Prototypes, demos and R&D.

TasteTrend Analytics

RAG proof of concept, demo available · AWS Bedrock, OpenSearch, Lambda, React

AI retrieval over restaurant reviews: heterogeneous review files are normalised, indexed in OpenSearch and queried through a Bedrock agent that returns source-linked answers in a web demo.

Where the claim stops: A documented PoC run over four review inputs. No production SLA, no published retrieval score, no client outcome.

MarketIQ

Hackathon case study · React, Express, TypeScript, Qdrant, LangSmith

LLM market-intelligence workflow built at a PwC hackathon: one strategic question becomes ranked hypotheses, evidence references, actions and a workflow trace.

Where the claim stops: Core company and market data is synthetic. Decision support, not causal proof.

Scoutbound

Agent workflow, prototype · TinyFish, TypeScript, Express, SQLite

Agentic prospecting workflow: research leads with a browser agent, keep the evidence, score fit with reviewable reasons, export to CSV/JSON or a CRM handoff. Live, mock and degraded modes are explicit.

Where the claim stops: Mock mode is the default. Extraction quality is not benchmarked publicly.

Revon

RevOps, under development · React, Express, Prisma, PostgreSQL, Python

Multi-tenant lead workspace with explainable rule scoring, background workers, reporting and ML training/inference scaffolding.

Where the claim stops: No trained model artifact, calibration report or production outcome is claimed.

AI Research Intelligence

Local RAG, R&D · Python, FAISS, MiniLM, Qwen, FastAPI

CPU-first research assistant: arXiv ingestion, exact FAISS retrieval, cross-encoder reranking and local Qwen synthesis with no paid inference API at runtime.

Where the claim stops: Model, corpus and index are external downloads. Quality evaluation is heuristic.