ResearchAgent
September–October 2026Search free sources first, then decide whether another piece of evidence is worth buying. This team demo tracks budgets, checks delivery proofs, and handles challenges and refunds with a synthetic corpus and testnet or simulated payments.
Presented on the main stage at AI Tinkerers Singapore, 10 October 2026. Top three Cloud Run demos; earlier SingHacks finalist.
Team project · main-stage presentation02 / Python + DuckDB
talktoit
2026Turn an Apple Health export into a database you can ask questions of. A streaming importer feeds DuckDB, dashboards, and natural-language queries, with local defaults and explicit opt-in remote modes.
The unglamorous importer became the interesting problem.
Self-hostable application03 / TypeScript + Bun
YALD
2026A read-only view of coding-agent usage, from local logs to portable reports. Explore tokens, cache use, and API list-price estimates while keeping missing measurements visible.
An estimate is useful when you can see its limits.
v0.1.0 GitHub preview04 / Python + NumPy
Plotline
October 2026How much of a forest map does a screening decision actually need? A geospatial prototype stores the facts a defined rule uses, then compares compact and original answers through reproducible checks and an inspection UI.
Keep the decision. Measure what was left behind.
Measured hackathon prototype05 / Python + Neo4j
StyleMind
April–May 2026Fashion recommendations with some memory of your taste. Conversation becomes a structured persona, which guides graph and vector retrieval before the assistant puts an outfit together.
Preferences have relationships, not just keywords.
Working prototype06 / Python + Airwallex
Handoff
September 2026Before an invoice becomes a draft, check it against the contract, accepted delivery, and prior billing. A team finance demo surfaces conflicts, asks for approval, and reads back the Airwallex sandbox result.
A known-error demo, with the evidence left in view.
Team hackathon prototype07 / Go + MCP
MandateFlow
August–September 2026An agent's permission should depend on where its information came from. Contributed the Go provenance and authorization gateway, policy checks, and demo hardening to this team build on the supplied Agent Launchpad starter.
A tool call needs a reason you can trace.
Team hackathon proof of concept08 / Python + FastAPI
Show Me Your Agents
August 2026A learning event becomes a bounded diagnostic job, then a mastery proposal with supporting evidence. This early tutoring slice keeps a human tutor in charge of reviewing the proposal.
A proposal can be useful without becoming a verdict.
Early vertical slice09 / Python + BM25
Payment & incident investigations
September 2026Two authored submissions to the Julius Baer challenge: payment investigation with retrieval tools, and deterministic incident correlation with quoted evidence and calibrated uncertainty. Built within the organizer's starter, with the submission still unmerged.
A useful answer should show why it might be wrong.
Submitted · unmergedFitback / Detect-T
2024–2025The public camera-guidance proof of concept behind the team Fitback chapter. Pose feedback and capture run in the browser, while an older model runtime lives in a worker so incompatible TensorFlow versions can coexist.
One camera flow, two model runtimes.
Historical component prototype