Hi, I'm Akash.I build agents, andthe loops that ship them.
Two-time founder (one exit), ex Senior ML Engineer at mirrAR and Founding Engineer at a YC company. Since October 2025 I run Motiontic, a solo applied AI lab: I work inside startup teams as an AI lead, founding engineer or product engineer, and run my own experiments on the side. I work in Claude Code and Codex, with 5 to 6 agents running in parallel loops on a normal day.
1000+PRs shipped through gstack-loops, up to 40 in a day
5+startups shipping agents built by Motiontic
8years building: two startups, one acquired in 2023
300+members in GTM Machines, the community I run
01 · Work
Where I've worked, newest first.
Oct 2025 - now
Motiontic
Founder · solo applied AI lab
I embed with startups in India and the US and ship agents and AI products into production, in the role each team needs.
AI lead: owns the AI work inside the client's teamFounding engineer: first engineer on a new productProduct engineer: owns a product end to end
Quantum Heaps
AI engineer · AI lead
Built an agentic revenue harness for AgentQ, their agentic CRM, that runs pipeline, inbound and outbound for salespeople.
PulseCorp
Founding engineer
Pulse, an AI user research platform. Launched a real-time UX interview agent that reads speech, video reactions and clicks and asks follow-up questions, plus theme and sentiment extraction and Slack, HubSpot and analytics integrations. Cleared security reviews for Remitly in 7 days and Groww in 5.
NeevCloud
Product engineer · Fring
Fring, a coding harness across desktop, CLI, VS Code and web. Built a 50-person superuser cohort that drives activation and turns power users into a feedback and distribution loop.
Bloom Markets
AI content
An agentic video and influencer harness that any marketer with basic Claude Code or Codex skills can run: 20+ accounts posting around the clock, 10K+ daily views from accounts under 100 followers.
mirrAR
AI lead
Led AI as a client of Motiontic, separate from my years on staff. Manufacturing agents for a 150-person team: an email order agent handling 500 orders a day, a QC agent checking 200+ deliveries a day against purchase orders, and a spec-sheet-to-SKU pipeline.
Jun 2025 - Sep 2025
Narrative (YC W23)
Founding engineer
A testing platform where AI browser agents run end-to-end tests. I built the browser infrastructure underneath it.
Remote-browser infrastructure on AWS: a pool of streamed browser sessions (NICE DCV) with near real-time interaction, so users could watch and debug agent tests live
Hundreds of parallel test runs per organisation: atomic session pool, autoscaling after inactivity, isolated and incognito sessions, race-free session assignment and cleanup
A recorder for no-code test creation, browser automation to run tests, and replay to debug them
Monitoring with Prometheus, CloudWatch and Sentry; GitHub and Slack integrations
Nov 2023 - May 2025
mirrAR
Machine Learning Engineer → Senior ML Engineer · employee
Virtual try-on and generative AI for beauty, jewellery and fashion brands.
Started with StyleGAN hairstyle transfer for clients like Lakmé, and an iOS ARKit SDK for in-store partners
Moved into generative AI: jewellery try-on, cloth try-on (10K+ try-ons a month on a 500K-visitor site), and AI jewellery design software (Spark, now used by 5+ enterprise customers)
Led the Rad Beauty product for six months with a team of 7 to 8: AI try-on cut client photoshoot costs by 60% and grew the app 4x in downloads and 160% in DAU in three months
Jul 2020 - Oct 2023
Durbin Technologies
Co-founder & CTO · acquired
Raised $125K and scaled an AI-enabled IoT asset-tracking platform from zero to acquisition: 500+ locations across India and a team of 10. Turned 2M+ daily telematics events into decisions (15% faster turnaround) and delivered LiDAR railway mapping for Indian Railways.
May 2018 - Jun 2020
Quinch Systems
Co-founder
Computer vision for industrial sites, started in college, for Skipper, Linde and Indian Railways. It led directly to Durbin.
02 · Motiontic lab
Motiontic is a solo applied AI lab.
Since October 2025 the lab has two halves. Client work: I join a startup's team as an AI lead, founding engineer or product engineer and ship agents into production (above, under Work). Experiments: everything below, most active first. Some became products, some stayed research.
01 · researchedition 1 out
ARI-42, Agent Risk Index
An open index of the 42 security requirements a product built on AI agents should meet, with evidence-linked ratings of how far 85 widely used agent frameworks, SDKs, harnesses and MCP tools take a builder toward each one. Every requirement is tied to the 2025-26 CVE and incident record that motivates it, and cross-walked to OWASP, MITRE ATLAS, CoSAI and ISO.
85agent tools rated
2,099ratings, each with a source
129,058CVEs screened
759AI incidents reviewed
Finding: tools meet a requirement by default in 2% of cases.
02building
GLoop
One chat for every coding agent and loop on your machine. You talk to one router agent; it writes the loop prompt for the right project and only interrupts you when a decision is yours. gstack-loops v2.
03 · most usedused by developers
gstack-loops
A loop engineering kit built on gstack. Run as many feature loops in parallel as you want: each gets its own worktree, its own set of ports so parallel builds never collide, and its own QA run. A loop goes spec → build → review → QA → ship and keeps going until the tests pass, so it finishes the whole development cycle and hands you a PR.
The shared context layer for humans and agents: one live page that every teammate's coding agent pulls, follows and publishes back to. This portfolio is a board on it.
