Clinical De-Risking Engine

Inspiration

Biotech risk often changes before the story is obvious from headlines. Important signals are scattered across messy public records: trial status updates, sponsor changes, milestone timing, phase progression, and negative events like terminated or withdrawn studies. I wanted to see whether an agentic workflow could turn that fragmented information into something more structured and actionable.

That led to Clinical Milestone De-Risking Engine: a workflow designed to answer one clear question:

Which small-to-mid-cap biotech companies appear to be becoming more or less de-risked based on public clinical-trial signals?

Rather than trying to predict approvals or generate investment advice, the goal was to build a transparent system that surfaces clinical momentum and risk changes from public evidence.

What it does

The project takes a curated universe of biotech companies (I focused on the Oncology segment but it can be applied to any segment) and uses public ClinicalTrials.gov data to:

• collect trial records by company and sponsor
• normalize messy sponsor, asset, phase, status, and date information
• distinguish direct company-led studies from partner-linked studies
• generate scoreable milestone signals from trial activity
• rank companies by a transparent de-risking score
• produce leaderboard views, company narratives, and presentation-ready outputs

The workflow is designed to be explainable. Every score is tied back to visible signals such as active recruiting status, later-stage programs, pipeline breadth, recent updates, catalyst timing, or negative trial outcomes.

How I built it

I built the project as an end-to-end workflow in Zerve, using a modular block structure so each stage of the logic was visible and debuggable.

The main workflow steps are:

1. Define a curated oncology biotech universe
2. Pull public trial data from ClinicalTrials.gov
3. Normalize raw trial records into a flat, structured table
4. Tag study ownership and distinguish direct vs partner-linked programs
5. Generate milestone signals from status, phase, recency, and timing
6. Filter low-confidence matches
7. Score trials and companies using transparent heuristics
8. Generate narratives and presentation outputs
9. Present the final leaderboard and featured-company view

I also used iterative QA blocks to validate ownership tagging, false positives, penalty logic, and presentation consistency before freezing the final flow.

Challenges I ran into

The hardest problem was not data access. It was entity resolution.

Clinical-trial records are messy. A company can appear under multiple sponsor variants, some studies are partner-led, some assets appear in combinations, and not every relevant field is cleanly populated. It was easy to create a system that looked analytical but was actually amplifying noise.

A lot of the work went into making the workflow honest:

• preserving provenance
• tagging direct vs partner-linked trials
• dropping low-confidence matches
• avoiding raw trial-count bias
• making the score interpretable instead of overfit

Another challenge was deciding scope. A broad biotech platform would have been more ambitious, but much harder to finish cleanly. I intentionally narrowed the MVP to oncology so the signal quality and storytelling would stay strong.

Accomplishments I'm proud of

I'm proud that the project became a true end-to-end workflow rather than just a static analysis. It starts with messy public trial data and ends with a ranked, explainable output that a user can actually follow.

I'm also proud of:

• building a full Zerve workflow that runs from raw data to final presentation
• creating a transparent scoring system instead of a black-box ranking
• preserving provenance so the workflow can distinguish direct company-led studies from partner-linked ones
• producing a presentation layer with a leaderboard, narratives, and featured-company view
• keeping the MVP narrow enough to finish cleanly while still making the project feel ambitious

What I learned

The biggest lesson was that good agentic workflows are not just about automation, but about traceability. It is not enough to say “AI ranked these companies.” The workflow has to show how raw public records become normalized evidence, how evidence becomes signals, and how signals become a final conclusion.

I also learned that a narrower, better-explained workflow is often stronger than a broader but noisier one. Focusing on one sharp question made the project much easier to build, validate, and present.

What's next for Clinical De-Risking Engine

This MVP focuses on oncology to keep the signal quality high and the workflow explainable, but the structure can be extended further.

Possible next steps include:

• expanding beyond oncology into other biotech segments
• adding more data sources such as regulatory or company disclosure signals
• turning the workflow into a lightweight app or API
• improving user inputs so someone can choose a company, asset, or therapeutic segment interactively

The core idea will stay the same: turn scattered public clinical records into a transparent, explainable ranking of de-risking momentum.

Built With

• api
• clinicaltrials.gov
• data
• pandas
• python
• visualization
• zerve