AI Coaching Assistant

Lead Engineer, Velotio (R Systems) · 2023 – 2025

A coaching assistant embedded directly in Slack, taken from design through to production. Rather than a separate web app people have to remember to open, it answers in the channel where the conversation is already happening, grounded in the organisation's own material through a retrieval pipeline built on LangChain and Vertex AI.

Most of the engineering went into the parts around retrieval: assembling context under a fixed prompt budget, reconstructing Slack threads into coherent conversations, and making the assistant decline rather than improvise when nothing relevant came back. Evaluation ran off real production questions kept as regression fixtures, not a synthetic set.

• Reach: Answers arrive in the channel the conversation is already in, with no separate app to open
• Cost: Serverless on GCP Cloud Functions, so cost tracked usage rather than headroom
• Trust: Grounded answers cite their sources, making them checkable by the reader

• LangChain
• Vertex AI
• GCP Cloud Functions
• Slack API

Engineering Manager for Agents

An open-source, MIT-licensed harness that turns a single terminal coding agent into an engineering manager for a fleet of autonomous IC agents. It plans and decomposes work, delegates each task to a worker with an isolated checkout, supervises the parallel runs in tmux, and ships through git and the GitHub CLI.

The load-bearing piece is the per-task budget. An autonomous agent has no sense of sunk cost and will circle an unsolvable problem indefinitely while producing plausible output the whole time. A hard token and wall-clock ceiling turns that silent failure into a reported one, which is what makes a fleet observable instead of merely parallel.

• Shape: One planner, many workers: fan-out with a single point of coordination
• Safety: Per-task token and time ceilings guarantee every delegation terminates
• Isolation: One task, one workspace, one branch, so collisions surface as merge conflicts

• AI Agents
• Multi-Agent Orchestration
• Claude Code
• tmux

OffGrid Vision: Local Multimodal CLI

An open-source, MIT-licensed CLI that analyzes images with a local multimodal model through Ollama instead of spending cloud multimodal tokens. The images never leave the machine, which makes it usable on material that could not be sent to a hosted API in the first place.

Built zero-dependency and scriptable so it drops into an agent harness as a local vision step: pipe it a path, get structured output back, no service to stand up and no key to manage.

• Privacy: Runs fully offline, so image data never leaves the machine
• Footprint: Zero runtime dependencies, installable and scriptable in one step

• TypeScript
• Node.js
• Ollama
• Multimodal LLMs

Production Scheduling Backend (Netflix)

Lead Engineer, Velotio (R Systems) · 2025 – Present

Backend work on Scope, the tool Netflix content teams schedule productions in. The job is mostly the unglamorous half of a live internal product: shipping features like recurring bookings and Airtable-driven scripting into something people are using that same day, without destabilising it.

I also own Scope's data pipelines, which means their reliability and monitoring rather than just their code. Separately, on the Creative Equity project, I built a translation service that processes multilingual input captured through no-code Salesforce forms and runs it through AWS Translate.

• Ownership: Scope's data pipelines end to end, including their monitoring and the freshness people depend on
• Delivery: Features shipped into a live scheduling tool without interrupting the teams working in it
• Process: Documented previously undocumented workflows and introduced Jira tracking, so handovers stopped depending on who remembered what

• TypeScript
• Node.js
• AWS
• Salesforce
• Airtable

Cloud Apps & APIs (Pure Storage GTM)

Lead Engineer, Velotio (R Systems) · 2021 – 2023

Cloud-native automation for a go-to-market platform, used day to day by more than five teams. The work spanned event-driven services on AWS, a workflow scheduler, and the data path feeding analytics through Kafka into Snowflake.

I led a 12-engineer team across the effort, which meant as much time on interfaces, review standards, and splitting work so people were not blocked on each other as on the systems themselves.

• Architecture: Event-driven services on AWS, with the analytics path running Kafka into Snowflake
• Scale: Built a scheduler running 1,000+ workflows daily
• Team: Led a 12-engineer team; platform adopted by 5+ teams

• AWS Step Functions
• Lambda
• EKS
• Kafka
• Snowflake

Digital Evidence Management (iCrimeFighter)

Lead Engineer, Velotio (R Systems)

A digital evidence management system for law enforcement agencies: collecting, storing, and organising the material a case is built from. I was the primary backend developer on it and contributed to the front end where it was the thing blocking a feature.

The constraints come from where the evidence is captured. Files are large, they are recorded in the field on unreliable connections, and the same one often arrives twice from two officers. So: resumable uploads that survive a dropped connection, duplicate detection so a case does not store the same footage repeatedly, case workflow tasks, and API integrations with partner evidence platforms so an agency is not locked to one vendor.

• Reliability: Resumable uploads, so evidence captured in the field survives a dropped connection mid-transfer
• Storage: Duplicate detection, so the same file is not stored repeatedly across a case
• Interoperability: Partner API integrations, so agencies can exchange evidence rather than commit to one vendor

• TypeScript
• Node.js
• Docker
• Serverless Framework
• AWS ECS
• Angular

Photo Distribution Platform (Boltpics)

Co-Founder & Lead Software Engineer · 2019 – 2021

A smart photo-distribution startup I co-founded. Event photographers upload a shoot, and a facial-recognition pipeline delivers each attendee only the photos they appear in, which removes the step where thousands of images get dumped into a shared folder nobody searches.

As co-founder I owned the technology end to end: the recognition pipeline and its throughput, the infrastructure bill, and the tradeoffs between the two. Running a compute-heavy pipeline on a startup budget set most of the architecture.

• Scale: Facial-recognition pipeline processing 1M+ images/day
• Business: Grew to ₹1 Cr (~$120K) revenue in two years

• Node.js
• AngularJS
• Firebase
• TensorFlow
• AWS

Yoga Pose Feedback

A real-time yoga coaching app that scores how closely you match an instructor's pose using on-device pose estimation with TensorFlow.js MoveNet. It runs entirely in the browser: no backend, no uploads, and the camera feed never leaves the device.

The interesting problems were a confidence-weighted matching algorithm that ignores joints the model is unsure about, an offline video pipeline to generate reference pose data from instructor footage, and a compact binary format for that data so the app loads quickly on a phone.

• Privacy: Fully client-side, so the camera feed never leaves the device
• Accuracy: Confidence-weighted scoring that discounts low-certainty joints
• Performance: Custom binary pose format to keep first load fast on mobile

• React
• Vite
• TensorFlow.js