Before
Reviewers had to manually stitch together fast-moving oncology evidence from scattered sources, including official guidelines, with keyword search offering limited help and little time to inspect the reasoning behind a summary.
Signal to system - case studies
The business value comes first.
The solution comes second.
These case studies show how I approach AI and machine-learning product discovery: not as a search for impressive technology, but as a disciplined way to improve decisions, workflows, efficiency and growth.
01 Discovery and organization's world model
02 Map owners and recommend initiatives
03 Beat the baseline in MVP
04 Ship, scale or kill.
Case index
Four ways AI becomes useful only after the problem is understood.
Healthcare evidence navigation, operations workflow, synthetic customer research and traffic forecasting. Different domains, same discipline: human accountability, baselines, provenance, observability and explicit stop conditions.
A deployed MIT AI/ML capstone for high-stakes oncology evidence navigation. The system supported medical reviewers with source-linked answers, uncertainty, reviewer workflow and explicit quality boundaries.
Controlled evidence sources
Semantic retrieval
Deterministic guardrails
LLM-assisted synthesis
Reviewer validation
A hybrid system: deterministic engine, semantic retrieval/RAG and LLM assistance.
Problem
NSCLC research grows faster than clinical review capacity, while new findings may take years to appear in formal guidelines. The problem was the cost and risk of manually synthesising patient-relevant evidence, not the absence of a chatbot.
Why it mattered
Confidence did not come from a polished AI answer. Reviewers needed to inspect source evidence, reasoning path, uncertainty and system boundaries before trusting the output.
MVP / pilot
A bounded evidence navigator where reviewers could ask focused questions and receive source evidence, rationale, uncertainty signals, provenance traces and workflow support.
Evaluation
- Retrieval correctness and citation faithfulness.
- Answer faithfulness and time-to-evidence.
- Reviewer agreement and error analysis.
Baseline
Manual reviewer research and simpler keyword/filter search. The AI layer had to improve real review work, not merely sound more fluent.
Provenance + kill criteria
Every answer should stay tied to source document, evidence fragment, retrieval logic, model/prompt configuration and review trail. Stop or redesign if citations are not verifiable, hallucinations appear, reviewers do not trust it, or governance/security requirements are not met.
After
Reviewers can work from source-linked and guidelines mapped evidence, visible uncertainty and a traceable review trail. The potential gain is faster evidence navigation without handing clinical judgment to the system.
A workflow discovery for an events and marketing organisation facing fragmented manual operations, unclear ownership and outsourced processing cost of approximately 30% of contract/order value.
Intake
Vendor data
Approved templates
Signature trail
Payout readiness
Risky full LLM contract generation was rejected. The safer direction was a deterministic lifecycle with human review.
Problem
The organisation could not reliably handle contract-to-payout work in-house because the process was fragmented across email, spreadsheets, Drive, task management, accounting and e-signature tools, creating an outsourced processing cost of approximately 30% of contract/order value.
Why it mattered
Status disappeared, handoffs did not scale, operational control was limited and contractors received an inconsistent experience. The problem was not one missing automation; it was the absence of a trusted workflow.
MVP / pilot
A dashboard and lifecycle system for intake, contractor data, document generation from approved templates, verification, signature flow, Drive references, accounting handoff and event history.
Evaluation
- Can the process run end-to-end in-house?
- Are ownership, status and handoff history visible?
- Does the workflow reduce manual duplication without increasing legal risk?
Baseline
The current outsourced process, with approximately 30% added cost, limited operational visibility and no reliable internal lifecycle.
Provenance + kill criteria
The source of each contract should be an approved legal template, with retained statuses, events, document references and decision history. Stop if outsourcing is still required, legal risk rises, status visibility does not improve, or cost does not improve against the baseline.
Before
Contract-to-payout work was fragmented across email, spreadsheets and disconnected tools. Ownership and status could disappear between handoffs, while outsourcing added approximately 30% to processing cost.
