Research,
made runnable.

An instrumented, browser-native research prototype for studying how response latency and progressive refinement affect exploratory analysis. Users cross-filter one million property transactions while switching between immediate sample-based feedback and conventional exact queries.

Experimental variable

Exact results versus progressive sample-to-exact refinement

Research system

Five linked views over partitioned Parquet, queried locally with DuckDB-WASM

Observable evidence

Timings, interactions, and exploration traces can be exported for analysis

A deterministic symbolic program synthesiser that turns small input–output examples into ranked, executable query candidates. It searches a compact transformation language in a worker, keeps every program consistent with the demonstration, and exposes ambiguity rather than hiding it.

Inference engine

Enumerative synthesis across filters, joins, groups, strings, sorting, and projection

Transparent output

Every surviving candidate is inspectable as a structured program and generated SQL

Evaluation harness

Benchmark tasks test intent recovery on full tables the examples never reveal

An interactive business analytics research prototype built from real historical retail transactions. It connects revenue exploration, behaviour-based customer segments, cohort retention and a transparent campaign simulator, allowing users to inspect the evidence and test how assumptions change the decision.

Research dataset

506,073 cleaned merchandise transaction lines from UCI Online Retail

Decision workspace

Linked market filters, customer definitions, cohort evidence and campaign scenarios

Reproducible evidence

Explicit metric definitions, source fingerprints and exportable interaction traces

A working research prototype connecting historical building electricity and weather data to operational and financial decisions. Users compare buildings, inspect overnight patterns and unusual days, evaluate a weather-informed model against a seasonal baseline, and test investment scenarios with every financial assumption exposed.

Research dataset

A reproducible subset of 17 building electricity meters from Building Data Genome 2

Measured comparison

Chronological holdout evaluation against a previous-week baseline on identical dates

Decision transparency

Visible observation coverage, tariff assumptions, savings sensitivity and exportable evidence

A research extension grounded in VRJ, the industrial VR cable-jointing platform developed by Hari Ragavan. VRJ already includes visual and voice instructions, assessment modules and a companion progress dashboard. This browser prototype makes a proposed fading policy inspectable and logs help, errors and successful actions; the new policy has not been integrated into VRJ or evaluated with learners.

Existing system

Hari-developed VRJ with procedural interaction, audio guidance, assessments and dashboard reporting

Working extension

Fixed versus adaptive guidance policy, independent-mode demonstration and exportable browser event trace

Evidence boundary

Existing VRJ footage is implementation evidence, not a recording of the proposed adaptive condition

A working instructor-dashboard research prototype using an explicitly synthetic, replayable classroom session. It combines task progress, repeated errors and learner-initiated help requests into an inspectable support queue. Instructors can acknowledge cues, record their intended support action and export the trace. This is a browser prototype, not footage of a connected multi-user VR classroom or evidence of improved learning.

Research interface

Dashboard and baseline views preserve explicit help while varying analytical support

Observable events

Deterministic task-event replay with visible cue rules and recorded instructor acknowledgements

Responsible interpretation

No attention scores, headset-movement proxies, real learner records or invented study outcomes