The Problem

LLMs hallucinate. In emergency triage, an under-rated urgency level can be fatal. Generic LLM-based triage tools give confident answers with no structural guarantee against incorrect urgency assignments.

The Solution

MedEval's core design constraint: the LLM is structurally barred from making the urgency decision. A three-node LangGraph pipeline (extract → triage → explain): GPT-4o-mini converts free-text into 95 structured booleans, a deterministic Python rules engine assigns the ESI level by firing 68 ACEP handbook rules, and a second LLM call rephrases the rationale into lay language. The LLM explains decisions — it never makes them.

Engineering Decisions

Rules engine evaluates recursive AND/OR conditions against a merged fact pool. First A or B rule match is definitive. C rules set a base level but D rules can upgrade to ESI 2. Missing fields return False — pediatric rules safely no-op for adult patients. Built and published medeval-harness on PyPI (v0.1.0) — 50-case eval suite with exact/tolerant accuracy, under-triage rate, and p50/p95 latency. CI runs weekly against live production via GitHub Actions with --fail-under 60.

Business Value

Dual UI: patients see plain-English explanation only; doctors see full decision path and rule IDs that fired — making every triage fully auditable. Under-triage rate reduced 20% → 10% after prompt tuning. The evaluation harness is on PyPI so accuracy claims are independently reproducible.

Key Results

• 68 deterministic ACEP rules — zero LLM influence on urgency decision
• Under-triage rate: 20% → 10% — the safety-critical metric
• medeval-harness v0.1.0 published on PyPI — independently reproducible eval
• CI weekly eval against live production; build fails below 60% accuracy
• Live at medeval.rajkumarai.dev — HTTPS, Docker Compose, AWS EC2, Nginx

Stack

The Problem

Financial analysts spend days compiling research from 10+ scattered sources. Manual work doesn't scale and introduces inconsistency.

The Solution

A 7-step async LangGraph pipeline that generates auditable BUY/HOLD/SELL recommendations in under 30 seconds. GPT-4o constrained to explanation-only: zero influence over the decision, ensuring fully reproducible and auditable outputs.

Engineering Decisions

Normalized multi-factor scoring engine (Technical 0–25, Fundamental 0–40, Sentiment 0–15) with conflict-aware logic — signal disagreement (variance > 0.15) overrides to HOLD, eliminating false conviction. 5-factor confidence model blocks recommendations below 20% confidence. Redis TTL 900s caching reduces repeated query latency from ~15s to <100ms.

Business Value

Any firm paying analysts $80k+/year offsets that cost with sub-30s deterministic decisions. Portfolio ranking mode supports 2–8 tickers with proportional capital allocation.

Key Results

• BUY/HOLD/SELL decision generated in under 30 seconds
• Repeated query latency: ~15s → <100ms via Redis caching
• Zero LLM hallucination risk — GPT-4o explains, never decides
• Portfolio ranking mode: 2–8 tickers with proportional capital allocation

Stack

The Problem

Film studios spend millions producing trailers with no way to quantify emotional impact before release. Traditional test screenings are expensive, slow, and limited to small samples — no second-by-second signal.

The Solution

CineNeuro uses Meta FAIR's TRIBE v2 brain encoding model to predict fMRI activations across 20,484 cortical vertices for every second of a trailer. Three feature extractors — V-JEPA2 (vision), Wav2Vec-BERT (audio), Llama 3.2 (transcript) — feed TRIBE v2, mapping to 5 emotion channels across 7 brain regions.

Engineering Decisions

Two-tier deployment: g4dn.xlarge GPU ($0.53/hr) handles inference on-demand then terminates. t3.micro serves the React dashboard 24/7 using pre-computed JSON results (<100ms). The three extractors total 12.89GB — barely fitting in 15GB VRAM, requiring careful sequential model loading. PR #20 submitted to Meta FAIR TRIBE v2 — added int8 compute_type fallback for WhisperX on non-CUDA devices. CLA signed, checks passing.

Key Results

• 20,484 cortical vertices predicted per second across 7 brain regions
• 5 simultaneous emotion channels: excitement, fear, joy, suspense, boredom
• Pre-computed demo results served in <100ms from disk
• Full 2-minute trailer analysis in ~55 minutes on Tesla T4 GPU
• Monthly hosting cost: ~$0 (t3.micro free tier + on-demand GPU only when needed)
• PR #20 submitted to Meta FAIR TRIBE v2 — CLA signed, checks passing

Stack

The Problem

Enterprise knowledge locked in PDFs and Word docs means employees waste hours on search. Generic chatbots hallucinate without grounding.

The Solution

Production RAG system: PDF ingestion pipeline, ChromaDB vector store, MMR retrieval, and GPT-4 generation with real-time streaming. Hallucination guardrails block LLM calls when retrieval confidence is too low.

Engineering Decisions

MMR (Maximal Marginal Relevance) retrieval over plain cosine similarity — prevents near-duplicate chunks reaching the LLM. Rule-based query classifier (no LLM cost) detects factual/complex/ambiguous intent and adjusts retrieval params. Three API cost guards: rate limiting, input cap, tiktoken token budget — all fire before any LLM call. Citation format [SOURCE N: filename — Page: X] forces the LLM to cite, never invent.

Key Results

• Zero hallucination on document-grounded queries
• Two interfaces: Streamlit (users) + FastAPI Swagger (developers)
• Live on AWS EC2 with Nginx reverse proxy and HTTPS

Stack