Ashwin M.S / Career & selected work
I build products, the systems behind them, and an understanding of how they work.
I’ve worked across consumer marketplaces, media infrastructure, manufacturing and industrial AI, often in small teams where engineering and product decisions happen together. My work spans backend architecture and production operations, knowledge and retrieval systems, and software that connects to physical machinery.
Alongside that, I explore machine learning through implementation: training transformers, investigating their behaviour, and working with robots. I like being close to the details while keeping sight of what the whole system needs to accomplish.
Career
- 17 April 2026–present
Allude
Every factory runs on expertise locked in experts’ heads. Allude captures and structures that knowledge, so performance no longer depends on who’s on shift.
I work with the founders on product direction and build the platform’s core data and intelligence backend. My work has helped move Allude from an MVP towards a more stable, modular platform used across multiple enterprise customers.
Contributions & outcomes
- Rebuilt the backend foundations. Built v2 across the database schema and backend implementation, reworking MVP features for stability and modularity so subsequent product work has a stronger base.
- Made SOPs versioned and editable. Introduced version history, rollback, draft editing and checkpointing in place of an unversioned workflow.
- Connected scattered knowledge. Reorganised the datastore and unified ingestion around modular, swappable components. Revamped indexing, automated tagging and labelling, and knowledge graph generation to support continual integration of new knowledge.
- Expanded what the intelligence layer could use. Built agent tools for accessing the underlying data, benchmarked and improved retrieval, and improved SOP generation by making use of previously untapped multimodal sources.
My ownership spans ingestion, knowledge representation, retrieval, access control and infrastructure.
- January–April 2026
Independent ML research & engineering
I dedicated this period to TEL, PyTorch and independent ML explorations: implementing models, running experiments, reading the relevant literature, and documenting what I learned. The aim was to build intuition through the full process of training and debugging, from data and optimisation to memory use and compiler behaviour.
That work lives in Transformer Efficiency Lab and my ML notes.
- August–December 2025
xLogic Labs
I worked across the software and hardware sides of a manufacturing platform, from configuring a fabrication order in the browser to integrating a robotic welding cell.
Contributions & outcomes
- Ordering and instant quoting. Built the web product for ordering, configuring and customising fabrication requirements, including the RFQ system for instant quotes. Collaborated on DFM checks and browser-based CAD workflows for manufacturability detection.
- Industrial integration. Set up the local factory network and programmatic control of welding units. Integrated their controllers with a custom welder system, a Fairino FR5 robotic arm and an additional horizontal axis that moved the arm over the welding bed.
- Launch. Contributed to the product that attracted 100+ users in its first month and processed 200+ instant RFQs during my time there. The welding cell was a prototype; early customer orders were fulfilled through vendors.
This was hands-on work across web software, networking, controllers and motion, with the practical constraints of a factory shaping the implementation.
- December 2022–August 2025
Headout
I built and maintained supply and booking systems in Kotlin and Spring, connecting suppliers to customer bookings in a consumer marketplace. My work included live entertainment, theatre and seat-map booking workflows, supplier integrations, and commission handling in pricing.
Integrations I wrote handled hundreds of thousands of bookings per month. Across the wider team, the integration footprint grew from roughly 15–20 to 45–50 integrations, with automated bookings reaching nearly 10,000 per day at peak.
I also built internal tools for listing management and inventory tracking, delivered integrations across operations, business and engineering teams, and contributed to analytics-driven operational workflows. A regular on-call rotation kept me close to how these systems behaved in production.
- August 2021–August 2022
Incredible.dev
I was a core member of the team that built and open-sourced Incredible. I built backend services in Go and TypeScript, including media generation, processing and streaming pipelines using FFmpeg and AWS MediaConvert.
My work also covered cloud infrastructure, containerised services deployed through Docker and Kubernetes on AWS, and product analytics integrations with Segment, Amplitude and Customer.io. It was an early opportunity to own work from the backend implementation through deployment and product behaviour.
Selected work
Transformer Efficiency Lab
TEL is my ongoing exploration of the systems inside a modern transformer. I implement and train decoder-only models, investigate architectural and optimisation choices, and measure the effects on training behaviour, throughput and memory.
The work spans tokenisation and data pipelines, gradient accumulation, activation checkpointing, diagnostic metrics and torch.compile. One investigation traced large training-time spikes to changing buffers and mutable state in diagnostic hooks; the fixes eliminated guard-failure recompiles in the tested runs.
I keep the experiments, failures and conclusions public so there is a trail from the question to the implementation and the result.
Explore TEL → · Compile investigation → · Code ↗
SO-101 / LeRobot
I assembled and calibrated an SO-101 setup, worked with teleoperation and two-arm control, collected demonstrations, and fine-tuned a policy for pick-and-place imitation learning. The project brought together data collection, training and physical robot rollouts.
Working with a real robot made the connection between perception and action tangible: the demonstration data, calibration and physical setup were all part of the learning system.
Project, talk & demos → · Code ↗
Background
B.E. Computer Science & Engineering · VTU · 2017–2021
My main tools have included Python, Go, Kotlin/Java and TypeScript; PyTorch for ML; AWS and Azure for infrastructure; and industrial controllers and robot interfaces for physical systems. The stack changes with the problem.
For the personal side, there’s Me. For more experiments and smaller projects, there’s Builds.
Have a product, systems or ML problem you'd like to work on together?
[email protected] ↗ GitHub ↗ LinkedIn ↗