Ideas Radar: 2026-04-27
Yesterday's idea radar leans heavy on AI-adjacent infrastructure that nobody is shipping yet — the trust layer for agents, the verification layer for organizational knowledge, the routing layer for prediction markets — and a smattering of consumer pain that's been sitting around long enough to be embarrassing. The recurring tell: the gap is rarely a missing model. It's the boring scaffolding that lets the model do real work without supervision.
#1
A research assistant trained on every publicly registered clinical trial protocol — arm by arm, endpoint by endpoint, eligibility criterion by eligibility criterion. ClinicalTrials .gov is a public goldmine, but the indexing is shallow and most BD/medical-affairs teams still bounce around it manually. A model that can answer "what trials have ever tested this combination at this dose in this population" in 30 seconds replaces a week of human work.
Source: https://x.com/heytambor/status/2048009637031714880
Source: https://x.com/heytambor/status/2048009637031714880
#2
Native Mixture-of-Experts liquidity routing for prediction markets. Today every Polymarket market is one shared brain — smart traders, noise, panic, insiders, random bets, all averaged into one price. The best wallets already specialize: one only catches arbitrage, one only trades politics, one only reads orderflow. That's already MoE happening manually. The product: route flow across specialized market makers, let experts compete on their own slice, isolate noise instead of letting it infect the whole price. Hundreds of small brains in parallel beats one smarter market.
Source: https://x.com/Marko_Poly/status/2048119817014530455
Source: https://x.com/Marko_Poly/status/2048119817014530455
#3
A non-Figma slide tool that Claude Code or Codex can natively drive. Today people pipe Claude Design through screenshots and HTML kludges to get slides done. What's missing is a real MCP or CLI for slides — a software-language design layer for slides — so a coding agent can iterate on a deck the way it iterates on a webpage. Slide generation is one of the most-requested business workflows; it shouldn't require a separate AI app.
Source: https://x.com/NiravJ3/status/2047933245694857646
Source: https://x.com/NiravJ3/status/2047933245694857646
#4
The persistent organizational memory and verification layer for enterprise AI. Every dollar of AI investment goes to applications — chatbots, copilots, agents, automations — but underneath there's nothing. No way to know if the context the model is reasoning over is accurate, current, or internally consistent. Same model. Wildly different outputs. The opportunity is the index that sits between an enterprise's docs/Slack/email/Drive and the agent that touches it, with verification baked in so a single bad source can't poison the whole knowledge graph.
Source: https://x.com/Raakin/status/2047935564469666255
Source: https://x.com/Raakin/status/2047935564469666255
#5
The trust layer for autonomous agents. Today any agent can act, there's no unified identity, no trust between agents, no accountability. As soon as agents start paying agents, trading 24/7, running businesses, the bottleneck is governance of machine capital. Concretely: a "Know Your Agent" registry, an oracle that validates agent actions and verifies execution, a permission layer that decides which agents can move which kinds of capital. Wallets execute, AI decides — but something has to authorize.
Source: https://x.com/genrih99999/status/2048068509670814025
Source: https://x.com/genrih99999/status/2048068509670814025
#6
A "DumbDetector" for AI models. Like DownDetector, but crowdsourced reports of which models are being measurably dumber than usual today. Anthropic spent six weeks shipping silent harness regressions before the postmortem caught it. A community signal layer would have flagged it on day three. Tens of thousands of devs hitting an outage page is more reliable than waiting for the lab's internal evals.
Source: https://x.com/narphorium/status/2047850795211894969
Source: https://x.com/narphorium/status/2047850795211894969
#7
A model-comparison benchmark site that surfaces intelligence, not just throughput. Current local-LLM dashboards rank by tokens/sec. The result is that 2GB toy models drown out genuinely smart 70B+ "hero quants" running on serious hardware. The product is a leaderboard segmented by model size class and intelligence-per-watt, with first-class space for hero quants that don't win on speed but win on every reasoning eval.
Source: https://x.com/banana_baeee/status/2048130332181123162
Source: https://x.com/banana_baeee/status/2048130332181123162
#8
The non-AI distribution gap. Law firms still run on referrals. Med spas still buy Google Ads at $8 CPM. Staffing agencies pay LinkedIn rates for candidates reachable through creator channels at a third of that cost. The category isn't underserved by product — it's underserved by 2010-era go-to-market. The opportunity is a vertical agent that lives where the customers actually are (TikTok, podcasts, niche newsletters) and treats the regulated channel as the legacy fallback, not the default.
Source: https://x.com/ItIsRaymo/status/2047980464934469900
Source: https://x.com/ItIsRaymo/status/2047980464934469900
#9
A unified social-event-to-calendar handshake. Every event posted on X requires the reader to either screenshot, transcribe, or hunt for the host's separate Eventbrite. The fix is laughably small: a recognized format that turns "this Saturday 7pm at Bar Bayeux" into a one-click ICS download. Twitter could ship this in a sprint. The fact that it isn't already standard is the clearest signal that the social platforms are leaving real product work on the floor.
