Savvyn Abstract Builder v1: Testing Workflows, Tiers, and What’s Next
- Catherine Del Vecchio Fitz
- Sep 25, 2025
- 11 min read
Updated: Sep 26, 2025
Abstracts as the proving ground for rethinking scientific writing
TL;DR
Abstracts are short, painful, and always due yesterday, which makes them the perfect low-hanging fruit to test workflows and tools. So I built v1 of the Savvyn Abstract Builder: a Landbot–Zapier–Airtable–ChatGPT workflow that turns shorthand notes into structured drafts. It flags missing input, enforces limits, and makes the process faster (and less soul-destroying) than starting from a blank Word doc. No, AI won’t replace your science — but it can replace your late-night formatting session (or your PI’s word count panic). And this is just the start.
Why Are We Still Writing Content Ourselves?
In academic science, the ritual has always been the same: you write a draft, your supervisor bleeds red ink all over it, then rewrites half the sentences in their own voice. And yet — the science doesn’t change. The numbers are the same. The conclusions are the same.
So what was the point? Most of the time, it was style preference. Hours arguing over phrasing when we should have been focused on data, interpretation, and next experiments.
And here’s the bigger truth: AI makes the process more of an equalizer. If the data are solid, the draft comes out solid — regardless of whether it’s written by a junior trainee, a non-native English speaker, or a seasoned PI. Good science no longer gets buried under uneven writing skills.
That’s the inefficiency I wanted to test. So I built a first version of the Savvyn Abstract Builder — a workflow that takes shorthand notes and outputs structured, submission-ready drafts.
Pick Your Tier (the Savvyn Way)
Before I show you v1, it’s worth explaining how I think about building this system in tiers:
Tier 0 — Prompt Pack
Smarter prompts that force IMRaD structure, enforce word limits, and tone down over-confident conclusions. Immediate upgrade from raw ChatGPT.
Tier 1 — Style Guide + Examples
Short artifacts (Do’s/Don’ts, glossaries, and “gold abstracts”) that give the outputs a consistent Savvyn voice.
Tier 2 — Tiny Retrieval
A lightweight Airtable knowledge base that injects curated snippets (conference caps, acronyms, phrasing) into every run.
Tier 3 — Custom GPT
Interactive drafting and polishing. Great for demos or quick iteration loops.
Tier 4 — Fine-tuning (future)
Once I’ve got enough gold examples, I can lock in tone and structure permanently.
Each tier reduces friction and improves reliability. Tier 0 is where I started — and abstracts were the obvious proving ground. They’re short, structured, and universal, yet somehow always dreaded: rebuilt from scattered notes and spreadsheets, always due yesterday, and always subject to nitpicky edits that change style, not meaning. If shorthand notes could be turned into submission-ready drafts here, then scaling to manuscripts, clinical protocols, and regulatory summaries was clearly within reach.
So I built the simplest version possible: a proof-of-concept workflow that could take whatever notes I threw at it and turn them into a structured abstract with the right sections, the right limits, and a clear output. Nothing fancy, just wiring tools together to see if the concept worked.
The Workflow I Tested (Tier 0)
Here’s what I built (no PhD in Zapier required):
Landbot submission — user drops in some basic inputs including the word or character limit, indication, study type, methods summary, results summary, and key takeaways.
👉 Key design choice: inputs don’t have to be polished. Bullet points, shorthand, even half-sentences work. You just need to throw the data on the page.
From here Zap runs
AI sufficiency check — each section labeled “sufficient” or “insufficient,” with missing pieces politely (well, algorithmically) flagged.
Airtable record created — because spreadsheets are love, spreadsheets are life.
Draft generated if inputs are sufficient.
Google Doc created — formatted, ready to share.
Email generated to either alert that the abstract is ready or this step could be used to send out the link to the google doc, etc.
If needed, a second Zap Revision loop runs — reviewers add notes, AI revises, new versions append to the same Doc.
Tier 0 wasn’t about polish — it was about proving that raw notes could be converted into a structured draft, reliably and repeatably.
Field Tests: Reverse Engineering
To stress-test the Builder, I flipped the process around. Instead of starting with notes and generating a draft, I took a published ASCO abstract and deconstructed it into the Builder’s input fields — essentially running the workflow backwards. That gave me a set of “inputs” I could then feed back through the Builder twice: once in messy shorthand form, and once as a more polished version.
Here’s the original (thanks to the authors for serving as my unknowing guinea pigs in this test — publicly available content, of course):
And below are the two sets of inputs — one in messier shorthand, and one a more polished version.