After
The organisation can now run - fully internally and without outsourcing costs - one visible lifecycle from approved templates through signature and payout readiness, with people reviewing the consequential steps. The potential gain is operational control and lower processing cost.
A confidential discovery case for AI-assisted concept screening before expensive panel research. The goal was preselection, not pretending synthetic respondents can replace real people.
Concept inputs
Audience segments
Synthetic responses
Semantic Similarity Rating
Panel candidates
Textual elicitation plus Semantic Similarity Rating preserved qualitative reasoning while creating a comparable signal.
Problem
The client needed a faster, cheaper way to screen early concepts, messages or offers before committing weak candidates to expensive panel research.
Why it mattered
The pain was the cost and time required to reach useful insight. The AI layer had to reduce waste before panel research, not become the main decision source.
MVP / pilot
A local pilot where teams input concepts, choose audience segments, generate synthetic qualitative responses, inspect score distributions and rationale, then select the best candidates for real validation.
Evaluation
- Do top synthetic candidates later perform better in real panels?
- Are results stable across runs?
- Does the workflow reduce weak concepts entering paid research?
Baseline
Traditional panel research: slower, more expensive and still the human benchmark. The synthetic layer sits before the panel as a preselection layer.
Provenance + kill criteria
Each result should retain prompt version, model/provider, audience segment, SSR anchors, raw responses, scoring, cost and latency. Stop if outputs do not correlate with panel outcomes, become unstable, or create false confidence.
Before
Early concepts either reached expensive panel research too soon or were filtered through intuition alone. Weak candidates could consume budget before a real signal was available.
After
Teams can use a calibrated synthetic screen to remove weaker options and take only the strongest candidates to human-panel validation. The potential gain is less wasted research spend, not a replacement for real respondents.
An in-progress recommendation for a traffic-forecasting decision-support MVP that helps managers plan staffing from demand signals, while keeping approval and operational responsibility with people.
Historical traffic
Weather + local events
Expected demand
Staffing recommendation
Manager approval
The MVP should not proceed unless retrospective validation shows it can beat a simple baseline.
Problem
The organisation could not reliably forecast customer traffic across venues, so staffing decisions were reactive. Some shifts risked overstaffing; others risked understaffing during peaks.
Why it mattered
Operations, reception, shift managers, HR/scheduling, leadership and customers all felt the consequences. The missing piece was a decision model, not simply "more AI".
MVP / pilot
A four-week decision-support MVP: import historical sales/traffic data, add weather and local event context, forecast expected demand, recommend staffing levels, show rationale and allow manager correction.
Evaluation
- Retrospective forecast error on a closed month.
- Understaffed and overstaffed shift count.
- Scheduling decision time and manager feedback.
Baseline
Manager experience, calendars, reservations, manual reports, day-of-week averages, seasonality and known events.
Provenance + kill criteria
Each forecast should show the historical period, data sources, weather/event inputs, model version and manager corrections. Stop if the model does not beat simple baselines, data is too incomplete, or managers do not find recommendations useful.
Before
Managers planned staffing reactively from experience, calendars and incomplete demand signals. The result could be costly overstaffing on quiet shifts and understaffing - causing reputational damage - when demand peaked.
After
Managers can review an evidence-based demand forecast, see its rationale and adjust the recommendation before scheduling. The potential gain is better coverage with approval and operational responsibility still held by people.
Operating pattern
The repeatable method is driven by value, not AI hype.
AI is useful when it improves a real decision, workflow or reliability problem and can be evaluated against the current way of working.
Rule 1
Start with the real decision. Identify who feels the pain and who owns the call. Map the workflow, handoffs and sources of truth.
Rule 2
Consider AI and non-AI solution classes. Choose the simplest credible system. Design the MVP with human accountability.
Rule 3
Compare against a baseline. Build in provenance and observability. Define kill criteria before scaling.