Akbar Juraev

Error Feedback in Immersive Voice Programming (DreamCodeVR)

University of Birmingham · Wizard-of-Oz user study

Manuscript in preparation · CHI 2027

When an AI writes your code and it breaks, can you still tell whose mistake it was? In VR, speech is the only tool you have to find out.

When an LLM writes your code and it fails, how do you work out what went wrong if speech is the only tool you have? Spoken commands become Unity C# compiled live in-scene, so a failed generation leaves the user to diagnose and repair by voice alone. The study compares three feedback conditions, no feedback, an explanatory text panel, and an embodied conversational agent, against error attribution, recovery strategy, and trust.

My role. Contributed to the original idea, implemented most of the technical work, and am currently contributing to the design of the user study.

Omni-Connect / DeviceSphere

University of Birmingham · multi-device HCI research system

Targeting CHI 2027

Close your fist over a function on the laptop, open it toward the tablet, and it is there. Three devices, one workspace, no setup.

A laptop, phone, and tablet operating as a single connected workspace. The laptop runs a native Unity app doubling as the coordination server, while phone and tablet join through zero-install WebGL clients over the local network. Cross-device copy-paste is driven by hand gestures captured with Meta Project Aria 2 glasses: close a fist over a function tile to grab it, open it toward a device to paste.

My role. Built the real-time WebSocket-to-UDP bridge, the gesture-streaming pipeline with automatic reconnection, build tooling for four client builds, and the infrastructure for two controlled user studies.

AgenticXR: Safe Agentic Authoring Inside Immersive Environments

University of Birmingham · builds on DreamCodeVR, originally developed at UCL

Targeting IEEE VR

Letting an AI edit the world you are standing in is only safe if it can be tested somewhere else first. Generated code runs against a hidden clone of its target before it is allowed to touch the live scene.

A person in a VR headset creates interactive behaviour by speaking, while AI agents generate, critique, and validate that behaviour before it changes the world they are standing in. The contribution is a runtime architecture keeping a live human and a slower asynchronous agent system coordinated safely, held together by a persistent Shared XR Memory, a Verification Space where generated code runs against a hidden clone of the target before touching the live scene, and five graded levels of agent initiative matched to the risk of each action.

My role. Own evaluation and deployment: the within-subjects study design (five tasks, four hypotheses, NASA-TLX, SUS, IPQ presence, human-automation trust), the study harness merged upstream, Quest deployment, zero-configuration LAN discovery, and the first live speech-to-text validation of the system.

LLM-Designed Affective VR Environments

University of Birmingham · MSc research collaboration

VR & affective computing

Can an LLM design a room that reliably makes you feel calm, and does the shape of the room change the answer?

Whether an LLM, constrained to a fixed pool of interior-design parameters, can compose virtual rooms that reliably induce a target emotion, and whether room shape moderates that effect. Four emotions drawn from the diagonal quadrants of Russell's circumplex model, within-subjects, crossed with two room geometries.

My role. Contributed to ideation and research meetings, and built the technical system: a single-source-of-truth Python pipeline where the prompt, JSON schema, validator, control arm, and Unity C# constants all derive from one definition, a three-layer validation gate ending in C# at scene load, and 73 automated tests requiring no API key or network. Unity procedurally generates both room shells so dimensional constraints are machine-verified rather than eyeballed.

Correlated Perception Error in Semantic 3D Mapping

Independent research · Jul 2026

A 70%-accurate labeller builds a better map (0.560 mIoU) than an 85%-accurate one whose errors repeat by viewing angle (0.502).

Semantic mapping fuses per-frame labels into a voxel map using Bayesian updates that assume each observation is independent. Real cameras violate this: a surface voxel is seen from a narrow band of viewpoints, and segmentation models fail consistently from particular angles, so one mistake is counted as repeated evidence. Holding accuracy fixed and varying only error correlation isolates the effect: a 0.115 mIoU gap and 4× worse calibration error. Standard accuracy benchmarks therefore do not predict map quality, and can rank two models backwards.

My role. Seven CI checks run on every push; four validate the benchmark itself and three recompute the headline numbers, so the README cannot drift from the result.

Policy Evaluation & Statistical Comparison Harness

Independent project · closed-loop policy comparison

A policy that looked 78% better than a do-nothing baseline was 27% worse than repeating the previous action, a rule that fits on one line.

Runs a candidate and an incumbent on identical seeded scenarios and returns three verdicts rather than two: better, worse, or underpowered, the last carrying the episodes needed to settle it. A second finding: a policy with very low action error and 98% gripper accuracy scored zero successes in 240 closed-loop rollouts. Never conditioned on target position, it had learned to imitate the shape of human motion with no representation of its purpose. The offline metric was answering whether the output looked human, not whether it accomplished anything.

Teleoperation Demonstration Pipeline

Independent project · versioned demonstration data

Nine signal-level checks identify a poor demonstration from the numbers alone, with no video review.

Seven stages from raw session to trained, traceable policy. Scoring detects operator hesitation, saturated commands, tracking divergence, gripper chatter, and dropped frames, grading each session gold, silver, or reject so rejects never reach training. Splitting is session-aware, so near-duplicate trajectories cannot straddle the train/test boundary, and every model is pinned to a data revision and a commit. The scorer is validated against a generator simulating four operators of differing skill without being told who is who.

Fleet Deployment and Continual Improvement

Independent project · multi-node learning with canary rollout

Over eight rounds a three-node fleet moves from 32% to 94%, uploading only 16% of collected data.

Built on the premise that a bad update on a robot moves an arm rather than crashing an app. A new policy is proven loadable and within its latency budget on that specific hardware before it goes near the control loop, then switched in atomically with rollback on failure. Each node scores its own attempts locally and uploads only what is worth uploading. New versions reach a canary subset first. The case that motivates the architecture: a regressed model passes every server-side gate, including scoring 100% in the server's own simulator, and is caught only once the canary nodes run it.

LLM Bias Mini-Evaluation

Independent research · Feb 2026

Reproduction of a published evaluation method

Independent reimplementation of a directional gender-bias evaluation method from the literature, applied across BERT, RoBERTa, and GPT-2. Across all masked prompts the top predicted token skewed to one gender in both BERT and RoBERTa, while generation results varied across runs, suggesting instability in how bias surfaces.

Wind-Turbine Health Analytics from Drone Imagery

Turin Polytechnic University in Tashkent · Jun 2026

Production

Computer vision system for wind-turbine health inspection from aerial imagery, classifying turbine conditions and detecting operational modes.

My role. Own the pipeline from dataset preparation through model training and evaluation to a production-ready exportable model, plus client-facing PDF health reports generated from model outputs.