Apr 15, 2026 · London, UK
Sentra is a real-time decision intelligence layer embedded in retail banking frontline interactions. It transforms fragmented customer data into a dynamic “digital twin,” enabling agents to understand customer context, assess vulnerability, and take the most appropriate action (proactively or during live interaction). Key Features: - Digital Twin: Real-time, contextual view of the customer (financial, behavioural, interaction signals) - Vulnerability Assessment: Explainable identification of financial stress and customer risk with clear guardrails - Intelligent Orchestration: Dynamic next-best-actions, including recommendations, suppression of harmful offers, and tailored guidance - Agent Interface: Real-time prompts, reasoning, and suggested actions during calls
In Common is an AI-powered workforce intelligence prototype that helps large organizations staff teams faster, more fairly, and with better evidence. It sits on top of existing workday employee and role data, enriches static profiles through conversational AI, and uses a layered matching pipeline to connect people to opportunities. Key features: Conversational profile enrichment to capture hidden skills, motivations, working style, and team dynamics Role-to-candidate and candidate-to-role matching Hard filters plus semantic matching plus LLM-powered reasoning Anonymized shortlists by default to reduce bias and fairer Staffability scoring and training recommendations to improve deployability Decision traces and audit logs for explainability and trust
Our project combines an agent-driven development framework (AI-Led SDLC) with a plugin called LegacyBridge to enable interaction with legacy systems that lack APIs. LegacyBridge allows agents to extract context from legacy interfaces and execute actions directly within them using tools like Playwright. This replaces traditional RPA with more flexible, intelligent, and adaptive agent-based automation. Key features include: Context extraction from inaccessible or legacy systems Agent-based interaction (navigation, form filling, workflow execution) Integration with development tools (Jira, Confluence, repos, Stitch & Figma) End-to-end delivery enabled by AI-Led SDLC Reusable across multiple industries and systems
Coachdex is a team intelligence layer on top of Codex. It turns isolated Codex sessions into shared team intelligence and reusable skills by capturing developer activity, extracting patterns, and surfacing four core capabilities: Sharing, Insights, Coaching, and Suggestions. Teams can explore synced Codex sessions across repos and machines, identify repeated workflows, generate coaching recommendations, and promote successful patterns into reusable playbooks, prompt templates, and draft skills.
Our project is an AI-powered travel insurance claim processing platform that automates the full claims journey from document intake to final decision. It uses a multi-agent pipeline to ingest claim files, extract and validate evidence, reason over policy coverage, run compliance and fraud checks, make an approve/deny/escalate decision, and handle payment and customer notification. The product also includes a polished real-time dashboard with live agent activity, history, and replay, making the whole process transparent and easy to follow.
An AI-powered software engineering assurance tool that turns a public GitHub repository into a structured review in minutes. A user submits a repo URL and the app clones the repository, runs a Codex(SDK)-driven assurance workflow, and produces a dashboard-backed report with weighted RAG scores across four pillars: security, maintainability, reliability, and architecture alignment. The output includes evidence-based findings, strengths, quick wins, and next steps, giving teams a repeatable and auditable alternative to manual review panels.
This project focuses on developing a Time and Attendance system that integrates alongside our existing payroll platform, bringing a critical function in-house to improve control, efficiency, and data consistency. The solution includes a user-friendly frontend application where employees can securely log in, view company-specific theming, and receive prompts for upcoming timesheet submissions. Users can enter their working hours, selecting from both project-specific codes and general activity codes, ensuring accurate and flexible time tracking. A key feature of the platform is its ability to integrate multiple data sources. Where available, external systems such as Jira or card-based sign-in systems can supply attendance data. This data is processed through a microservices-based architecture, enabling each data source to be handled independently. As an example, a card reader microservice has been developed that takes raw CSV input and converts it into a standardised JSON format. This format is then consumed by a central backend API, which stores and serves the data for frontend use (currently using JSON as a working model, with database integration planned). This architecture provides a centralised data platform while remaining highly flexible. It allows the system to ingest, transform, and unify data from multiple sources, ensuring consistency across the organisation while supporting different attendance capture methods. Overall, the solution is designed to be scalable, modular, and extensible, creating a strong foundation for future enhancements and broader adoption across teams and clients.