Source: https://x.com/roger/status/2047845164132213173
Source: https://x.com/roger/status/2047845164132213173
#10
Letterboxd for music. Recommendation has fragmented across Spotify auto-recs, TikTok memes, Instagram Reels, and aging blogs. Letterboxd nailed the hybrid of "personal log + community ratings + curated lists" for film. Music has Discogs (collector-focused), Last.fm (data-focused), Album of the Year (review-focused) — none of them feel like a place to socialize taste. Whoever rebuilds Letterboxd's product DNA for albums catches the next decade of music discovery.
Source: https://x.com/kapam3s/status/2048151033965125649
Source: https://x.com/kapam3s/status/2048151033965125649
#11
A media-only filter for X. Right now if a video clip is annoying you across the timeline, the only remedy is muting every poster. Block/mute by media (per-clip, per-video-id) is the missing primitive. It would make the platform meaningfully usable during news cycles where the same low-effort clip gets reshared 200 times in 24 hours. Twitter's existing infra already fingerprints media for copyright; flipping that toward end-user mute is one feature flag.
Source: https://x.com/nickclark2311/status/2048051889498132609
Source: https://x.com/nickclark2311/status/2048051889498132609
#12
A scrollbar-thickness preference at the OS or browser level. macOS uses thick. iOS, visionOS, Android, Windows all use thin. Designers and devs working across platforms hit this every day. A single user-pref toggle (or web platform feature) would close it. Small idea, but the kind of small that compounds — every OS-level inconsistency is a hidden tax on cross-platform engineering.
Source: https://x.com/lachlanjc/status/2048172434764050838
Source: https://x.com/lachlanjc/status/2048172434764050838
#13
A multi-game ship-design marketplace. People who play space games (Eve, Star Citizen, Elite Dangerous, etc.) build elaborate ship configurations and then have nowhere to share or sell them. The user explicitly says: "I wish there was a place to post your ships and for me to buy them." Real money already flows in these communities through grey-market discord channels — a clean Stripe-backed marketplace would skim a meaningful slice. Adjacent: Genshin/Wuwa team builds, factory layouts in Satisfactory/Factorio, base designs in Rust.
Source: https://x.com/50PlusGamer1970/status/2047860066032992307
Source: https://x.com/50PlusGamer1970/status/2047860066032992307
#14
Voice-to-dream-journal. The premise: "I had a banger dream last night and can't for the life of me figure out how to write it down." Right now the workflow is wake-up → fumble for a notes app → memory already half gone. The product is a sleep-side voice recorder that listens for the user mumbling/talking on wake, transcribes, time-stamps, and feeds into a searchable dream archive. Pair with an LLM that asks targeted clarifying questions during the first 60 seconds of consciousness while the dream is still vivid.
Source: https://x.com/TaigataRRRo/status/2048156066111004955
Source: https://x.com/TaigataRRRo/status/2048156066111004955
📡 Eco Products Radar
Eco Products Radar
Polymarket open-source agent stack — py-clob-client, polymarket/agents, poly-market-maker, clob-client. Cited as the obvious entry point for prediction-market product ideas (MoE liquidity routing, AI hedge bots, settlement-rail integration).
Perplexity Computer / Comet — At least two posts in this batch reference "give Perplexity Computer one prompt" workflows for market research. The agent is becoming the de facto research assistant for solo founders mapping product gaps.
Claude Design + Claude Code handoff — The pattern shows up as both an opportunity and a complaint. People want a design-to-implementation pipe; the slide-tool gap is the cleanest unmet need.
ClinicalTrials .gov public dataset — The clinical trial protocol corpus is 100% public and almost completely under-served by AI tooling. Every BD/medical-affairs team and biotech KOL desk needs an analyst on top of it.
Cerberus / agent-security tooling — Mentioned across multiple "agent accountability" posts. The "warn vs act" gap (other tools warn, Cerberus revokes) is the wedge.
Polymarket open-source agent stack — py-clob-client, polymarket/agents, poly-market-maker, clob-client. Cited as the obvious entry point for prediction-market product ideas (MoE liquidity routing, AI hedge bots, settlement-rail integration).
Perplexity Computer / Comet — At least two posts in this batch reference "give Perplexity Computer one prompt" workflows for market research. The agent is becoming the de facto research assistant for solo founders mapping product gaps.
Claude Design + Claude Code handoff — The pattern shows up as both an opportunity and a complaint. People want a design-to-implementation pipe; the slide-tool gap is the cleanest unmet need.
ClinicalTrials .gov public dataset — The clinical trial protocol corpus is 100% public and almost completely under-served by AI tooling. Every BD/medical-affairs team and biotech KOL desk needs an analyst on top of it.
Cerberus / agent-security tooling — Mentioned across multiple "agent accountability" posts. The "warn vs act" gap (other tools warn, Cerberus revokes) is the wedge.
Comments