Blue MatterSI

Applied AI for Life Sciences / Est. 2026

Start a project
N° 001 bluemattersi.com Pharma · Biotech · Medtech Scroll to open ↓

Medicine has a language.

BM·SI
500 mg

We build AI that is
fluent in it.

Custom, pharma-tuned models and agents for the commercial, medical and market access teams who can't afford a confident wrong answer.

Scroll
§ 1The problem with generic AI

General-purpose AI has never derived a line of therapy, read a payer policy, or sat through an MLR review. So it guesses. In pharma, a confident guess ends up as a number in somebody's board deck.

Fig. 1

Anatomy of a
pharma-grade AI system.

01 · Your data
02 · Definitions
03 · Pharma-tuned model
04 · Agents & tools
05 · Evaluation & guardrails
Your data, untouched. Claims, EHR, specialty pharmacy, CRM and payer documents. De-identified, and it never leaves your cloud.
Definitions as code. Lines of therapy, market baskets and KPIs are versioned, testable rules, not paragraphs inside a prompt.
A model that knows the field. Open-weights models fine-tuned on therapy-area language and your terminology, then distilled to be small, fast and yours.
Agents that don't invent numbers. Specialist agents call deterministic tools. Every figure traces back to a query and every caveat is surfaced.
Proof, not promises. Expert gold sets, calibrated confidence, a human review tier and a full audit trail on every answer.
01 / 05

§ 2 — Core AI

The elements of
pharma-grade AI.

Train Build Trust
01FtFine-tuning

Supervised fine-tuning with LoRA and QLoRA on open-weights models, trained on your therapy areas.

02PaPreference alignment

DPO and reward modelling against expert feedback, so the model answers the way your analysts would.

03DiDistillation

Frontier-model quality compressed into a small, private model that is fast and cheap to run.

04SySynthetic data

Persona-driven conversations and rare-case generation where real labels are scarce or sensitive.

05RgRetrieval

Hybrid search that checks permissions on every query and cites every passage it returns.

06AgAgents

Multi-agent orchestration with typed tool calls, MCP servers and deterministic decision steps.

07DxDocument AI

Vision-language extraction that turns labels, policies and PDFs into strict, validated JSON.

08TxText-to-SQL

Governed analytics over Snowflake and Databricks, with a golden question set and per-step scoring.

09EvEvaluation

Expert gold sets, LLM-as-judge and regression suites that run on every change.

10CfCalibration

Confidence scores you can route on, so uncertain answers go to a human reviewer.

11GrGuardrails

PHI redaction, claim checks, refusal policies and a full audit trail.

12OpLLMOps

Serving, monitoring, drift detection and cost control, all inside your cloud.

Fig. 2 — Inside a fine-tuning run

Same question.
Two very different models.

train.pystep 0 / 1200
  1. $ base = load("open-weights-8b", quant="4bit")
  2. $ data = curate(lot_cases=18_400, synthetic=6_200)
  3. $ lora(r=64, targets=[q,k,v,o,gate,up,down])
  4. > sft  epoch 1 · loss 2.14 → 1.02
  5. > dpo  pairs 3_900 · reward margin +0.41
  6. > eval gold_set=312 · pass
eval loss2.140
0noise floor1200
Prompt

EGFR+ mNSCLC patient on osimertinib. Progression at month 14, then carboplatin + pemetrexed starts 70 days later. Which line is the new regimen?

Generic LLM✕ Inconsistent

This appears to be a continuation of first-line therapy. Chemotherapy is often added to targeted therapy, so the patient is likely still on first-line treatment, though this may vary…

BM·SI pharma-tuned✓ Schema-valid
{ "line": 2,
  "trigger": "regimen change after progression",
  "rule": "client_v4.2 · gap 63d",
  "confidence": 0.94,
  "evidence": ["PD m14", "carbo+pem d70"] }
Schema-valid output
58% → 100%
Line exact-match
61% → 89%
Cost per 1k calls
−92%
Median latency
2.1s → 0.3s

Illustrative benchmark. All figures are invented for this example.

§ 3 — What we build

Eight practices.
Built for the way pharma actually works.

Keep scrolling →

01

Fine-tuning

Pharma-tuned language models

Small, private models trained on therapy-area semantics, your brand vocabulary and your commercial definitions. They're cheaper, faster and more consistent than a frontier API, and they run where your data already lives.

Methods
SFT, DPO, LoRA, distillation
Ships as
Open weights in your VPC
Measured by
Expert-annotated gold sets
02

Real-world data

Patient journey & line-of-therapy AI

Line-of-therapy derivation, cohort building, persistence and switching on claims and EHR data. Each patient-level decision traces back to its evidence, and every definition carries a version number.

Covers
LOT, regimen, progression
Also
Note-derived biomarkers
Finds
Undiagnosed & at-risk patients
03

Market access

Payer & policy intelligence

Agents read payer policies, formularies and prior-authorisation criteria and turn them into structured records you can compare across plans. When a policy changes, the record updates and you get an alert.