SkillPilot is an AI-powered engineering learning platform that adapts to each engineer’s skill level, role, and growth goals. It helps in onboarding and upskilling engineers across the spectrum identify development gaps and next steps in areas such as technical depth, production readiness, code quality, evaluation rigor, collaboration, communication, mentoring, and leadership. To do that, SkillPilot can learn from various enterprise signals like pull requests, mentor-mentee input, project feedback, retrospective notes, internal knowledge bases, as well as personal info and predefined growth paths. By turning those fragmented inputs into personalized growth journeys, SkillPilot helps organizations move faster toward an AI-native way of working: engineers do not just get generic learning content, but contextual guidance based on how work actually happens inside the company. This also accelerates enterprise rollout of tools like Codex, because it helps teams learn where AI can create leverage, teaches people how to use those tools in role-relevant ways, and reinforces company-specific best practices, guardrails, and feedback loops. In that sense, SkillPilot is not just a learning product; it is an enablement layer for enterprise AI adoption. It uses Trust-by-Design principles so recommendations are explainable, humans stay in control, and feedback continuously improves both learning outcomes and responsible AI usage.
Our project is an artifact and integration-driven modernization framework designed to transform legacy applications into modern architectures safely and incrementally using structured, multi-agent workflows using Codex.
Our project aims to create a set of hierarchical skills for migrating legacy Python code bases using Codex and subagents. By having more modular skills we are able to delegate work from a higher-level abstraction skill to more implementation oriented skills. The goal is to keep the same functionality and feel of the application whilst improving the underlying tech stack.
Upload a proposal deck and walk away. A short while later, you'll have a shareable link to a fully deployed proof of concept on AWS - built to Artefact's engineering standards, ready to demo to a client or hand directly to the technical team for production.
Today, tools like Codex already let you split development into multiple worktrees per feature. But merging those worktrees is still a manual, slow process: you create PRs, wait for human review, and merge step by step. What we’re building is a mastermind layer on top of that. Instead of treating worktrees independently, the mastermind orchestrates them: It manages multiple parallel worktrees automatically It handles merges between them without heavy manual intervention And most importantly, it enables combinatorial testing of features Concretely, if you have Feature A in Worktree A and Feature B in Worktree B, you can instantly spin up a combined environment (A + B), test them together, and validate that they actually work as a system—not just in isolation. The mastermind coordinates these combinations, merges the compatible ones, and keeps everything else isolated and reversible. In short: 👉 From manual, linear merging → to automated, parallel, and composable development. For the demo, we’re using Strudel because it represents a real, tangible codebase—but in a way that’s intuitive. In Strudel, features are like music tracks: bass, drums, melody, etc. Each track can have multiple variations, just like different implementations of a feature. Instead of working on a single version, we generate multiple possibilities per track. Users can then listen to and combine these variations in real time: Bass v1 with Drums v2 Melody v3 with Bass v2 or any combination they want This is exactly like combining features from different worktrees. On top of that, the mastermind orchestrates everything: It manages all these parallel “tracks” (worktrees) It enables seamless combination and testing of them together And it helps surface the best overall composition, not just the best individual parts So instead of validating features in isolation, you experience and evaluate them as a system, just like music. In short: 👉 Features become tracks 👉 Worktrees become variations 👉 And the mastermind becomes the composer optimizing the final piece
TalentLens is an AI-powered CV screening solution that transforms unstructured resumes into structured, comparable candidate profiles. It enables faster, more consistent hiring decisions through standardized evaluations, evidence-backed skill assessments, and transparent scoring. Key features: - LLM-based extraction of relevant skills and requirements directly from job postings - Converts PDF resumes into structured candidate profiles using LLM-based parsing and information extraction - Generates evidence-backed assessments of skills and experience from CV content - Scores candidate fit against role requirements to highlight the most promising matches
Using Codex to prototype a Codex-powered solution focussed on data migrations, particularly around semantic reasoning over legacy databases
Our Project - Codex Morph is a Codex-powered legacy/POC discovery and transformation planning workflow that turns a codebase into a structured, developer-ready modernisation plan in minutes. The product is built as a guided Next.js workspace with a multi-agent planning engine behind it. A user submits a codebase and transformation brief, Codex analyses the system, asks targeted clarification questions, orchestrates specialist section agents, and produces a reviewable transformation output covering architecture, migration strategy, security, CI/CD, testing, integration, product planning, and publish-ready documentation. What makes the approach strong is that Codex is not used as a one-off assistant. It is the workflow engine. It handles large codebase analysis, maintains context across multiple planning stages, and uses compaction patterns to preserve context for later iterations. Each run persists local Markdown artifacts such as shared run context, per-agent `context.md`, and `output.md` files so the system can rehydrate only the relevant context on reruns instead of starting from scratch. That makes the workflow more reliable, iterative, and reusable across delivery engagements. Key features: - Codebase intake from GitHub URL, ZIP upload, or local folder - Guided clarification flow before planning starts - Multi-agent planning orchestration using the Agents SDK - Specialist section agents for architecture, migration, infrastructure, security, testing, database, integrations, and product delivery - Local artifact generation for context persistence and future iteration - Human-in-the-loop review, section updates, and targeted reruns - Publish-ready outputs including hosted report views and structured export paths This is designed as a reusable accelerator, not a one-off demo. The workflow can be applied to different codebases, different transformation briefs, and different partner scenarios with the same structure: - Intake - Clarification - Multi-agent analysis - Review - Publish Because context persisted in local artifacts and outputs are modular by section, the system is well suited to repeated use across discovery engagements, internal accelerators, and future productisation. It can evolve into a partner-facing modernisation assessment tool, an internal engineering accelerator, or a pre-sales and delivery discovery workflow.
