidirectional DC/DC Converter with Constant Power Load
    Atul Agarwal, Koyinni Deekshitha, Suresh Singh, Deepak Fulwani. 2015 IEEE First International Conference on DC Microgrids (ICDCM), Atlanta, GA, USA, June 2015. IIT Jodhpur · doi:10.1109/ICDCM.2015.7152056

    Nonlinear robust control of a plant that is destabilising by construction: a constant-power load presents negative incremental impedance, so current rises as voltage falls and conventional linear PID control is insufficient. Designed a sliding-mode controller for the bidirectional buck-boost converter interfacing storage in an islanded DC micro-grid, proved existence of the sliding mode and stability of the chosen switching surface, and validated it in real-time hardware-in-the-loop simulation on an OPAL-RT digital simulator.

Skills

Control & Signals

  • Sliding-mode and nonlinear robust control: published, HIL-validated
  • System identification: frequency-domain, multisine excitation, best linear approximation
  • Kalman and particle filtering; recursive least squares for online estimation
  • Computed torque / inverse dynamics; minimum-jerk trajectory generation

Real-Time & Robotics

  • C++17 on PREEMPT_RT: SCHED_FIFO, mlockall, absolute-deadline loops
  • ROS 2 Jazzy; Pinocchio and MuJoCo for dynamics and simulation
  • Serial servo bus bring-up: protocol, timing budgets, encoder calibration
  • From one project, not from production experience: see the writeup

Languages

  • TypeScript, Python: backend services, bots, APIs, analysis
  • Solidity: Foundry, Hardhat, EVM internals
  • Move: Aptos framework, resource model
  • C++, Go, Rust

Digital Hardware

  • Verilog / RTL design for production SoCs: three years, multiple tapeouts
  • Static timing analysis and timing closure
  • RTL verification; physical design flow
  • Reading a datasheet and a register map as a first instinct

Protocol & Systems Design

  • Perpetual futures: funding rates, margin systems, liquidation engines
  • Lending & borrowing: interest rate models, collateral management, risk parameters
  • Order book architecture on a custom L1 / Avalanche Subnet
  • AMM invariants: StableSwap, CurveCrypto, Newton's method solvers in Solidity

Reliability & Security

  • Audit coordination end-to-end: Dedaub, Sherlock, Code4rena, MoveBit
  • Distributed backends: indexers, liquidator bots, risk monitoring, REST services
  • Retry logic, state management and failure handling for production uptime
  • Instrumentation that proves a claim rather than asserting it

AI Tooling

  • MCP servers exposing execution strategies as agent-callable tools
  • LLM agents for first-pass code review and pre-audit analysis
  • Agent-driven debugging with full protocol context

Writing

Technical deep-dives on robotics control, and earlier work on protocol mathematics.

Education

Indian Institute of Technology (IIT) Delhi
M.Tech — Control & Automation
2017
Indian Institute of Technology (IIT) Jodhpur
B.Tech — Electrical Engineering
2015

Get in Touch

Looking for work in controls, robotics and real-time systems, and open to roles that want both the control-theory side and the production-software side. Based in Goa, India. Open to international relocation.

Location:
Goa, India. Open to relocation
Twitter:
@0xshinobii