AI projects for early talent
Hands-on projects pair an experienced mentor with early-career talent. Each collaboration is scoped to finish inside its timeframe and produces concrete statistical deliverables, including a visualization dashboard and documented evaluation results.
A catalog of achievable, industry-relevant AI projects that pair a mentor with early-career talent. Every project is scoped to finish inside its timeframe and produces concrete statistical deliverables — an interactive visualization dashboard and documented evaluation results — so students leave with something real to show.
Every project needs a mentor and one or more students. Short-term projects are focused sprints; long-term projects are phased builds. Expand any project to see its full scope and plan.
Clinical Operations
AI for clinical operations and process optimization — document generation, programming automation, study design, and end-to-end workflows.
Regulatory Document Auto-Drafting
The gap. Writing Informed Consent Forms and SAP sections is slow, templated, and repetitive. Teams re-draft the same boilerplate for every study.
What you build. A retrieval-augmented LLM pipeline that drafts a single, well-bounded document type from structured inputs, grounded on prior approved examples, with a human-in-the-loop review step.
Deliverables
- Interactive review dashboard that shows the draft next to its source passages and a section-by-section completeness score
- Evaluation results: rubric scores (completeness, accuracy, tone) vs. gold documents, with agreement statistics
- Reusable drafting pipeline and prompt library
Plan · 4 weeks
- Week 1 — Pick one document type; collect a small corpus of public/sample templates
- Week 2 — Build a retrieval index and a baseline drafting prompt
- Week 3 — Prompt-chain a full draft from structured study inputs
- Week 4 — Score drafts against gold examples and build the review dashboard
Industry impact
- Cuts drafting time and improves consistency for one high-volume document
- Faster study start-up, with the author still accountable for sign-off
Skills you gain
Statistical Programming QC-Diff Assistant
The gap. QC double-programming of CDISC datasets is costly, and errors accumulate across the SDTM to ADaM to TLF flow.
What you build. For a single ADaM dataset (e.g., ADSL), a developer step generates derivation code and an independent QC step re-derives the output and diffs the two, producing an output-consistency score.
Deliverables
- Dashboard visualizing variable-level discrepancies between developer and QC outputs
- Evaluation results: consistency score and a discrepancy report across repeated runs
- Reproducible dev + QC code and a sample specification-to-dataset workflow
Plan · 4 weeks
- Week 1 — Choose one ADaM dataset and spec; stand up sample SDTM input
- Week 2 — Generate derivation code from the spec
- Week 3 — Build an independent QC re-derivation and an automated diff
- Week 4 — Compute the consistency score and build the discrepancy dashboard
Industry impact
- Reduces programming and QC hours on a concrete, bounded task
- Directly targets the top named trust blocker: validating non-deterministic output
Skills you gain
Study-Design Recommender
The gap. Designing a new study often reinvents prior work; it is hard to quickly find analogous trials and learn from their endpoints, arms, and sample sizes.
What you build. A similarity search over a slice of ClinicalTrials.gov: given a draft synopsis, return the most similar prior trials with a summary of their design decisions.
Deliverables
- Interactive dashboard: enter a synopsis, see ranked similar trials and a visual comparison of their design features
- Evaluation results: retrieval relevance (precision@k) on a set of known analogous trials
- Reusable embedding index and search code over the trial corpus
Plan · 4 weeks
- Week 1 — Pull a ClinicalTrials.gov slice; parse key design fields
- Week 2 — Embed protocol text and build cosine-similarity search
- Week 3 — Return top-k similar trials with a design summary for a query synopsis
- Week 4 — Evaluate relevance on known cases and build the lookup dashboard
Industry impact
- Speeds up study design and grounds it in precedent
- A reusable internal tool for protocol and biostatistics teams
Skills you gain
Protocol to Structured Study Definition (USDM)
The gap. Protocols are free text, but downstream automation needs a structured study definition (CDISC USDM). Manual conversion is slow, and this structured backbone is a prerequisite for automating the rest of the pipeline.
