AI/ML Engineer in London, currently at Acutro. I have 6+ years shipping production systems and now work end-to-end on GenAI — agentic architectures, RAG at scale, and LLM-powered products — with the MLOps to keep them alive after the demo.
- 🔭 Building agentic systems with LangGraph, Claude, and MCP — most recently ortelius/pdvd-aiops; plus time-series and predictive-maintenance ML on high-dimensional IoT sensor data.
- 🎓 MSc Big Data Science, Queen Mary University of London — Distinction.
- 🏆 Top CDF Contributor 2024 (The Linux Foundation), Ortelius Governing Board member, and the first-ever Gold Legend — see Open Source.
- ✍️ I write about the things I build — see Writing below.
- 📄 Résumé (PDF) — the full detail behind everything below.
- 📫 Reach me on
LinkedIn · Twitter · utkarshkviim@gmail.com
⭐ Featured Work — ortelius/pdvd-aiops
Ortelius AIOps — an autonomous, agentic pipeline for dependency health and supply-chain security, built for the CDF's Ortelius project. Point it at a GitHub repo and it updates every outdated dependency, builds and tests the result, bisects and rolls back whatever breaks, patches CVEs, and opens a PR with an LLM-written risk assessment. Sole author — all 45 commits.
The design principle is that most of a pipeline shouldn't need an LLM. Ecosystem detection, update application, build/test, scanning, and PR creation are all deterministic; the model is invoked only for the nine judgement calls that genuinely require it. Total cost lands at $0.001–$0.01 per run.
| Orchestration | LangGraph state machine with conditional routing — analyse → detect commands → prepare → build → test → rollback ⟲ → integrations → security audit → patch → PR/Issue, with an orchestrator agent routing tasks and a typed state object threaded throughout. |
| Binary-search rollback | When tests fail, packages are ranked by suspicion (error mentions + major-bump risk). Low confidence bisects the suspicious half, isolating the culprit in log₂(n) retries instead of n. |
| Nine LLM analysers | Changelog risk · source-code impact · config drift · security triage · reachability · failure root-cause · maintainer PR summary · update grouping · multi-repo synthesis. |
| Reachability analysis | Rather than reporting every advisory, it greps for imports of the vulnerable package and asks whether the vulnerable path is actually reachable from your application — turning CVE noise into a ranked action list. |
| Supply-chain scanning | Seven tools in concert — pip-audit, npm audit, govulncheck, cargo audit, Trivy, OSV-Scanner, Semgrep, Bandit — with fixable CVEs auto-patched and unfixable ones tracked in a persistent GitHub Issue. |
| Breadth | Eight package ecosystems (pip, poetry, npm, pnpm, yarn, cargo, go, plus extras); six interchangeable LLM providers (Anthropic, Gemini, OpenAI, Groq, HuggingFace, Ollama); GitHub access via the GitHub MCP server. |
| Surfaces | CLI, a FastAPI async job API with OpenAPI docs, Docker Compose, and org-wide batch mode that synthesises shared CVEs across repos and ranks which to fix first for maximum blast-radius reduction. |
- Predictive maintenance at scale — 200+ regression and tree-based models over 3,000+ IoT sensors across 50 sites, plus an N-BEATS network for HVAC fault detection. 70% faster training, 60% less unplanned downtime.
- Production GenAI assistant — agentic orchestration on Claude with LangChain tool calling and RAG at scale, served via FastAPI streaming + Next.js. Search time 10 min → 10s.
- Agentic IaC provisioning — a repo-aware agent that proposes an architecture, renders it for review, then auto-provisions post-RBAC. Turnaround 5 days → 1 day.
- GenAI NER on Vertex AI — Gemini 1.5 Flash + RAG with domain-adaptation fine-tuning and RLHF alignment. ~95% extraction accuracy.
📄 These are summaries. The stack, scale, and full context behind each one — plus six years of prior work — are in my résumé.
