Future improvements
Ordered by how much they would change an answer, not by how interesting they are to build.
Would change the conclusions
Patent and exclusivity status
The largest gap in the analysis. The system can identify that a space is originator-dominated from its competitive structure, and does so for saroglitazar and inclisiran. It cannot say whether entry is legally possible or when exclusivity ends.
Every entry recommendation currently carries a caveat that a freedom-to-operate review is needed. Removing that caveat would require joining an IP dataset on molecule and geography and adding an eligibility gate ahead of the right-to-win pillar, so a blocked space is excluded rather than ranked low.
Highest value, and the only item here that would change a recommendation rather than sharpen one.
Longer history
Two annual observations support a direction and nothing more. Four or five years of MAT would allow a real trend decomposition, a variance estimate on each growth rate, and detection of whether 2025-26 was anomalous. It would also let the mean-reversion rate be estimated from the data rather than assumed.
The forecast module is written so the base-rate weights and decay parameter are configuration, which is where they would be replaced.
Institutional and tender channel data
The dataset is retail only. India’s public screening programme has put tens of millions of patients on hypertension therapy through primary care, none of it visible here. A space that looks small in retail may be materially larger in total, and the current analysis would systematically understate first-line therapies against premium combinations.
Cost and margin
Every conclusion is about revenue opportunity. Adding transfer price, gross margin and an estimate of promotional cost per share point would turn “this space is attractive” into “this space is attractive and worth funding”, which is the question a commercial leader actually asks.
Would sharpen the analysis
Field-force modelling
Right to win currently uses observable proxies: molecule adjacency, brand franchise, sub-segment presence. Actual detailing capacity, prescriber coverage and share of voice would make the pillar considerably stronger, and would let the system say how much investment a share target implies rather than only whether a route exists.
Geographic and specialty granularity
State-level or specialty-level splits would show where a national opportunity concentrates. Hypertension prevalence ranges from roughly 15 per cent to over 40 per cent across Indian states, so a national average conceals a lot.
Brand-level right to win
Adjacency is measured at molecule and umbrella-brand level. Modelling prescriber overlap between specific brands would distinguish a genuine extension from a nominal one.
Scenario modelling on the framework itself
The sensitivity analysis perturbs weights. A useful extension would perturb the inputs: what if volume growth in a space were half what is reported, or the leader defended aggressively on price. That moves the question from “is the ranking robust to my weights” to “is the recommendation robust to the world behaving differently”.
Would strengthen the system
Systematic signal curation
Fourteen documents chosen for relevance is curation, not a literature review. A defined search protocol, inclusion criteria and a refresh schedule would make the corpus defensible as evidence rather than as illustration. Signal magnitudes are currently assigned by judgement and would benefit from a stated rubric.
A dense retrieval backend
Not needed at fourteen documents, and the interface already accepts one through SignalRetriever.register_dense_ranker. Worth adding if the corpus grows past roughly a hundred documents, at which point lexical retrieval starts missing paraphrases.
Claim-level verification
Numeric grounding checks that every number appears in the evidence. It cannot detect a real number attached to the wrong space, or a correct figure used to support a conclusion it does not support. Verifying claim triples, of the form (space, metric, value), against the evidence would close that gap and is the most valuable remaining guardrail.
Multi-turn analytical sessions
Conversation memory currently stores intent and which spaces were discussed, deliberately excluding figures so a number cannot survive past the guardrail that checked it. A richer session model would let a user refine a shortlist across several turns, which is how this analysis would actually be used.
Automated regression on the scorecard
The test suite verifies the framework computes what it says it computes. A snapshot test on the full scorecard would catch an unintended ranking change from an unrelated edit, which is the failure mode most likely to go unnoticed.
Deliberately not planned
A larger model, or fine-tuning. The model does not produce numbers, so a better model produces better prose and nothing else. The constraint on answer quality is the data, not the language model.
More space levels. Six already produce 326 spaces, of which 108 survive filtering. More levels would add combinatorial noise rather than insight.
A general-purpose pharmaceutical analytics platform. This is built for one dataset, one therapy area and one set of questions. That focus is why the tools can return interpreted metrics rather than raw rows, and why the numeric guardrail is enforceable at all. Generalising it would cost both.
An LLM-as-judge evaluation. Cheap to produce and impossible to defend when asked how the judge was validated. Every current metric is a deterministic assertion, and that property is worth more than the additional coverage.