What you build. An extraction pipeline that reads a protocol and emits a USDM-aligned JSON of design elements, variables, and values, validated against the schema, with coverage and accuracy metrics.
Deliverables
- Dashboard visualizing extraction coverage and confidence across USDM entities
- Evaluation results: field-level extraction accuracy, coverage, and error-mode breakdown vs. a hand-labeled set
- Validated extraction pipeline producing schema-conformant USDM JSON
Plan · 10–12 weeks
- Weeks 1–2 — Study the CDISC USDM structure and target entities; assemble sample protocols
- Weeks 3–5 — Build information extraction over protocol text (NER + mapping)
- Weeks 6–8 — Map extracted entities to USDM JSON and validate against schema
- Weeks 9–10 — Measure extraction coverage/accuracy and characterize error modes
- Weeks 11–12 — Extend toward Analysis Results Data (ARD) and build the coverage dashboard
Industry impact
- A structured backbone unlocks automation across the whole pipeline
- Foundation for AI-native, machine-readable submissions
Skills you gain
Agentic Orchestration Harness
The gap. Individual AI tools exist, but handoffs between spec, programming, QC, and review are still manual. The future operating model requires agents that collaborate reliably.
What you build. An orchestration harness where specialized agents (spec-reader, programmer, QC, reviewer) collaborate to take a small study spec through to a validated TLF, with logging and human checkpoints.
Deliverables
- Dashboard visualizing the end-to-end run: agent steps, handoffs, checkpoints, and where humans intervened
- Evaluation results: end-to-end success rate, error-recovery rate, and human-intervention counts across runs
- Reusable multi-agent harness with logging and human-in-the-loop checkpoints
Plan · 10–12 weeks
- Weeks 1–2 — Define agent roles and an agent-to-agent communication protocol
- Weeks 3–6 — Build the harness on an orchestration framework
- Weeks 7–9 — Run it on a small end-to-end toy pipeline
- Weeks 10–12 — Add human checkpoints, logging, and the run dashboard with success metrics
Industry impact
- Demonstrates the future operating model of augmented statistics teams
- A flagship, portfolio-defining systems project
Skills you gain
Evidence Generation
Advanced methods and AI for generating credible evidence — modern modeling, causal inference, real-world evidence, and multimodal data.
Baseline-Drift Monitoring Dashboard
The gap. Baseline covariate distributions can drift as enrollment progresses, threatening comparability and interpretation — but drift is rarely monitored in real time.
What you build. A monitoring tool that tracks baseline distribution drift over cumulative simulated enrollment, with statistical drift metrics and alerting.
Deliverables
- Interactive monitoring dashboard showing distribution drift over cumulative enrollment with alert flags
- Evaluation results: detection sensitivity and false-alarm rate across injected drift scenarios
- Reusable drift-metric library and simulation harness
Plan · 4 weeks
- Week 1 — Simulate enrollment with injectable drift scenarios
- Week 2 — Implement drift metrics (standardized differences, PSI, tests)
- Week 3 — Add alerting logic and thresholds
- Week 4 — Build the dashboard and evaluate detection across scenarios
Industry impact
- Early warning on comparability issues during a live trial
- A monitoring pattern that generalizes to safety and data-quality dashboards
Skills you gain
Deep-Learning Survival Models
The gap. Traditional Cox models struggle with high-dimensional and multimodal data, yet teams need better prognostic modeling without sacrificing interpretability.
What you build. Implement and benchmark deep-learning survival models (e.g., DeepSurv, transformer-based) against a Cox baseline on public data, adding interpretability via SHAP or attention.