Table 1: Inputs (Polished vs. Less Polished)
Less Polished Input | Polished Input |
1. Study Type Retrospective | 1. Study Type Retrospective analysis |
2. Indication / Disease Nasopharyngeal carcinoma, NPC | 2. Indication / Disease Nasopharyngeal carcinoma, NPC |
3. Cohort / Sample Size 45 pts, all had ctDNA analysis, all got systemic tx | 3. Cohort / Sample Size 45 patients with nasopharyngeal carcinoma, all underwent ctDNA analysis, all received comprehensive systemic treatment |
4. Methods Summary logistic regression, looked at ctDNA impact on ORR | 4. Methods Summary ctDNA analysis performed, logistic regression used to assess factors impacting objective response rate ORR |
5. Results Summary increase ctDNA → higher risk non-ORR OR=1.051 95%CI 1.001–1.104 p=0.046, persistent ctDNA+ = poor outcomes, ctDNA clearance = better outcomes, ctDNA + EBV DNA helps screen NPC, ctDNA higher in N3 vs N0–N2, ctDNA higher in EBV+ vs EBV–, ctDNA correlated w lymph node load GTVnd, ctDNA vs tDNA mutations concordant esp in common NPC genes | 5. Results Summary Increase of ctDNA associated with higher risk of non-ORR OR=1.051 95% CI 1.001–1.104 p=0.046, persistently ctDNA-positive patients had poor outcomes, ctDNA clearance associated with better outcomes, ctDNA assists EBV DNA in screening for NPC, ctDNA levels higher in N3 vs N0–N2 stage patients, ctDNA levels higher in EBV positive vs EBV negative patients, ctDNA levels correlated with lymph node load GTVnd, ctDNA and tDNA mutation profiles showed strong concordance especially for frequently mutated NPC genes |
6. Key Takeaway Message ctDNA useful biomarker for systemic tx in NPC, tied to stage and lymph node load, can guide tx escalation/de-escalation, minimally invasive personalized marker | 6. Key Takeaway Message ctDNA is an effective biomarker for systemic treatment in NPC, associated with late stage and lymph node tumor load, ctDNA can guide treatment escalation and de-escalation, minimally invasive personalized monitoring tool |
With those inputs in place, the Builder then generated full abstract drafts. Below is how the outputs compare: the original published version, the draft built from unpolished shorthand, and the draft built from polished input.
Table 2: Outputs (Original, Unpolished, Polished)
Original Abstract | Output (Unpolished Input) | Output (Polished Input) |
Title ctDNA-guided treatment stratification for nasopharyngeal carcinoma patients: A retrospective cohort. | Title Circulating Tumor DNA as a Biomarker in Nasopharyngeal Carcinoma: Implications for Treatment and Prognosis | Title Circulating Tumor DNA as a Biomarker in Nasopharyngeal Carcinoma |
Background Liquid biopsy based on circulating tumor DNA (ctDNA) has recently been used to treat various cancers. Studies on the use of ctDNA in nasopharyngeal carcinoma (NPC) are not available. | Background Nasopharyngeal carcinoma (NPC) presents diagnostic and therapeutic challenges, particularly in evaluating treatment response and prognosis. Circulating tumor DNA (ctDNA) has emerged as a potential biomarker for systemic treatment monitoring in NPC. This study investigates the relationship between ctDNA levels, treatment outcomes, and disease characteristics. | Background Nasopharyngeal carcinoma (NPC) presents challenges in monitoring and treatment stratification. This study evaluates circulating tumor DNA (ctDNA) as a biomarker for systemic treatment efficacy in NPC. |
Methods Our retrospective analysis encompassed 45 patients diagnosed with nasopharyngeal carcinoma, all of whom underwent ctDNA analysis and received comprehensive systemic treatment. Logistic regression analysis was employed to evaluate the impact of relevant factors on the objective response rate (ORR). | Methods In this retrospective analysis, we evaluated 45 patients with NPC, all of whom underwent ctDNA analysis and received systemic treatment. Logistic regression was employed to assess the impact of ctDNA on the objective response rate (ORR). Additional analyses examined the correlation between ctDNA levels and various clinical parameters, including Epstein-Barr virus (EBV) status and lymph node involvement. | Methods This retrospective study included 45 patients with NPC who underwent comprehensive systemic treatment. ctDNA analysis was performed to assess its role in treatment response. Logistic regression was used to evaluate factors impacting the objective response rate (ORR). |
