🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
DrugLens - Geriatric Polypharmacy Safety Auditor
DrugLens addresses the geriatric polypharmacy crisis — but the real problem isn't that EHRs miss interactions. It's that they flag everything, so clinicians ignore up to 90% of alerts and the dangerous ones get lost. We attack precision, not volume. A deterministic rules engine runs on EVERY request in ~3ms: 102 curated drug-drug interactions, 61 AGS Beers Criteria 2023 rules, and 38 STOPP + 20 START v3 criteria. Three things make it different from a lookup table: 1. TRUE COMBINATION LOGIC. "Opioid + benzodiazepine" and the renal "triple whammy" fire only when EVERY component is present — never on a single drug. This AND-semantics is the core defence against alert fatigue, and it's locked by tests. Our safe demo case produces ZERO alerts. 2. eGFR AND AGE GATING. Metformin is silent at eGFR 90 and flagged at eGFR 25. The same drug list at age 30 produces 0 Beers violations; at age 85 it produces 2. Change one field and watch the rules activate live. 3. TOKEN-EFFICIENT ROUTING. Low-risk patients are answered entirely offline for ZERO LLM tokens. Only MODERATE/HIGH cases escalate to the cloud. For the "novel drug" blindspot, every drug pair with no database entry is evaluated by the cloud model, grounded in each drug's PubChem SMILES structure — surfacing interactions no lookup table contains. The rules can never be hallucinated or skipped: the deterministic engine always runs first; the LLM only narrates findings it cannot invent or suppress. React 19 + FastAPI, single-origin deployment, 75 offline tests green in CI, deployed live and open. Every statistic in the UI is read from the loaded rulesets at runtime, never hardcoded.
13 Jul 2026
Loan Shark: Multi-Agent AI Banking Pipeline
Loan Shark is an enterprise-grade, multi-agent AI pipeline designed to automate and secure the loan processing workflow. Traditional lending pipelines are slow, fragmented, and vulnerable to compliance oversights, often taking 3 to 14 days to pass applications across credit, fraud, and legal departments. Loan Shark replaces this manual chain with a pipeline of 9 specialized, task-focused AI agents: Intake, Document, Credit, Fraud, Risk, Compliance, Decision, Pricing, and Communication. Coordinated entirely via stateless P2P messaging using the Band SDK, each agent parses the application, appends its unique analysis, and @mentions the next agent. The entire execution takes under 5 seconds. Crucially, the system enforces strict decision safety for high-stakes workflows: 1. Compliance Agent: Enforces regional guidelines (such as Indian RBI exposure limits and KYC rules) and acts as a mandatory gate that can issue overrides. 2. Immutable Audit Trail: Because agents communicate in a shared Band chat room, the persistent message history functions as a tamper-proof audit log. 3. Human-in-the-Loop Gate: The system does not dispatch decisions autonomously. Instead, a custom React officer dashboard displays the real-time pipeline status and presents a drafted sanction or regret letter for physical review and authorization.
19 Jun 2026