Deliverables
- Dashboard comparing model risk predictions, survival curves, and feature attributions
- Evaluation results: concordance index and calibration across models, with confidence intervals
- Reproducible training/benchmark code and an interpretability analysis notebook
Plan · 10–12 weeks
- Weeks 1–2 — Reproduce a Cox proportional-hazards baseline on a public dataset
- Weeks 3–6 — Implement a DeepSurv / transformer survival model
- Weeks 7–9 — Compare via concordance index and calibration
- Weeks 10–12 — Add interpretability and build the comparison dashboard
Industry impact
- Better risk prediction for prognosis and patient selection
- Protects the traditional-statistics rigor half of the program's balance
Skills you gain
Prognostic-Score Covariate Adjustment
The gap. Trials need more statistical power without more patients. Prognostic covariate adjustment (PROCOVA-style, digital twins) is EMA-qualified but under-used, and its regulatory limits are poorly understood by newcomers.
What you build. Train a prognostic model on historical control data, use it as a covariate in a simulated RCT analysis, and quantify the power gain while confirming type-I error control.
Deliverables
- Dashboard visualizing power and type-I error across effect sizes and adjustment strategies
- Evaluation results: power gain vs. unadjusted analysis and type-I error simulation tables
- Reproducible simulation study and a short methods write-up on regulatory limits
Plan · 10–12 weeks
- Weeks 1–3 — Train a prognostic model on historical/control data
- Weeks 4–7 — Simulate RCTs and analyze with ANCOVA + prognostic covariate
- Weeks 8–10 — Quantify power gain and run type-I error simulations
- Weeks 11–12 — Summarize value/limits and build the simulation dashboard
Industry impact
- Smaller, faster trials using an EMA/FDA-recognized method
- Bridges machine learning and rigorous trial statistics
Skills you gain
RWD/RWE Patient–Trial Matching & Site Selection
The gap. Patient recruitment is the single largest bottleneck in clinical research. Real-world data can identify eligible patients, rank sites, and check representativeness.
What you build. Using a public claims/EHR-like dataset, build an eligibility matcher against trial criteria and a site-selection ranking based on eligible-patient density, plus representativeness metrics.
Deliverables
- Dashboard mapping eligible-patient density and ranking candidate sites with representativeness views
- Evaluation results: matching precision/recall vs. labeled cases and cohort representativeness metrics
- Reusable eligibility-matching pipeline
Plan · 10–12 weeks
- Weeks 1–3 — Encode a set of trial eligibility criteria; profile the dataset
- Weeks 4–7 — Build structured + semantic matching against the data
- Weeks 8–10 — Rank candidate sites by eligible-patient density
- Weeks 11–12 — Compute representativeness metrics and build the site dashboard
Industry impact
- Faster enrollment and reduced per-trial recruitment cost
- Better, more representative trial populations
Skills you gain
Literature-to-Hypothesis Research Agent
The gap. The literature grows faster than any researcher can read. Identifying gaps, generating hypotheses, and tracking evidence is manual and slow.
What you build. An agent that ingests abstracts (PubMed / Europe PMC), summarizes a field, identifies gaps, proposes candidate hypotheses, and links supporting evidence — judged by a human-quality rubric.
Deliverables
- Dashboard exploring the evidence map: topics, gaps, proposed hypotheses, and supporting citations
- Evaluation results: human-rated quality/novelty scores with inter-rater agreement
- Reusable ingestion + summarization + hypothesis-generation pipeline
Plan · 10–12 weeks
- Weeks 1–2 — Scope a disease/target area; pull abstracts via a public API
- Weeks 3–5 — Build retrieval + summarization and gap identification
- Weeks 6–8 — Generate hypotheses and link supporting evidence
- Weeks 9–12 — Define a quality rubric, run human evaluation, and build the evidence-map dashboard
Industry impact
- Compresses background research and surfaces non-obvious connections
- A springboard for target identification and mechanism discovery
Skills you gain
Generative Models for Trial Design & Patient Selection
The gap. Exploring trial-design and patient-selection choices is slow and often limited to a handful of scenarios, so promising options go unexplored.
