Generate, refine, search, and review — in one conversation.
Asthra is the chat-native studio for regulatory writers. Plan, retrieve, draft, cite, hyperlink, QC, and hand off — every step inside the same chat. Or initiate a run, step away, and come back to a finished bundle.
What makes Asthra a studio, not a chatbot.
State-backed authoring means the .docx you see is a render of structured backend state — not the source of truth itself. Four properties make it work in practice — and together they are what the parity-language category (“closed retrieval”, “sentence-level citation”, “audit ledger”) does not on its own deliver.
Word renders. Asthra holds the state.
The Dossier State is Asthra's backend representation of the document — every section, claim, retrieval, approval, edit. Word is one render of it; the audit ledger, the citation graph, and the ability to replay any moment of the run all fall out of having state as the truth.
Walk-away → bundle, not transcript.
Initiate, step away. Asthra works through plan, retrieve, draft, and QC in the background — typically 60–120 minutes. Come back to one audit-ready bundle: the draft with ledger embedded, the citation graph, the QC report, the run metadata. Reopen the .docx and the studio reloads with the full agent history — not a chat transcript you have to dig through.
Cross-section consistency by the citation graph.
Every claim is bound to its source through the citation graph. Within a single drafting run, that means the same statistical value stays the same across §3, §9, and §13; defined terms drift-free; cross-references that actually resolve. Cross-document live-dossier propagation — a source change re-flowing into every affected draft — is the next horizon, not what ships today.
A skill library, not a vendor feature list.
Skills are not features tacked onto a chat. They are first-class objects that mutate the same Dossier State the agent does. Your team adds new skills — turn any freeform chat into a saved /command, available across every run. Reviewer personas — FDA Clinical, CMC, CER Device, CSR Writer QC — are on the roadmap, built on the same skill substrate.
The five-step loop, on top of state-backed authoring.
Plan → Retrieve → Draft → Review & refine → Hand off. Each step mutates the Dossier State; the .docx in Word re-renders. The agent does steps 01–03; the writer drives step 04 with skills and mid-draft tools; step 05 is the seal-and-deliver.
Asthra plans which sources to use for each section.
Before generating any text, Asthra produces a plan: which source documents it intends to consult for each section of the deliverable. Writers review the plan, swap sources, exclude material out of scope, and approve.
- ✓
Section-by-section retrieval mapping with source-file pinning
- ✓
Edit, swap, or exclude sources before drafting begins
- ✓
Plan is captured in the audit ledger and reproducible
Asthra pulls excerpts from the approved sources.
Asthra indexes the files you upload and pulls the right excerpts via semantic retrieval. No unsupervised internet access, no fallback to pretrained knowledge, no cross-tenant leakage — only the documents you specified. If an external lookup is genuinely needed, Asthra pauses and asks the writer to approve a specific request before fetching anything; the approval and the result are recorded in the audit ledger.
- ✓
Section-aware semantic search, scoped to your document set
- ✓
Tables, figures, and structured data ingested alongside narrative
- ✓
Every retrieval returns a passage, page, and offset — recorded in the ledger
Asthra writes each section from those excerpts, with citations.
Asthra drafts each section against module-specific writing instructions your team validated during onboarding. Every claim is bound to a retrieved passage. Where data is missing, Asthra surfaces an explicit gap flag rather than inventing plausible-sounding text.
Mid-draft, refine through chat: tighten a paragraph, swap a source, ask the agent to find recent literature on PubMed, ClinicalTrials.gov, or bioRxiv to support a specific claim. Every external lookup is writer-approved and ledger-logged.
- ✓
Module-specific writing rules (ICH E3, ICH E2C, EU MDR, ICH M4Q) applied per document type
- ✓
Two-level citations: document-level by default, sentence-level on demand
- ✓
Mid-draft literature search across PubMed, ClinicalTrials.gov, bioRxiv — approval-gated
- ✓
Inline gap flags for missing or contradictory source data
Asthra brings the writer in to review and refine the draft.
The agent has produced a first draft. The writer takes over inside the studio: every claim, every citation, every gap flag surfaces on the section spine. Refine through chat, or run a skill over any section (“/verify-citation”, “/rephrase-concise”, “/match-style-of §7”) — each returns a structured result the writer accepts or rejects.
