MIT Cancer Navigator
A transparent evidence engine for NSCLC treatment navigation, designed to show exactly where deterministic safety leads and where hybrid AI can genuinely help.
The Problem
Current LLM-heavy clinical tools often look impressive right up until you ask what exactly drove the answer. In highly structured oncological settings, hallucinations, non-deterministic outputs, and missing provenance are not cute product quirks. They are trust failures.
Our Approach
Deterministic scoring
Evidence is scored against strict parameters like methodology, robustness, recency, and credibility, so ranking remains inspectable instead of vibe-based.
Guideline mapping
Retrieved evidence is mapped against curated ESMO topics, making alignment, conflict, and silence visible as explicit labels rather than hidden prompt behavior.
Maximum explainability
The interface shows what was retrieved, why it matched, and where uncertainty remains, so the demo can be audited instead of merely admired.
How We Use AI
Deterministic gives us safety and auditability. Hybrid gives us lift, but it does not get to cosplay as magic.
That is the whole point of this demo posture. We want MIT to see where the rules engine is intentionally strict, where hybrid retrieval materially improves recall, and where the limits still are. If the semantic path helps, we show it. If it does not move the decision layer enough, we show that too.
This is the safety rail. Deterministic logic owns cohort matching, ERS ranking, guideline mapping, label assignment, and visible uncertainty handling.
This is the lift engine. It uses embeddings, dense plus sparse retrieval, and hybrid search to find better candidates, then hands them back to deterministic guardrails.
Deterministic
Built for consistency, traceability, and explainable failure. It is deliberately strict because the safer failure mode in this domain is visible limitation, not smooth improvisation.
Hybrid + embeddings
Built to improve recall when structured metadata is sparse, patient-fit signals are thin, or topic matching needs semantic help. It is valuable exactly because it is bounded.
Normalize the case first
We turn the patient profile into explicit fields like disease setting, histology, therapy line, and biomarker buckets before any retrieval starts.
Run deterministic cohort matching
The rules engine screens out studies that do not fit the actual patient cohort, which keeps obviously off-target papers out of the main path.
Layer in hybrid retrieval where it helps
A second engine can use dense embeddings, sparse vectors, and hybrid search to recover relevant evidence when structured metadata is incomplete or too generic.
Fuse retrieval back into guardrails
Hybrid candidates are merged back into deterministic ranking and guideline logic, so semantic lift is constrained by explicit safety rails instead of replacing them.
Escalate uncertainty instead of bluffing
If a study is ambiguous, weakly typed, or still not patient-fit, it stays visible as uncertainty or manual review instead of being turned into fake confidence.
Deterministic engine
Typed patient schema, deterministic cohort matching, ERS ranking, and guideline labeling remain the source of truth for the clinical-facing result.
Hybrid retrieval engine
A separate hybrid path can widen candidate recall and semantic topic matching, but it still feeds back into deterministic decision logic instead of bypassing it.
Qdrant Cloud
Our external vector backend stores the hybrid retrieval index for chunk-level search across PubMed evidence and ESMO guideline projections.
OpenRouter + text-embedding-3-small
The live semantic path uses OpenRouter to generate OpenAI `text-embedding-3-small` vectors. That gives us real embeddings without pretending embeddings are the whole product.
Hybrid rank fusion
Dense and sparse results are combined through rank-fusion logic rather than raw-score magic, which makes the merge safer and easier to reason about.
Supabase Postgres
Imported evidence, guideline topics, semantic metadata, and operational state live in Postgres so the demo runs on tracked runtime data rather than invisible temp files.
Codex + Next.js + FastAPI on Vercel
AI Engeneered with Codex, The interface is a Next.js app and the backend is FastAPI, both deployed on Vercel so the web demo and API stay aligned as one inspectable system.
Gemini Assistive LLM hooks
We keep optional import-assist and grounded explainability hooks as explicit toggles, but they are not the decision authority and should never be confused with the core engine.
We are not presenting a magical all-knowing oncology copilot. We are presenting a transparent system with two engines: one that is deliberately deterministic because trust matters, and one that is deliberately hybrid because retrieval breadth matters. When hybrid improves the result, great. When it only improves candidate recall and not the final label, we say that out loud too.
This is the honest build order. We did not begin with a magic copilot. We began with deterministic rules, tiny test packs, structured cleanup, and only later earned the right to add hybrid retrieval and assistive layers.
Deterministic Python core
The first real product move was not chat. It was deterministic validation, typed rules, and a safety-oriented core that could be audited study by study.
Tiny evaluation packs
This gave us a sane sandbox: roughly 10 to 15 patient vignettes and a tiny bundle of papers and guideline examples, small enough to inspect manually and break on purpose.
Data sanitation and structuring
That meant normalizing fields, enforcing shaped records, and treating missing metadata as an explicit problem rather than a silent shrug.
Shared clinical vocabulary
This is where we learned that a deterministic system can still fail for semantic reasons when tags like `egfr-tki`, `tki`, and `targeted` mean the same thing to humans but not to machines.
Next.js + FastAPI on Vercel
This is where the demo became a product surface: Next.js on the front, FastAPI behind it, both deployed on Vercel so the flow could be exercised live.
Deterministic runtime online
This phase matters because it locked in the central promise: the online demo should be inspectable, repeatable, and willing to show uncertainty instead of bluffing.
Hybrid retrieval and vector infra
The semantic path grew in stages: semantic chunk storage, external vector search, dense plus sparse retrieval, rank fusion, and rescue back into deterministic decision logic.
Assistive LLM hooks
This keeps the story honest. We can use an assistive LLM layer where it adds ergonomics, but it does not get to impersonate the deterministic engine or rewrite the evidence contract.
Contributors
MIT AIML - Group 3:
BUNDYRA, WIDMER, ESPELAND, LEŚNIEWSKI, RIEKEN, THEIS.
Designed to demonstrate how a mission-critical evidence navigator can be honest about both its strengths and its limits.