What you build. Use generative/diffusion models to synthesize realistic virtual patients and explore how design and selection choices affect operating characteristics, validated against real distributions.
Deliverables
- Dashboard comparing synthetic vs. real covariate distributions and design trade-offs
- Evaluation results: fidelity metrics (e.g., distributional distance) and privacy/utility checks
- Reproducible generative-model code and a virtual-patient dataset generator
Plan · 10–12 weeks
- Weeks 1–3 — Assemble a public dataset and profile real distributions
- Weeks 4–7 — Train a generative/diffusion model for virtual patients
- Weeks 8–10 — Validate fidelity and run design/selection scenarios
- Weeks 11–12 — Build the comparison dashboard and document limits
Industry impact
- Explore many more design options before committing resources
- Enables privacy-conscious data sharing and prototyping
Skills you gain
Multimodal Data Fusion Prototype
The gap. Digital pathology, wearables, and labs each carry signal, but they are usually analyzed in isolation, leaving predictive value on the table.
What you build. A prototype that fuses two or more public modalities (e.g., imaging + tabular labs) into a single predictive model, with an ablation showing each modality's contribution.
Deliverables
- Dashboard visualizing per-modality contributions and fused predictions
- Evaluation results: single-modality vs. fused performance with an ablation table
- Reproducible multimodal training pipeline
Plan · 10–12 weeks
- Weeks 1–3 — Select public modalities and build per-modality baselines
- Weeks 4–7 — Implement a fusion architecture
- Weeks 8–10 — Run ablations quantifying each modality's contribution
- Weeks 11–12 — Build the contribution dashboard and write up findings
Industry impact
- Higher predictive value from data already being collected
- A template for combining emerging data sources
Skills you gain
Go/No-Go Phase 2 to Phase 3 Prediction
The gap. Progression decisions from Phase 2 to Phase 3 are high-stakes and often rely on static, point-in-time reads that miss signal in richer data.
What you build. A progression-scoring model that predicts Phase 3 success from Phase 2 data on a public/simulated dataset, with calibrated probabilities and clear decision thresholds.
Deliverables
- Dashboard turning model scores into calibrated go/no-go probabilities with threshold controls
- Evaluation results: discrimination (AUC), calibration, and decision-curve analysis
- Reproducible modeling pipeline and a decision-rule write-up
Plan · 10–12 weeks
- Weeks 1–3 — Assemble a public/simulated Phase 2 to Phase 3 dataset
- Weeks 4–7 — Build and tune the progression-scoring model
- Weeks 8–10 — Calibrate probabilities and evaluate decision rules
- Weeks 11–12 — Build the decision dashboard and document thresholds
Industry impact
- More consistent, evidence-based progression decisions
- Applicable across oncology and non-oncology indications
Skills you gain
External Control Arm Feasibility Explorer
The gap. External control arms can accelerate trials, but their validity depends on comparability and bias control that teams struggle to assess up front.
What you build. A workflow that constructs an external control arm from public data, balances it against a treated cohort, and quantifies residual bias and sensitivity.
Deliverables
- Dashboard visualizing covariate balance before/after weighting and sensitivity to unmeasured confounding
- Evaluation results: balance diagnostics, estimated effect with confidence intervals, and bias sensitivity analysis
- Reproducible external-control construction and analysis code
Plan · 10–12 weeks
- Weeks 1–3 — Define the target estimand; assemble treated and external cohorts
- Weeks 4–7 — Balance cohorts (matching/weighting) and check overlap
- Weeks 8–10 — Estimate effects and run sensitivity/bias analyses
- Weeks 11–12 — Build the feasibility dashboard and summarize regulatory limits
Industry impact
- Faster trials where a concurrent control is hard to enroll
- Clear-eyed assessment of when external controls are credible
Skills you gain
Trust & Validation
Making AI outputs trustworthy — reproducibility, benchmarking, human-in-the-loop sign-off, and quality control across the data flow.