Every refinement mutates the Dossier State and is recorded in the audit ledger. Reviewer personas — a reviewer’s structured findings (severity, rubric anchor, suggested action) summoned over a section — are on the roadmap, built on the same skill substrate.
- ✓
Run skills over any section — /verify-citation, /rephrase-concise, /match-style-of
- ✓
Save your own freeform chats as new skills the team can re-use
- ✓
Every refinement audit-logged on the section spine
- ○
On the roadmap: reviewer personas — FDA Clinical, CMC, CER Device, CSR Writer QC — for rubric-anchored findings
The draft is delivered with the audit ledger embedded.
The draft lands in Word with an immutable transaction ledger. Every plan change, retrieval, draft step, gap flag, and edit is recorded with a timestamp, actor, and target section. The ledger stays with the document — independent of Asthra — so audit readiness survives the vendor.
- ✓
Append-only ledger embedded in the .docx as a side-channel attachment
- ✓
Each citation resolves to a file, page, and exact passage
- ✓
Track changes capture every writer edit alongside the agent record
What you can do mid-run — without leaving the studio.
The five-step loop is what every run does. The capabilities below are composed on top of it — walk-away drafting, skills you run mid-draft (with reviewer personas on the roadmap), table and figure analysis in the chat, hyperlinks resolved at write-time, end-of-run QC and an audit-grade run bundle. Click a chip to jump to it.
Initiate the run. Step away. Come back to the studio, fully briefed.
Start a CSR or PSUR run, close the laptop, head to a meeting. Asthra works through the plan in the background — retrieving evidence, drafting sections, resolving cross-references, running the QC pass, packaging the bundle. Return in 1–2 hours to a first draft with the bibliography, hyperlinks, and quality report all in place.
The draft is a starting point, not the finish line. Regulatory work needs the writer's judgement — and the studio is where that judgement gets applied. Open the document and the studio loads everything the agent did while you were gone: the full chat history, every retrieval, every tool call, every gap flag, every QC finding. Continue the conversation as if no time had passed. Refine §9.4, swap a source, regenerate a paragraph, ask the agent why it chose a particular citation, summon a reviewer persona over the section that came back flagged. Walk-away delivers a fast first draft; the studio is where you take it to submission-ready.
- ✓
Background execution with progress visible per stage
- ✓
Approval gates surface as notifications when the agent needs your input
- ✓
Resumable — interrupt the run mid-stage, resume from the last checkpoint
- ✓
Bibliography, hyperlinks, and QC report all in place when you return
- ✓
Native continuity — the studio loads the full agent history when you reopen the document; nothing is lost between the run and your review
✓ What it is
A background-execution mode that produces a first draft plus bibliography, hyperlinks, and a QC report in 1–2 hours, with the full agent history available in the studio when you return.
× What it is not
Autonomous AI submission. The draft is a starting point for the writer's review, not the finished deliverable. Regulatory work still needs the writer's judgement.
Author skills your whole team can run — reviewer personas are next.
Skills are live today: run /verify-citation, /rephrase-concise, or /match-style-of §7 over any section, and turn any freeform chat into a saved /command your team re-uses across every run. Each skill is a first-class object — it mutates the same Dossier State the agent does, and every invocation is recorded in the audit ledger.
Reviewer personas are on the roadmap, built on the same skill substrate: reviewer perspectives — FDA Clinical, CMC, CER Device, CSR Writer QC — summoned over a section to return rubric-anchored findings (severity, snippet, suggested action). SME-authored, so your domain experts shape them as the rubrics mature.
- ✓
Skills today — run over any section, ledger-logged per invocation
- ✓
Save your own freeform chats as new team skills — no Python PR required
- ○
On the roadmap: reviewer personas with rubric-anchored structured findings
✓ What it is
Skills writers run mid-draft, authored as markdown with YAML frontmatter, so domain experts add new ones without engineering involvement. Reviewer personas extend the same model — on the roadmap.