Extracts
PA criteria, step edits, tiers
Watches
Plans, PBMs, policy diffs
Feeds
Pull-through analytics
04

Omnichannel

HCP targeting & next-best-action

Segmentation, propensity and next-best-action that field teams actually trust, because each recommendation comes with a plain-language reason. Includes a rep copilot for pre-call planning and call summaries.

Builds
Segments, propensity, NBA
For reps
Pre-call briefs, summaries
Ranks
Field alerts by value
05

Medical · Legal · Regulatory

MLR review copilot

Pre-screens promotional material against approved claims, label language and references, and flags fair-balance gaps before the committee sees the piece. The result is fewer review cycles, and the reviewers stay in charge.

Checks
Claim ↔ reference matching
Flags
Fair balance, label drift
Reports
Review-cycle analytics
06

Safety

Pharmacovigilance & AE detection

Finds adverse events in call notes, patient-support programmes and social listening. Supports case-intake triage, deduplication and MedDRA coding, with a human signing off every case.

Detects
AEs in unstructured text
Triage
Intake, dedup, coding aid
Monitors
Emerging signals
07

Analytics

Conversational brand analytics

Ask your commercial warehouse a question in plain English. Answers are tiered by confidence, checked nightly against a golden question set and linked to the exact SQL behind each figure.

Engine
Text-to-SQL, regression-tested
Writes
Data narratives, cited
Governs
KPI & definition catalogue
08

Insight

Research & competitive intelligence

Agents code primary-research transcripts against your codebook, mine MSL insights and track competitor trials, launches and label changes. Your team gets a short, cited brief every Monday.

Codes
Transcripts at scale
Mines
MSL & field insights
Tracks
Pipelines, launches, labels

Fig. 3 — A real question, end to end

It answers like
your best analyst.

That means showing its working, citing the definition it used, and telling you what you can't yet put in front of a client.

Brand lead

What was second-line share for our brand in EGFR+ metastatic NSCLC last quarter? Can I show it to the client?

BM·SI

Cohort first: 1,847 patients qualify. 304 are excluded because the setting can't be determined from their records.

↳ shows the denominator

Under definition client_v4.2 (63-day gap, maintenance doesn't start a new line), second-line share is 23.4% (CI 21.5–25.3).

↳ names the definition

The vendor figure of 27.7% isn't a conflict. It's fully explained by a 90-day gap (+2.9 pts) and counting maintenance as a new line (+1.4 pts).

↳ reconciles, doesn't argue

Before it goes to the client: 214 patients have an ambiguous line boundary and are queued for steward review. The worst-case swing is 1.9 pts.

↳ surfaces the caveat

Illustrative. All figures are invented for this example.

§ 4 — Across the value chain

From first molecule to the hundredth refill.

01Research & Clinical
  • Trial site selection
  • Protocol feasibility on RWD
  • Literature review agents
02Medical & Regulatory
  • MLR pre-review
  • Medical information drafting
  • MSL insight mining
03Launch & Forecasting
  • Patient-based forecasts
  • Launch-readiness monitor
  • Synthetic HCP personas
04Pricing & Access
  • Policy & PA extraction
  • Formulary change watch
  • Contract analytics
05Commercial & Field
  • Next-best-action
  • Pre-call planning
  • Incentive comp analytics
06Patient & Safety
  • Adherence risk
  • Hub & SP analytics
  • AE detection

§ 5 — Method

A prompt is a

prototype product.

We build the second one.

  1. 01

    Define & measure

    Your experts define the task. We build the gold set and the evaluation harness before writing any model code.

    week 1–2
  2. 02

    Curate the domain

    De-identified claims, documents and transcripts, plus synthetic data wherever real labels are scarce.

    week 2–4
  3. 03

    Fine-tune & align

    Supervised fine-tuning followed by preference alignment, on open models you own outright.

    week 4–7
  4. 04

    Evaluate honestly

    Per-step accuracy, calibrated confidence, LLM-judge plus expert review, and a regression run on every change.

    week 6–8
  5. 05

    Deploy & compound

    Deployed privately in your cloud. Every analyst correction becomes training signal for the next version.

    ongoing
0+

AI use cases mapped across the pharma value chain

0wks

to a first measured pilot on your own data

0%

of our models can run inside your own cloud

0

client records ever used to train a shared model

§ 6 — Ways to engage

Start small. Scale what works.

2 weeks

Opportunity Sprint

We map your use cases, score them on value, feasibility and data readiness, and leave you with a ranked roadmap.

6–10 weeks

Pilot to Proof

One use case on your own data, with a gold set and an accuracy number your stakeholders can sign off on.

3–6 months

Build & Transfer

A production system with monitoring and governance, handed to your team with full code ownership.

Ongoing

Managed AI & CoE

We run, retrain and govern your models, and help you build an internal AI centre of excellence.

HIPAA-alignedGxP-aware validationDe-identified data onlyYour VPC, your weightsSnowflake · DatabricksAWS · Azure · GCPFull audit trail

§ 7 — Contact

Bring us yourhardest question.

Hello, I'm from . We're trying to , and I'd like to know whether AI can help. You can reach me at .