Organised by question, not by chronology

Research question

When a practitioner has to choose one open-weight language model out of dozens for a classification task, what evidence actually supports that choice — and how much of the apparent difference between leading models is real rather than noise?

Why it matters

Model selection is usually settled by a single accuracy column on a single benchmark. That column hides three things that matter in deployment: whether the gap between two models is statistically distinguishable, whether a model's confidence means anything, and whether the benchmark still has headroom left to measure with. Getting this wrong is expensive — teams ship a larger model than they need, or trust a confidence score that is not calibrated.

Approach

Broad, controlled zero-shot evaluation rather than fine-tuned leaderboard chasing: a fixed prompting protocol applied across model families and parameter scales, evaluated on multiple datasets at once, and analysed with calibration measures, input perturbations, pairwise significance testing and deployment-cost accounting alongside raw accuracy.

Findings so far

• — Instruction-tuned 3B models can outperform several evaluated 7B base models, so parameter count alone is a poor selection criterion within this benchmark suite.
• — Differences among the leading models on MASSIVE are statistically indistinguishable under pairwise McNemar tests — a ranking by raw accuracy overstates how much separates them.
• — Widely used benchmarks such as SNIPS have saturated and no longer meaningfully discriminate among current open-weight models.

Where it goes next

Extending the protocol beyond single-label intent classification to tasks where label ambiguity is intrinsic, and treating benchmark saturation itself as something to measure rather than discover after the fact.

Research question

How much of the performance of a video model on early detection tasks comes from explicitly representing the interaction between people, rather than from appearance and motion features that any strong backbone already captures?

Why it matters

Papers in this area routinely add an interaction module and report an improvement, but the improvement is rarely isolated from the backbone, the training schedule or the evaluation protocol. Without that isolation the field accumulates architectures rather than knowledge. Early detection raises the stakes further: a system that must commit before an event completes is operating on partial evidence by construction, and its errors fall on people.

Approach

A controlled ablation study rather than a new state-of-the-art claim — hold the backbone, data and protocol fixed and vary only the interaction representation, so the attributable contribution can be reported as a benchmark-scoped quantity rather than a general capability claim.

Findings so far

Nothing published yet. Results will appear here once the associated work is public — no provisional numbers are shown.

Where it goes next

Reporting attributable contribution as a function of how much of an event has been observed, and characterising where interaction representations stop helping.

Research question

What does a retrieval pipeline over a large, heterogeneous corpus of research papers need in order to answer expertise questions — "who here works on this, and what have they actually shown?" — rather than merely returning topically similar text?

Why it matters

Institutional expertise is real but effectively unsearchable: it is distributed across thousands of PDFs with inconsistent metadata. Making it queryable in natural language changes who can find a collaborator, not just how fast an existing search runs.

Approach

An end-to-end ingestion pipeline — scraping, parsing, metadata structuring and embedding generation orchestrated with LangGraph — feeding semantic retrieval over the full corpus, with summarisation grounded in retrieved documents.

Findings so far

• — A corpus of roughly 25,000 research papers across 50+ faculty members runs end to end through the retrieval pipeline in about 30 minutes.

Where it goes next

Deployment for faculty and research groups to query publications in natural language, supporting interdisciplinary team formation.

Research question

If an agent's memory is a context window, it forgets on restart and on handoff. What specifically does an operational agent need to persist — and what changes when memory becomes a queryable, transactional store instead?

Why it matters

In safety-adjacent settings the costly failure is not a wrong fact but a repeated action: recommending the thing that caused damage last time, because the consequence was never available at the moment of decision. That is a memory-architecture problem, not a prompting problem.

Approach

Separating episodic memory (what happened, in order) from semantic memory (what a given action caused), storing both in one database so that retrieval returns a precedent together with its provenance, and making every model-produced recommendation pass a post-model safety check before a human sees it.

Findings so far

• — Keeping episodic and semantic memory in a single database makes retrieval and the human decision that follows transactional against the same rows, removing a consistency boundary from the middle of a safety-critical path.
• — A post-model safety floor — re-checking model output and filtering citations down to memory IDs that were genuinely retrieved — catches failures that prompting alone does not.

Where it goes next

Closing the learning loop by writing back whether an approved action produced its predicted outcome, and evaluating retrieval quality with a labelled set and precision@k tracking.

Research question

What has to be true of the data layer — schemas, provenance, quality classification — before a machine-learning or agentic system on top of it can be trusted?

Why it matters

Most of the difficulty in applied AI is upstream of the model. Standardised metadata, traceability between records, and an explicit notion of record completeness determine whether downstream outputs can be audited at all.

Approach

Building the data layer as a first-class artefact: implementing a formal conceptual information model, deriving entity relationships and classification hierarchies from it, and automating data-quality classification rather than leaving it to reviewers.

Findings so far

Nothing published yet. Results will appear here once the associated work is public — no provisional numbers are shown.

Where it goes next

Extending automated quality classification and traceability across integrated engineering workflows.