Results The multivariate Logistic analysis showed that the increase of ctDNA (OR=1.051, 95%CI 1.001–1.104, p=0.046) increased the risk of non-ORR after systemic treatment. We also demonstrated that patients with persistently ctDNA-positive had poor outcomes and that patient outcomes were associated with ctDNA clearance. ctDNA can also assist plasma Epstein-Barr virus (EBV) DNA in screening for NPC. Accordingly, we explore the longitudinal relationship between ctDNA levels and patient characteristics. The results showed that ctDNA levels were significantly greater in patients with tumor stage N3 than in patients with stage N0–N2. ctDNA levels were significantly greater in patients with positive-EBV DNA than in patients with negative. And the GTVnd (lymph node load) is associated with ctDNA levels. The comparison of ctDNA and tDNA revealed good concordance of gene mutation levels, especially for genes with a high frequency of mutations in NPC. | Results An increase in ctDNA levels was associated with a higher risk of non-response to treatment (OR=1.051, 95% CI 1.001–1.104, p=0.046). Persistent ctDNA positivity was linked to poorer outcomes, whereas ctDNA clearance correlated with improved prognosis. ctDNA, combined with EBV DNA, proved effective in NPC screening. Higher ctDNA levels were observed in patients with N3 stage compared to N0–N2 and in EBV-positive versus EBV-negative cases. A significant correlation was found between ctDNA levels and lymph node gross tumor volume (GTVnd). Concordance was noted between ctDNA and tissue DNA mutations, especially in genes commonly altered in NPC. | Results An increase in ctDNA levels was significantly associated with a higher risk of non-ORR (OR=1.051, 95% CI 1.001–1.104, p=0.046). Patients with persistent ctDNA positivity experienced poorer outcomes, whereas ctDNA clearance correlated with improved outcomes. ctDNA proved effective in conjunction with EBV DNA for NPC screening. Higher ctDNA levels were observed in N3 stage compared to N0–N2 stage patients and in EBV-positive versus EBV-negative patients. A correlation was found between ctDNA levels and lymph node load (GTVnd). The mutation profiles of ctDNA and tumor DNA (tDNA) showed strong concordance, especially in frequently mutated NPC genes. |
Conclusions In conclusion, ctDNA is an effective biomarker for comprehensive systemic treatment for NPC, which is associated with late clinical stage, late N stage and high lymph node tumor load. ctDNA can guide the treatment of escalation and de-escalation in NPC. This approach will lead us to a minimally invasive and personalized era. | Conclusion ctDNA serves as a valuable biomarker for assessing systemic treatment response in NPC, reflecting disease stage and lymph node burden. Its minimally invasive nature allows for personalized treatment strategies, potentially guiding therapeutic escalation or de-escalation. | Conclusion ctDNA serves as a valuable biomarker for systemic treatment in NPC, reflecting disease stage and lymph node tumor load. It offers a minimally invasive method for personalized monitoring, potentially guiding treatment escalation and de-escalation. |
Results and Limitations
Taken together, the side-by-side shows how the Builder handles real-world input: even messy shorthand is enough to produce a structured, conference-ready draft, while polished input tightens style but ironically risks losing some detail.
📌 What changed?
Unpolished input → Captured the full scientific content, including subgroup details (EBV status, N stage, lymph node burden). Reads longer, but the science is intact.
Polished input → Accurate on the core findings, but trimmed away some exploratory detail and novelty.
For context: the original abstract was scientifically complete but dense and harder to parse.
Big win? Both Builder outputs preserved the core science. The difference is mostly style and granularity, not accuracy — proving that even shorthand notes are enough to generate a structured, scientifically accurate draft.
Beyond content, the Abstract Builder reliably cut down the time it takes to get to clean, structured drafts. Instead of staring at a blank Word doc, you can generate a coherent draft in minutes with the right structure and all core data intact. The revision loop worked smoothly too: every iteration stayed linked in Airtable and the shared Google Doc, avoiding the dreaded final-final-v3 chaos.
The weaknesses? Backgrounds skew generic unless you feed detail. Conclusions overstate confidence unless edited. And when data is missing, AI guesses. That’s why human scientists still need to hedge, validate, and interpret. AI clears the formatting grind; you still own the science.
Tools That Powered v1
Part of building v1 wasn’t just wiring the workflow — it was pressure-testing the tools themselves. Here’s what worked, what didn’t, and why it matters.