Trust Benchmark for Non-Deterministic AI
The gap. The same prompt can yield different answers, so AI cannot be trusted or validated for GxP work. This is the single most-named blocker across the advisory interviews.
What you build. A benchmark for one concrete clinical-stats task, with an evaluation matrix, repeated-run reproducibility measurement, and majority-vote scoring.
Deliverables
- Dashboard visualizing accuracy, rationale quality, and run-to-run reproducibility across models
- Evaluation results: an evaluation matrix and majority-vote scoring with reproducibility statistics
- Reusable benchmark harness and gold-standard test set
Plan · 4 weeks
- Week 1 — Define one bounded task and build a small gold-standard set
- Week 2 — Build the evaluation harness and metrics (accuracy / rationale / latency)
- Week 3 — Run multiple models and seeds; log all outputs
- Week 4 — Build the reproducibility + accuracy dashboard with majority-vote scoring
Industry impact
- Provides the evidence base needed for validation and regulatory acceptance
- Turns 'we can't trust it' into a measurable, defensible metric
Skills you gain
Human-in-the-Loop Sign-off Workflow
The gap. When AI drafts analyses or documents, accountability blurs. Teams need a clear review-and-sign-off flow that records who approved what and why.
What you build. A lightweight review workflow that routes AI outputs to a human reviewer, captures decisions and rationale, and produces an auditable sign-off trail.
Deliverables
- Dashboard showing review queue status, approval rates, and turnaround time
- Evaluation results: reviewer agreement and time-to-sign-off metrics on a sample set
- Reusable review workflow with an auditable decision log
Plan · 4 weeks
- Week 1 — Define the review states, roles, and audit fields
- Week 2 — Build the routing and decision-capture flow
- Week 3 — Add the auditable sign-off trail and export
- Week 4 — Build the status dashboard and evaluate on sample outputs
Industry impact
- Keeps accountability with humans as AI adoption grows
- A reusable pattern for GxP-aligned AI use
Skills you gain
Error-Propagation Tracker
The gap. Small errors introduced early (raw to SDTM to ADaM to TLF) can silently propagate and amplify, and teams lack visibility into where they originate.
What you build. A demo pipeline that injects controlled errors upstream and traces how they propagate to downstream outputs, quantifying amplification at each stage.
Deliverables
- Dashboard visualizing error propagation from source to final outputs across the flow
- Evaluation results: stage-by-stage error amplification metrics under injected-error scenarios
- Reproducible mini-pipeline with an error-injection harness
Plan · 4 weeks
- Week 1 — Build a small raw to SDTM to ADaM to TLF demo pipeline
- Week 2 — Add a controlled error-injection harness
- Week 3 — Trace and quantify propagation at each stage
- Week 4 — Build the propagation dashboard and summarize findings
Industry impact
- Shows where QC effort has the highest leverage
- Makes the case for checkpoints at the right stages
Skills you gain
Regulatory Science
Regulatory science in the AI era — applying FDA/EMA/PMDA guidance, submission quality, and AI-native regulatory workflows.
FDA 'Good AI Practice' Credibility Assessor
The gap. The FDA's 2025/2026 AI guidances define a risk-based credibility framework, but teams lack practical tooling to apply it to a specific use-case.
What you build. An interactive credibility-assessment tool that walks a use-case through the framework, assigns a risk tier, and generates a documentation artifact, with retrieval-backed citations.
Deliverables
- Interactive dashboard that scores a use-case and visualizes its risk tier across framework dimensions
- Evaluation results: consistency of tiering across sample use-cases and reviewers
- Generated credibility-assessment report template with guidance citations
Plan · 4 weeks
- Week 1 — Digest the guidances into structured principles and questions
- Week 2 — Build a questionnaire that maps answers to a model-risk tier
- Week 3 — Generate a credibility-assessment report template
- Week 4 — Add retrieval over the guidance docs and build the scoring dashboard
Industry impact
- Accelerates compliant AI adoption inside statistics/regulatory teams
- Produces artifacts teams can attach to submissions
Skills you gain
Submission & Define-Document Quality Checker
The gap. Define documents and derivations are hard to review for completeness and provenance, and quality issues surface late in the submission process.