Multimodal retrieval & recommender systems
| Repo | What it is |
|---|---|
| GenAIFashionStore | CLIP embeddings in a ChromaDB vector store for visual, text, and hybrid search; GPT-driven styling chat; recommendations blending visual similarity, category, and collaborative filtering. |
| fashion-visual-search | The retrieval engine done properly: ResNet50 features, FAISS similarity index, and MMR-based diversification for complementary and complete-the-outfit suggestions. Methodology documented in METHODOLOGY.md. |
| deep-learning-fashion-search | Neural Networks & Deep Learning coursework — an image-based fashion search system, with the written report alongside the notebook. |
Quantum & classical optimisation
| Repo | What it is |
|---|---|
| tsp-quantum-algorithm | MSc research: a QAOA solver for the Travelling Salesman Problem in Qiskit, configurable across problem sizes, optimisers, and penalty weights, running on simulator or real quantum hardware. |
| tsp-playground | The classical control group in C++ — Branch and Bound, Held-Karp, Genetic Algorithm, and Simulated Annealing, benchmarked head to head. |
| quantum-research | Held-Karp baselines and the benchmarking harness that makes the quantum-vs-classical comparison measurable. |
ML & data foundations
| Repo | What it is |
|---|---|
| ClassifierML · Gaussians_Mixture_model | Text + image feature fusion for cuisine classification; GMMs over the Peterson & Barney vowel formant dataset. |
| predictive-analysis · spark-taxi-analytics | Insurance-cost regression; distributed Spark analytics on a JupyterHub/Kubernetes cluster. |
Applied GenAI & open source
| Repo | What it is |
|---|---|
| llm-product-attribute-extraction | LLM-driven product attribute extraction over scraped retail catalogues, running Ollama locally against a category/attribute taxonomy — the working notebooks behind the NER pipeline. |
| GPT5VideoSubtitleGeneration | Video → transcript → translated subtitles, wrapped in a small web service around the processing and translation stages. |
| Ortelius microservices | CNCF/CDF software supply-chain evidence store — the contributions behind the two Linux Foundation awards. |
Before and alongside the ML work — full-stack products, distributed streaming, and platform tooling.
| Project | What it is | Stack |
|---|---|---|
| shop-delivery-manager | Shop and delivery operations platform — manage wholesalers and retailers, build multi-SKU orders, track deliveries, and cost batches. My largest codebase (~650 KB), with business logic pushed into PostgreSQL functions. | React · TypeScript · Material UI · Clerk · Supabase |
| ecommerce-app | Full storefront for Indian and ethnic products — catalogue, cart, real-time auth and orders, split across a Next.js front end and a separate TypeScript API server, with end-to-end tests. | Next.js 15 · TypeScript · Tailwind · Supabase |
| mil-can-platform | MIL-CAN — a network connecting educators, creators and literacy advocates to counter misinformation. | Next.js · TypeScript · Tailwind · Netlify |
| flink-kafka | A Flink streaming job consuming and transforming Kafka topics, packaged as a custom Flink job-cluster image and deployed to Kubernetes with Helm charts. Local stack (Kafka, Zookeeper, KafDrop) via Compose. | Scala · Apache Flink · Kafka · Docker · Kubernetes · Helm |
| PaymentWalletSystemWithJWT | Payment wallet service with JWT-secured endpoints and an Azure Pipelines build. My most-forked repository. | Java · Spring Boot · JWT · Azure Pipelines |
| gmaps-business-finder | Google Maps business-data finder — resolves every postcode in a UK city via postcodes.io, then runs a threaded scrape across them, streaming live progress to the browser over SSE. | Python · Flask · Docker · Railway |
| docker-management-api | REST API for container lifecycle management, wrapping the Docker CLI behind HTTP endpoints, with a companion tester application. | Python · Flask · Docker |
| go-abs | Database abstraction layer built around a datasource factory — swap the backing store (NFT.Storage ↔ PostgreSQL) without touching handler code. One Compose file per implementation. | Go · PostgreSQL · Docker |
GenAI & LLMs
ML & Deep Learning
MLOps & Platform
Data, Backend & Cloud
|
|
|
|
| Top CDF Contributor Award 2024 The Linux Foundation · 2024 |
Ortelius Most Valuable Contributor 2024 The Linux Foundation · 2024 |
Program Committee — cdCon 2026 The Linux Foundation · 2026 |
Program Committee — OSS + ELC North America 2026 The Linux Foundation · 2026 |
- 🥇 Gold Legend — the first-ever recipient of Ortelius' highest recognition, which requires reaching Gold in both the Champion (technical contribution) and Ambassador (community outreach) tracks. Announcement
- 🏛️ Governing Board member, Ortelius — leading the project's machine-learning direction, after starting out by volunteering at an architecture meeting and shipping the Flask backend for the frontend service.
- 🎤 Program Committee for cdCon 2026 and Open Source Summit + Embedded Linux Conference North America 2026 — reviewing and rating submitted speaking proposals.
- 📰 Featured in the CDF Continuous Spotlight interview series, and named Top CDF Contributor 2024.
- 🛠️ Contributor to Ortelius, Jenkins X, and Keptn.
"Don't think too much and don't be afraid to reach out to people."
Microsoft Azure — AI-900 (AI Fundamentals) · DP-900 (Data Fundamentals) · AZ-900 (Fundamentals) · AZ-204 (Developer Associate) · AZ-303 (Solutions Architect)
- 14 Must-Have Claude Code Plugins That Will 10x Your Developer Workflow
- Abstracting out Data Layer Implementation in Go
- Kafka 2.7 in Docker and Spring Boot | Let’s develop a pub-sub application
- Get started with Keptn — Multi-stage delivery with Quality Gates(Demo)-Part 1
- Keptn — Multi-stage delivery with Quality Gates(Demo)-Part 4