× What it is not
A replacement for human review or for the regulatory affairs team's sign-off. Skills and (later) personas make the human review faster, not bypassed.
Read more: From prompt engineering to standards engineering →
Analyse huge tables in the chat — with code you can audit.
Adverse-event line listings, sales and exposure data, lab parameter shifts — the data behind a CSR or PSUR runs into thousands of rows. Most writers spend hours filtering, pivoting, and reformatting in Excel before they can write a sentence about it. Asthra's data analysis agent handles that step inside the chat: ask for a frequency table by system organ class, a cumulative incidence over time, a sales-exposure chart for the reporting period — the agent does the analysis and returns the table or figure ready to insert into the draft.
The analysis is run as deterministic Python code, not LLM-fabricated arithmetic. The LLM proposes the analysis plan; Python executes it against your line listing. Every step — the data slice, the filter, the aggregation, the chart spec — is logged in the audit trail with the exact code that produced the output. If a reviewer asks “how was this Grade 3 AE table generated,” the answer is a Python script you can re-run.
- ✓
Deterministic Python execution against your data — no LLM-fabricated numbers
- ✓
Use cases: AE line listings, lab shifts, exposure / sales data, demographics
- ✓
Outputs include aggregated tables and figures (bar charts, line plots, forest plots) ready to drop into the draft
- ✓
Every analysis step audit-logged with the exact code and inputs — fully reproducible
✓ What it is
Agent-driven analysis of line listings, sales data, and lab parameter files using deterministic Python — every step audit-logged with the exact code that produced the output.
× What it is not
An LLM doing arithmetic. And not a substitute for the statistician on pre-specified primary or secondary analyses — those still come from the SAP.
Cross-references wired up as you write — not weeks later.
Every “see Table 14.3.1.2” or “as discussed in §11.4.1” is pinned to a stable anchor at draft time. When sections move, numbers re-derive automatically — references stay correct because the link is to a structural identifier, not a number that might shift. eCTD validation rule B21-ready by construction.
- ✓
Deterministic anchor IDs derived from caption text + section number
- ✓
Word bookmarks and resolved hyperlinks written into the .docx at handoff
- ✓
Section reorder safe — numbers and links re-derive on rebuild
- ✓
EU eCTD B21 validation, FDA PDF spec, EMA two-click rule — all green by default
✓ What it is
Cross-references bound to stable structural anchors at write time, with real Word bookmarks and resolved hyperlinks shipped in the .docx.
× What it is not
A publishing-phase fixup. Sections can be reordered without breaking links; numbers re-derive from current document state, not from a parser running over prose weeks later.
Read more: Hyperlinks belong in the draft, not the publishing phase →
Every run ships a QC report — and an audit-grade bundle.
When the agent finishes drafting, Asthra runs a structured quality pass: cross-reference integrity, statistical-claim source binding, defined-term ordering, citation density, gap detection, and hyperlink resolution. The result is a structured QC report packaged alongside the draft — one place to triage every flag before review.
Every run produces an audit-ready package: the draft itself, the full trail of how it was produced, the QC report, and the reproducibility metadata for the run. The package is portable, survives the vendor, and is the deliverable you can hand to an inspector.
- ✓
QC checks span integrity, sourcing, density, gaps — schema-typed findings
- ✓
Bundle is reproducible by replay from the event log
- ✓
Traceability survives the vendor — the .docx and JSON are yours to keep
✓ What it is
A deterministic quality pass over six structural checks (cross-reference integrity, source binding, defined-term ordering, citation density, gap detection, hyperlink resolution), packaged with the draft as a portable run bundle.
× What it is not
A copy edit or a prose-quality review. QC checks structure; the writer handles substance in the studio — reviewer personas, on the roadmap, will assist that pass.
Built for regulated writing — every layer earns its keep.
Foundation models alone don't ship in regulated environments. Asthra adds the chat-native studio surface, the orchestration layer, the closed-system retrieval, the persona library, the citation graph, the end-of-run QC, and the audit ledger that turn a capable model into an audit-ready writing agent.
See Asthra
on your documents.
30-day pilot. We benchmark Asthra's output against your existing process — and show you the writer-hours saved, line for line.