NOT AN LLM

Natural language in.
Cited verdicts out.

ALFA is a deterministic reasoning engine for regulated decisions. Ask in plain language; get a verdict — and the exact rule and number behind it. Same question, same answer, every time.

Probabilistic AI can’t sign off.

Modern AI assistants improvise. Ask the same question twice and you can get two different answers; now and then one states a limit, with total confidence, that simply isn’t true.

In regulated work — where every decision must be reproducible, traceable, and defensible — that is disqualifying. The decisions that matter most are exactly the ones today’s AI can’t make.

The hard part isn’t deciding the difficult cases. It’s performing, documenting, and defending the high volume of routine, rule-bound ones — quickly, identically, and with a citation every time.

Understanding in. Determination out.

ALFA reads a plain-language question and your case facts, maps them to typed parameters, and checks them against a curated, human-verified rule-base drawn directly from the regulation. It returns a verdict with the exact clause and number that produced it. The language understanding is the flexible part; the determination is fixed, cited, and reproducible. No opaque model judgment sits anywhere in the decision path.

01

Ask in plain language

The way you'd ask a colleague. No codes, no special format.

02

ALFA checks every applicable rule

Not just the first one it finds.

03

Get a verdict with its citation

The rule and number behind it, every time.

It shows its work.

Verdict Trace
Question
“Is Room 4 compliant for Grade A during operation?”
Checks
✓0.5 µm, in-operation: 3,480 — limit 3,520
✗5 µm, in-operation: 24 — limit 20
Verdict
NOT COMPLIANT

One check failed; the larger-particle limit was exceeded.

Cited: EU GMP Annex 1, airborne-particle classification table

A simpler tool checks one limit and says “all good.” ALFA checks everything the rule requires, and shows why.

Daily hygiene checks

Did today's Grade A line checks pass?

Four samples; three clean, the glove showed one colony where the limit is essentially zero.

Action neededEU GMP Annex 1
Releasing purified water

Can we release today's water-for-injection?

Three things must be in range; two were fine, organic content came in at 0.62 against a 0.50 limit.

Don't releaseEuropean Pharmacopoeia
Tablet dissolution

Did the batch pass first time?

Each of six tablets needs ≥85%; one came in at 82%.

Not yet — wider follow-up testUSP dissolution test
Dose consistency

Is the dose even across the batch?

The first ten units scored 16.2, just over the 15.0 ceiling.

Test more unitsUSP uniformity test
Sterile-process trial

Did the media fill pass?

12,000 units filled; target zero contaminated; one positive.

InvestigateEU GMP Annex 1
Equipment cleaning

Release the train after cleaning?

Must look clean and be under the residue limit; residue 4.2 against 3.8.

Not releasedCleaning limit
Air pressure between rooms

Is the airlock cascade right?

At least 10 units between steps; one step fine at 12, the next only 8.

Not compliantSite air spec / Annex 1
A fridge went warm

2–8 °C product reached 11 °C for six hours — a problem?

Inside the approved excursion allowance.

All goodApproved stability rules

Every example is multi-condition — ALFA selects which rules apply, checks each, and combines them. That’s reasoning, not a look-up.

One engine, not one checker.

Across a regulation, the deterministic logic reduces to a small, finite set of reusable rule types — and roughly 80% of deterministic GMP rules are covered by just three of them. As each type is built and validated, it covers a broad, predictable swath of the regulation. The same engine extends to new frameworks by encoding new rule-bases, not rebuilding.

01

Threshold checks

Particle and limit comparisons

02

Time-based requirements

Requalification, retention, stability timing

03

Conditional requirements

If-this-then-that obligations

04

Banded acceptance criteria

Staged and tiered limits

05

Sequence-counting

Run and event tallies against a cap

06

History-scoped impact

Look-back assessments

Next to a probabilistic model.

 A probabilistic modelALFA
Ask the same thing twiceAnswer can varyIdentical verdict, every time
A rule it wasn't givenMay infer a plausible oneRefuses and flags the gap
Audit trailReconstructed afterwardsIs the output: clause + number
Validation under CSVUnsettled, very hardQualifiable as deterministic software
Where the data livesOften the cloudOn your own site
Its jobPredict & optimiseDecide & defend

Built to be defended.

Inspectable, not a black box

A deterministic, traceable engine your validation team can qualify under CSV.

On-device, private

Runs on modest hardware on your own site. Your data never leaves the building.

Knowledge you can query

The rules live in an inspectable rule-base, not baked opaquely into model weights.

We’re precise about the frontier. The engine is proven and bounded. Robustly parsing the messiest multi-clause regulatory language — the natural-language front end — is the active research problem we’re advancing. And every judgment-based requirement — risk assessments, “adequacy” or “suitability” calls — is explicitly flagged, with citations, to stay with a human. ALFA earns its place exactly where probabilistic AI is disqualified.

Where it earns its place.

Quality Assurance & QP

Clear the routine, rule-bound checks so QPs spend time on judgment.

CSV / Validation

A deterministic engine your team can qualify, not a black box to argue about.

Manufacturing & Operations

Instant, cited pass/fail on in-process checks — no waiting for someone to look it up.

Regulatory Affairs

Every verdict traces to the clause. Inspection-ready by construction.

Digital & CDMO / CMO

One engine, multiple frameworks. License per site, deploy on-prem.

A relationship, not an exit.

ALFA is offered on a licensing basis — per site and per regulatory framework — preserving everyone’s independence. Engagements start as a scoped, low-risk pilot that runs alongside your team, never in the release path. We are not seeking acquisition.

Start with a conversation.

A 45-minute scoping call to pick one bounded, high-value area and define what a successful pilot looks like.

Book a scoping call