A terminal coding agent written from scratch in Rust, aiming to match the leading harnesses while starting faster, using less memory and spending fewer tokens.
07building
Witness
Deploys a feature's whole stack, across every repo and PR it needs, into one cloud sandbox, then hands the URL to Claude Code once every service is healthy.
08500+ users
Sudosu
An AI workspace with 100+ integrations and a Python CLI. Connect Gmail, Slack and Linear and build agents for long-running tasks.
An AI content creator framework. Generates UGC video at scale with Higgsfield or fal.ai, manages many devices and the Instagram accounts on them, and automates posting end to end.
Computer use with TypeSafe's Jev decision model: Jev makes the typed judgements (yes/no, choices, scores with calibrated probabilities) while my code drives the browser. Use cases for lead generation, a browser agent, a game agent and ad-library research.
11npm: legendai
Legend
A coding agent that runs Linear from Slack through a six-stage pipeline. It let a CSM who doesn't code ship 70+ production PRs in a month.
12spec
BranchDB
A database with branching built in, and native vector search as the second selling point.
13open source
Launch Film Kit
Launch films that are clear on first watch. Claude writes and directs, ElevenLabs voices and scores, HyperFrames renders.
14done
Greenlight
An agent-run security and compliance gate, and the research behind it into the gaps in today's agentic AI vendor reviews. Led to ARI-42.
15building
Panelist app
An app for people who get paid to take part in research studies: verified, one account per person.
16building
SkillHunt
Product Hunt for AI skills: discover, upvote and install skills for Claude Code and Codex.
17done
Glassbox
Live hiring screens: watch candidates work with coding agents inside a deliberately broken codebase, then get an evidence-linked report.
18prototype
Proactive Agents
An agent engine on LangGraph that acts before you ask: it watches Linear and your tools and relays work to a native app.
19prototype
AgenticMail
An AI inbox with real-time sync, folder filtering, attachments and suggested replies.
20prototype
Hiroscope
AI recruiting that scores candidates against a job, one at a time or in bulk, with the reasons shown.
03 · Writing
Notes on building with agents. Shorter takes on X and LinkedIn.
From a first-year student with zero coding experience to running a one-person lab.
Nov 2026Intraface goes live with two startupsAnd this portfolio becomes a board on it.
Oct 2026Launched a YouTube channelMy own content harness makes the videos for The Global Political and The Global AI News. One crossed 28K views.
Sep 2026Drevon launches publiclyAfter a month of beta with 100+ users and 3 paying customers.
Aug 2026The Growth Hackathon by GTM Machines200+ registrations, about 50 builders and 30+ teams shipping GTM systems in two hours at Persistence, Bengaluru.
Jan 2026Hosted Lenny's Newsletter's first Kolkata meetupAnd started writing at akashmunshi.com.
Nov 2025Launched SudosuBuilt in public: 200+ users by December, 500+ by the time v3 shipped.
Oct 2025Started Motiontic, a solo applied AI labAgents and AI products for startups in India and the US, and my own experiments on the side.
Jun 2025Founding engineer at Narrative (YC W23)Built the browser infrastructure for AI browser-agent testing: hundreds of parallel streamed browsers per organisation, a test recorder, automation and replay.
2024Senior ML Engineer at mirrARLed a team of 7 to 8 for six months on the Rad Beauty product: AI try-on, StyleGAN hairstyles, ARKit.
Nov 2023Machine Learning Engineer at mirrARStarted with StyleGAN hairstyle transfer for clients like Lakmé and an ARKit SDK, then moved into generative AI: jewellery try-on, cloth try-on and AI jewellery design software.
2023Durbin is acquiredMoved from Kolkata to Bengaluru.
Feb 2023Started consulting with Altor Smart MobilityProduct for an IoT rider-safety platform used by 150K Zomato delivery partners: 1.5M+ annotated data points a day, ETL and MLOps for on-device accident detection.
2021Graduated: B.Tech in Computer ScienceAcademy of Technology, 9.26 DGPA. TensorFlow Developer Certification. Taught Python and ML to 500+ students along the way.
Jul 2020Co-founded Durbin Technologies as CTORaised $125K and scaled AI-enabled IoT asset tracking to 500+ locations and a team of 10. LiDAR railway mapping for Indian Railways.
May 2018Co-founded Quinch Systems in collegeComputer vision for Skipper, Linde and Indian Railways, in dust, bad light and old equipment.
2017First year of college, zero coding experienceAndroid docs, then sklearn, OpenCV, TensorFlow and PyTorch. A statistics nerd who found machine learning and was sold.
05 · About
I don't come with big company or college tags. I learned by reading every page of documentation I could find, and later by having long conversations with ChatGPT. Distribution is the problem I've struggled with my whole career, so it is the one I most want to solve with agents.
How I work: Claude Code and Codex, 5 to 10 sessions in parallel, each in its own worktree, each looping until its tests pass. It changed what I can ship in a week more than anything else.
Outside the terminal I run GTM Machines, a community of 300+ GTM practitioners and founders in Bengaluru (50+ sessions so far), and The Baseliners Club, a lawn-tennis group with 100+ members.
Tools I live in
Claude Code and Codex Python, TypeScript, Node, Rust LangGraph, RAG, MCP AWS, Cloudflare, Vercel, Postgres Computer use on cloud phones and remote browsers, with TypeSafe's Jev decision model