Landbot for intake. Google Forms is fast but rigid; Typeform looks slick but is built for marketing. Landbot felt right for abstracts because it behaves like a guided conversation, nudging users to fill in the exact pieces a workflow needs. That’s the difference between a half-broken submission and a usable draft.
Airtable for records. Sheets and Notion can hold data, but they collapse under version control. Airtable was built for structured workflows: every submission became a clean record with inputs, outputs, and revisions linked together. No more “FINAL_FINAL_v3.docx” spirals.
Zapier as the glue. Zapier isn’t elegant, but it’s flexible, transparent, and easy to debug — exactly what you want in a v1. Make (Integromat) is more powerful for branching, and custom code will eventually win on scale and cost. But to learn fast, Zapier was the right call.
AI layer. Raw ChatGPT with a smart system prompt got me 70–80% of the way. A Custom GPT makes interactive polishing easier, and fine-tuning comes later, once I’ve built a library of “gold” Savvyn outputs to lock tone and structure.
The sandbox stack I used — Landbot, Airtable, Zapier, and ChatGPT — works well for abstracts and other non-PHI use cases like manuscripts and publications. These don’t trigger HIPAA requirements, so basic data security is sufficient. That makes this workflow not just a prototype, but also immediately practical for teams submitting conference abstracts or preparing scientific writing.
At the same time, this stack is only Tier 0. It’s the foundation for testing how much value each additional layer adds — style guides, curated retrieval, custom GPTs, and eventually fine-tuning. Each tier reduces friction and makes outputs more reliable, which is where the real differentiation lies.
The real constraint shows up once you cross into clinical territory. Protocols, patient-facing documents, regulatory submissions — those demand HIPAA compliance, audit trails, and secure hosting. Intake tools like REDCap or Formsort, compliant databases with audit logging, and AI models deployed in secure environments (Azure OpenAI with BAA, or on-prem LLMs) become non-negotiable.
Side note: Can we please get more HIPAA-compliant no-code tools — the Zapier/Make equivalents? If they already exist and I’ve missed them, send them my way 🙏.
What’s Next: Beyond Abstracts
Savvyn Abstract Builder v1 is just a starting point. Abstracts are the lowest-hanging fruit — short, structured, and universal enough to road-test workflows. But the vision is bigger, and the path forward has two tracks:
1. Moving up the tiers
Tier 0 proved that even basic wiring (prompt + workflow) works.
Tier 1–2 will add style guides, glossaries, and curated retrieval so every output sounds Savvyn, not generic GPT.
Tier 3 enables interactive polishing with a Custom GPT.
Tier 4 locks in tone and structure with fine-tuning once enough gold examples exist.
2. Expanding to new document types
The same framework extends well beyond abstracts:
Full manuscripts with sections, citations, and figure legends.
Clinical documents like protocol synopses and consent forms.
Regulatory submissions where strict formatting and audit trails are non-negotiable.
The future state also means moving past rigid Background/Methods/Results boxes. Instead of structured fields, users should be able to dump raw notes — shorthand, bullets, even “vomit drafts” — and let the system do the structuring automatically.
Closing Thought
The era of scientists writing their own abstracts — and most scientific content — is over. Controversial point, perhaps, but I stand by it. Abstracts have always been a peculiar kind of bottleneck: short, formulaic, universally dreaded — the one document every scientist must write, but no one wants to. That’s why they made sense as the test case for v1 of the Savvyn Abstract Builder. The experiment showed that even unpolished shorthand is enough to produce a structured draft, and that the core science comes through intact. Style, formatting, and polish can be layered later.
But the point isn’t just about abstracts. This experiment showed that workflows can offload the repetitive, inefficient parts of scientific writing without touching the substance. And in doing so, it levels the playing field: strong data yields strong drafts, no matter the author. The next steps — scaling up the tiers, adding style guides, retrieval, and fine-tuning — are about making outputs more consistent and reliable. The broader challenge is extending this to manuscripts, clinical documents, and regulatory submissions in ways that are compliant, auditable, and secure. That’s where the real future lies.
👉 Want to see the prototype in action? Reach out and I’ll share a demo version so you can try it yourself. Or follow along here at Savvyn Insights as I keep pushing from Abstract Builder v1 toward full scientific document automation.
Thanks for reading,
—Savvyn (your partner in ruthless efficiency)
🌀 ➝ 📊 ➝ 💡 ➝ 🚀
From Chaos to Clarity. Amplify Your Impact.
Comments