Our project is Briefly.ai, a personalized news product for people who need to keep up with the fast-moving agentic AI ecosystem without drowning in noise. It combines a user preference profile with a relevance-ranked briefing feed, so each person sees updates tailored to the topics, vendors, products, and tools they actually care about. Key features include personalized onboarding, saved interest profiles, role-aware summary styles, a filtered briefing feed, and a verification panel that shows the original source, supporting references, and trust signals behind each update.
Codex Compliance is a plugin that extends Codex into a compliance-aware coding agent for regulated industries. It scans codebases against frameworks like HIPAA, GxP, and the EU AI Act, identifies compliance gaps, suggests and applies fixes, and generates audit-ready evidence. By embedding compliance directly into the development workflow, it enables organizations to safely adopt AI coding agents, reduce manual reviews, and accelerate delivery in high-stakes environments. Note: Video shows as 8 min but is only 1.:12min long
Codex Tree is a branch-aware environment for AI-assisted software development. It turns a normally linear Codex chat into a tree of connected conversations so users can split a complex task into separate branches, keep the right context in each branch, and still merge the results back together later. The project includes a VS Code custom editor that visualizes the workflow as a tree and a Python backend that manages the actual branch sessions against a target codebase. The core idea is that software work is rarely linear. A user might ask Codex to plan an app, then break that work into branches such as API, UI, tests, or bug fixes. Each branch inherits the relevant prior context, maintains its own chat history, and runs in its own git-backed worktree so filesystem changes stay isolated. This avoids the common workflow of creating a new feature branch, starting a brand new AI session, and then re-prompting the model with all of the same background context just to continue related work. When work is complete, Codex can merge branches back into a new combined branch using both the branch workspaces and the saved conversation histories as context. Key features include visual tree-based chat editing in VS Code, branch creation from the current tip, the root, earlier turns, or specific nodes, per-branch filesystem snapshots and rollback, persisted session state, Codex-driven branch merges, streaming runtime events for assistant output, commands, tools, and file changes, command and file-change approval handling, and the ability to adjust model and reasoning level across turns. Compared with the regular Codex app fork and worktree logic, Codex Tree adds a much more usable layer on top: branches are visible as part of one connected workflow, branch points are tied to conversation nodes rather than being managed as isolated threads, users can branch from historical points with matching filesystem state, and rolling back a branch restores both the chat history and the files together. In the regular Codex app, once a fork is made the relationship to the original chat is much easier to lose because it appears as a separate chat thread. In Codex Tree, that relationship stays visible as a linked branch in the same view, so related ideas, experiments, and follow-up work remain grouped together instead of being scattered across separate chats. The visual tree also makes it much easier to keep track of what branches exist, what conversations happened in each branch, what was done, and roughly when different lines of work diverged or were completed. Because the interface is graphical and user-friendly, it can also make complex coding workflows more accessible to non-technical or less technical users who may find a purely terminal- or chat-driven workflow harder to follow. The graphical display makes plans, ideas, and workflow structure easier to explain, present, and understand. Together, these features make Codex feel less like a single chat window and more like a structured workspace for managing parallel AI development workflows.
A key advantage for regulated industries: the framework only optimizes the pipeline logic itself — no data ever leaves the client environment. This makes it directly applicable to healthcare, legal, and financial clients who are required to run local, on-premise models due to data privacy regulations. Weaker local models become significantly more capable through autonomous iteration — without any data exposure.
Loqui is an agentic interview platform for capturing human knowledge at scale. It helps teams turn a rough research brief into a deployable interview mission, schedule and run adaptive AI-led interviews in parallel, and convert those conversations into a structured, auditable insight graph. Key features include mission intake from natural language, an Agent Studio for refining interview objectives and target lists, adaptive interview execution over chat or voice, a deployment cockpit for tracking response and completion rates, and an insight layer that links findings back to source interviews so teams can drill down, interrogate the data, and generate outputs like executive summaries or slide decks.