What you build. A checker that reviews a define.xml-style package for missing derivations, provenance gaps, and inconsistencies, and reports quality issues with severities.
Deliverables
- Dashboard visualizing quality issues by type and severity across the package
- Evaluation results: precision/recall of issue detection against a seeded set of known problems
- Reusable quality-check rules and report generator
Plan · 4 weeks
- Week 1 — Model a sample define package and catalog quality checks
- Week 2 — Implement provenance/completeness/consistency checks
- Week 3 — Seed known issues and measure detection
- Week 4 — Build the quality dashboard and issue report
Industry impact
- Catches submission issues early, reducing review cycles
- Improves derivation and variable provenance quality
Skills you gain
Global Regulatory Alignment Tracker
The gap. FDA, EMA, and PMDA positions on AI evolve independently, and teams lack a single, structured view of where they align and diverge.
What you build. A structured, comparable knowledge base of FDA/EMA/PMDA AI guidance with a side-by-side view of alignment and gaps across key topics.
Deliverables
- Interactive dashboard comparing agency positions topic-by-topic with alignment indicators
- Evaluation results: coverage and reviewer-checked accuracy of the structured entries
- Structured, source-linked dataset that can be maintained over time
Plan · 4 weeks
- Week 1 — Define comparison topics and a structured schema
- Week 2 — Populate FDA/EMA/PMDA entries from public sources
- Week 3 — Build the side-by-side alignment view
- Week 4 — Review accuracy and finalize the comparison dashboard
Industry impact
- A single reference for multi-region AI strategy
- Reduces surprises when aligning global submissions
Skills you gain
AI-Native Submission Prototype
The gap. Submissions are document-heavy PDFs; derivations and variable provenance are hard to trace, and review is slow.
What you build. A prototype that represents analysis results and provenance as ARD-aligned YAML/JSON, with a retrieval agent that answers reviewer questions using traceable citations.
Deliverables
- Dashboard for reviewers to query results and see traceable provenance for each answer
- Evaluation results: answer accuracy and citation-traceability rate on a reviewer question set
- Structured, machine-readable submission package with a provenance graph
Plan · 10–12 weeks
- Weeks 1–3 — Model ARD-aligned outputs and a provenance graph for a toy analysis
- Weeks 4–7 — Build the structured submission package with derivations linked
- Weeks 8–10 — Add a retrieval agent over the package
- Weeks 11–12 — Build the reviewer Q&A dashboard and evaluate traceability
Industry impact
- Advances the review-time compression aspiration
- Audit-ready, machine-readable submission content
Skills you gain
Real-Time Trial Safety Monitoring
The gap. Safety signals can emerge mid-trial, but monitoring is often periodic and manual, delaying detection of important events.
What you build. A safety-first monitoring prototype that continuously scans simulated accumulating trial data for signals and surfaces them with statistical support and clear caveats.
Deliverables
- Real-time safety dashboard with signal alerts, trends, and drill-down by event
- Evaluation results: detection timeliness and false-signal rate across simulated scenarios
- Reusable signal-detection logic and simulation harness
Plan · 10–12 weeks
- Weeks 1–3 — Simulate accumulating trial safety data with planted signals
- Weeks 4–7 — Implement continuous signal-detection methods
- Weeks 8–10 — Tune sensitivity and control false alarms
- Weeks 11–12 — Build the safety dashboard and evaluate timeliness
Industry impact
- Earlier detection of safety signals with statistical rigor
- A safety-first template that respects regulatory caution
Skills you gain
Workforce & Skills
The augmented workforce — evolving roles, blended stats/ML/domain skills, and structured upskilling paths.
