Savvyn Abstract Builder v2

Catherine Del Vecchio Fitz
Oct 16, 2025
5 min read

From Proof to Practical Use

TL;DR

Savvyn Abstract Builder v1 showed unpolished notes could be turned into structured abstracts. Now, v2 has implemented standardized tone with curated style guides and embedded conference rules and study-type templates so compliance is automatic. But here’s the epiphany: speed isn’t the point — trust is. Abstracts are permanent record, and scaling to protocols or regulatory docs requires proof of absolute reliability. Next up for us: validate accuracy, pilot unstructured uploads, and migrate to HIPAA-ready infrastructure.

Quick Recap: What We’re Building

The Abstract Builder is being developed in tiers—each layer adds reliability and functionality. Think of it as scaffolding: each tier builds on the last to make outputs more consistent, compliant, and trustworthy.

Tier 0 proved the concept:

Landbot intake → Zapier orchestration → Airtable storage → ChatGPT generation
Shorthand notes in, structured drafts out
Word/character limits enforced, missing inputs flagged
Even messy bullet points produced scientifically accurate outputs

That validated the core idea: capture the right inputs, and AI handles formatting and structuring. Scientists don't need to spend nights in Word.

This post covers Tiers 1 and 2—the infrastructure that transforms a proof-of-concept into something researchers can actually trust and use.

Tier 1: Locking in Scientific Voice

Following Tier 0, Tiers 1 and 2 build the scaffolding to make this reliable and scalable. Tier 0 outputs were accurate but the presentation drifted. Backgrounds turned generic when inputs were thin. Conclusions occasionally oversold findings.

Tier 1 introduced structured style guidance. Each study type now comes with curated DO/DON'T rules and standard terminology. Phase I abstracts focus on safety and dose escalation without overselling efficacy. Phase III abstracts emphasize definitive endpoints and statistical power. Retrospective analyses hedge appropriately on causality.

This isn't micromanaging the AI—it's codifying what "Savvyn voice" means: precise, hedged when appropriate, and free of vague filler that signals AI authorship. The result: outputs that maintain professional tone regardless of input quality. Less drift, fewer revisions, more predictable baseline quality.

Tier 2: Embedding Conference Compliance

Scientific accuracy isn't enough. Abstracts have to clear submission portals: character limits, required sections, title formatting, table restrictions. Miss one rule and you're fixing it at 11:47 PM on deadline day.

Tier 2 embedded conference-specific retrieval. The system now stores submission requirements for major conferences. When a user selects a venue, it automatically enforces:

Title requirements and whether they count toward caps
Mandated section headings
Word or character limits
Table and figure rules
Formatting quirks specific to each portal

Instead of manually checking 40-page submission manuals or re-entering guidelines for every abstract, the rules live in the system. You set them once, they apply forever. This is the difference between a demo and a tool researchers actually use.

Trust as the Foundation

Early pilots suggest meaningful efficiency gains from Tier 2, though systematic measurement is still pending. But speed isn’t the primary goal yet. Trust is.

An abstract isn’t practice—it’s often the first public disclosure of your data, presented at major conferences, published in proceedings, and cited in the literature. Errors don’t just cause embarrassment; they damage scientific credibility, harm careers, and erode institutional reputation. A miscalculated percentage presented in an ASCO oral session doesn’t fade quietly—it follows you.

Skepticism in life sciences isn’t a hurdle — it’s the default operating mode. Scientists are trained to question every number, every conclusion, every method. Regulators assume mistakes until proven otherwise. In that environment, one wrong percentage or a fabricated subgroup doesn’t just dent confidence — it ends it. That’s why trust isn’t a feature for Savvyn, it’s the gating factor for adoption. Without absolute reliability, no one in this field will use it, no matter how fast or slick the outputs look.

That’s why Tiers 1 and 2 focused on reliability infrastructure before efficiency optimization:

Style guides prevent overconfident conclusions that can’t be supported
Conference retrieval ensures submission requirements are met without exceptions
Validation catches missing data before generation
Every number, every statistic, every claim must be traceable to source input

The unstructured file pilot raises the stakes further. When researchers upload raw ClinicalTrials.gov exports and data tables, the system has to parse, validate, and preserve every number exactly as provided. No rounding, no approximations, no hallucinated subgroups. Early results are promising—but this is where trust is won or lost.

Abstracts are presented publicly, scrutinized by peers, and become permanent scientific record. Protocols and regulatory submissions carry even higher stakes. We can’t scale to those use cases without proving absolute reliability first. Trust isn’t built by moving fast—it’s built by getting it right, consistently, every time.

What We’re Measuring

Next on our list: systematic validation focused on trust over efficiency.

Numerical accuracy: every number in the output must match the input exactly. No rounding, no missing decimals, no invented values.
Zero hallucination: every claim traceable to a user-supplied source. Any invented subgroup or endpoint is an automatic failure.
Portal compliance: first-pass success rate for meeting word/character caps, required fields, and formatting rules.
User trust: would the draft be submitted without rewrite? do users return for repeat use? do PIs sign off without line-by-line edits?
Revision burden: number and type of edits required to reach submission-ready status (substantive vs. cosmetic).
Efficiency (secondary): hands-on time versus turnaround time once data is ready.

Not all errors are equal. Formatting quirks are annoying. A miscalculated percentage is serious. An invented subgroup is disqualifying. That hierarchy matters as much as the raw metrics.

Beyond Structured Fields

Structured intake is a useful scaffold, but it’s not reality. Researchers don’t keep their work in tidy bullet points — they keep:

ClinicalTrials.gov exports with registration details.
Interim DSMB tables with response and safety signals.
Annotated spreadsheets with enrollment and baseline data.
Lab slides with half-formed conclusions scribbled in notes.

That’s what has to work next.

Early pilots already test this: drop in a registry export and the system pulls indication, endpoints, enrollment. Upload an interim table and it extracts efficacy and safety. Attach messy slides and it surfaces methods and key findings.

🚨 Spoiler alert: the early runs are promising. But this is where validation must be uncompromising — parse accuracy, calculation precision, completeness checks. Without that, unstructured input is just a parlor trick. With it, you have a real workflow replacement.

From Sandbox to Scale

The current stack — Landbot, Airtable, Zapier, GPT — is good enough for abstracts and manuscripts. These don’t contain PHI, so HIPAA isn’t triggered. That makes iteration cheap and fast.

But the scope is bigger. Protocol synopses, consent forms, regulatory submissions, patient-facing materials — all require HIPAA-compliant intake, orchestration with audit trails, and secure AI deployment. The workflow logic doesn’t change — only the environment does.

That’s what comes next:

Systematic trust validation — measure accuracy, compliance, and user confidence across study types and conferences.
Unstructured file handling — extend pilots with registry exports, interim tables, and messy lab slides.
HIPAA-ready deployment — migrate the validated workflow into secure infrastructure for clinical document use.

If these hold, abstracts are only the entry point. The same framework extends to manuscripts, protocols, regulatory submissions — anything repetitive, structured, and high-stakes enough that mistakes matter.

Closing Thought

Tier 0 proved the idea worked. Tier 2 shows it can be used. The lesson along the way: speed isn’t the point — trust is. Abstracts are permanent record, and protocols or regulatory docs raise the stakes even higher. The only way forward is proving absolute reliability, every time.

👉 Early access is open. If you're submitting abstracts to major conferences or want to pilot unstructured file uploads with your own data, reach out.

Thanks for reading,

—Savvyn (your partner in ruthless efficiency)

🌀 ➝ 📊 ➝ 💡 ➝ 🚀

From Chaos to Clarity. Amplify Your Impact.

Savvyn