We have built a product which will allow users of applications in the early stages of development to report bugs and issues they find with the system. An agentic system in the background will then attempt to recreate the bug, find the probable cause of said bug and then create a plan to fix the bug. The agentic system will then test the proposed fix, and should it work create a PR on the repository to implement said changes. This ultimately will greatly streamline the process of bug -> patch in the software development lifecycle.
A self-hostable M&A intelligence platform, allowing M&A analysts to perform their required initial due diligence and strategic fit assessment for potential target companies in minutes instead of days. Key features are: 1) Mini-CRM to store data on target companies 2) Automated research workflow that integrated web deep resaearch (powered by OpenAI models) and data from various company research APIs (e.g., SEC, EDGAR, Companies House, OpenSanctions). 3) Ability for user to upload strategy documentation of the acquiring company 4) Ability for user to set and define strategy fit criteria 5) Automated strategy fit assessment workflow that utilizes the automatically generated research, proprietary company strategy documentation and the strategy fit criteria to run an initial strategy fit assessment (powered by OpenAI models).
This project reimagines what it feels like to explore a city on foot. Instead of opening a map, searching for landmarks, and piecing together your own route, the user simply drops themselves anywhere they want to begin. From that moment, the app becomes a live AI tour companion: it understands the chosen starting point, builds a personalized walking or running route around it, and guides the user forward through spoken narration, interactive map pins, and timely nudges as they move.
CXecute accelerates contact center agentification by turning customer support transcripts and enterprise API specs into a deployment-ready Google Gemini Enterprise for Customer Experience GECX package. Instead of spending months manually moving from discovery to deployment, teams can use CXecute to automate the path from customer conversations and backend systems to validated AI agent artifacts. Key features: automated intent discovery from support transcripts API-to-intent mapping from enterprise specs payload flattening for agent-ready context generation of a real GECX package validation and self-repair loops when issues are found gateway bundle generation for deployment readiness live orchestration dashboard with stage status, logs, and agent handoffs
SpringGuard is an agent-native SDLC workflow for reducing a bounded static-analysis backlog in Spring Boot repositories. It uses a real Java codebase, a curated SonarQube issue set, repo-local instructions, and hook-driven validation so Codex can fix a narrow set of issues without widening scope. Key features are bounded remediation in the target repo, fast compile and static-analysis checks inside the agent loop, SonarQube checkpoint rescans, durable repo-local SDLC artifacts under `.agent-sdlc/`, and a reusable plugin that carries the workflow to the next repository.
Domino is a dependency-upgrade impact analyzer for Python codebases. Given a package and a version change, it builds a static dependency graph across the repo, diffs the package API between versions, matches the changed symbols to real code usage, and shows the blast radius across files, callers, tests, and entrypoints. It outputs a conservative upgrade verdict and structured evidence, and uses Codex to explain the findings in plain English.
One command palette for every internal tool. Built for the McKinsey intranet, this product helps teams navigate fragmented internal and proprietary products by turning enterprise tools into plugins inside one unified, system-level command palette with a fast keyboard entry point.
Agentic Data Migration Engine is a Codex-powered demo that turns a legacy user-domain schema into an inspectable migration specification and deterministic SQL. It orchestrates four agentic phases: schema inference from DuckDB fixtures, semantic enrichment grounded in profiling and target-model documentation, mapping generation with explicit transformation intent and approval state, and SQL compilation. Key features include a Kuzu graph-based digital twin, per-phase Markdown reports, JSONL audit logging, a machine-readable approval artifact, and compiled SQL models for dim_user, dim_role, bridge_user_role, and dim_address.
Touchless Control Assistant is a hands-free presentation and meeting controller powered by gestures, voice, and presence awareness. It lets a user move slides, pause or resume, mute audio, and track actions without touching a keyboard or mouse. Key features: Gesture-based slide navigation Voice commands for next, back, mute, start, and stop Away-state detection with visible pause status Unified command router for gesture and voice inputs Live event log with command, action, source, and timestamp Reusable architecture for accessibility and frontline workflows Extendable for other domains
A self-healing PR Guardian
Context-aware discovery and recommendation tool that turns natural-language data requests into actionable guidance for data teams. It analyzes an organization’s existing datasets, schemas, and pipelines to identify what can be reused and suggest what needs to be created when reusable components do not already exist.