AI-Skills Competency & Upskilling-Path Tool
The gap. Statisticians need to know what to learn for the next 3–10 years as roles converge across stats, ML, and domain expertise — but there is no structured, personalized path.
What you build. A competency framework plus a self-assessment that identifies gaps and recommends a tailored upskilling path with concrete resources.
Deliverables
- Interactive dashboard visualizing a person's competency profile and recommended path
- Evaluation results: face-validity feedback from sample personas and reviewers
- Reusable competency framework and self-assessment instrument
Plan · 4 weeks
- Week 1 — Define the competency framework from advisory themes and literature
- Week 2 — Build the self-assessment questionnaire
- Week 3 — Implement gap-to-resource recommendation logic
- Week 4 — Build the profile dashboard and test with sample personas
Industry impact
- Directly serves the PharmaDS talent program and its mentees
- A shareable resource for career development across the community
Skills you gain
Data Manager to Data Engineer Transition Kit
The gap. As data management shifts toward data engineering, practitioners lack a concrete, hands-on bridge from their current skills to the new ones.
What you build. A guided transition kit: a skills map plus a small, realistic hands-on project that takes a learner from data-management tasks to a modern data-engineering workflow.
Deliverables
- Dashboard tracking progress through the transition project and skills acquired
- Evaluation results: pre/post skills-check outcomes from a small pilot
- Reusable curriculum, skills map, and hands-on project template
Plan · 4 weeks
- Week 1 — Map data-management to data-engineering skills and gaps
- Week 2 — Design a small hands-on data-engineering project
- Week 3 — Build guided materials and a pre/post skills check
- Week 4 — Pilot with a learner and build the progress dashboard
Industry impact
- A concrete path for a workforce transition already underway
- Reduces ramp time into data-engineering roles
Skills you gain
Strategy & Adoption
Adoption, strategy, and business context — scaling from pilot to enterprise, vendor landscape, and serving small and mid-size organizations.
AI-in-Pharma-DS Vendor Landscape
The gap. The vendor and startup landscape is moving fast, and it is hard to track capabilities, gaps, and partnership opportunities.
What you build. A categorized landscape of AI tools/vendors for pharma DS with a structured comparison and a short opportunity analysis.
Deliverables
- Interactive dashboard to filter and compare vendors by function and capability
- Evaluation results: coverage of the space and a reviewer-checked gap analysis
- Structured, source-linked landscape dataset the community can maintain
Plan · 4 weeks
- Week 1 — Define a taxonomy by function (doc-gen, programming, RWE, regulatory, ...)
- Week 2 — Collect and structure vendor data from public sources
- Week 3 — Build a comparison and filtering view
- Week 4 — Analyze gaps/opportunities and finalize the dashboard
Industry impact
- Informs build-vs-buy and partnership decisions
- A living reference the community can maintain
Skills you gain
Minimal AI Environment for Small/Mid Pharma
The gap. Small biotechs and CROs cannot afford large AI infrastructure, yet still need a reproducible, compliant environment to run AI-assisted analyses.
What you build. A reference open-source AI environment — open LLM + retrieval + evaluation + GxP-style audit logging — packaged as a deployable template with a cost/benefit analysis and adoption playbook.
Deliverables
- Dashboard summarizing cost, performance, and usage of the reference stack
- Evaluation results: cost/performance benchmarks vs. a hosted baseline
- Deployable open-source template and an adoption playbook
Plan · 10–12 weeks
- Weeks 1–3 — Assemble an open-source stack (open LLM, vector store, eval)
- Weeks 4–7 — Add audit logging and versioning for GxP-style traceability
- Weeks 8–10 — Benchmark cost and performance vs. a hosted baseline
- Weeks 11–12 — Write the adoption playbook and build the cost dashboard
Industry impact
- Democratizes AI for organizations without big budgets
- Enables the AI + CRO model to serve small biotech
Skills you gain
Want to take on an AI project?
Join the PharmaDS mentorship program and tell us which project or track best fits your goals.