Julian's ideas

Notes

Founders ship AI features and only discover the inference bill after launch: cost per user is 10x the price, or one power user burns a month of revenue. Existing token calculators multiply a guessed token count by a list price — they miss what actually drives cost: context growth across a session, RAG chunk sizes, agent loop iterations, retries and failed generations, system-prompt bloat, cache hit rates, and which model each call actually routes to. The idea: a pre-launch token-consumption and profitability estimator. Input: a product spec (features, expected session shape, candidate models) plus, when available, real API traces from a prototype (OpenTelemetry / Langfuse / Helicone format). The estimator simulates usage — sessions per user, context-window growth, retrieval payloads, loop depth, retry rates — and tokenizes with the real provider tokenizers (not a chars/4 heuristic), producing a cost distribution per action, per user, and per month (p50/p95, not just the mean). From that it derives unit economics: break-even price, gross margin at candidate price points, and a sensitivity map showing which lever moves cost most (model choice, context trimming, prompt caching, distilling easy calls to a smaller model). It also recommends concrete margin plays — cacheable prefixes, routing easy queries to a cheap model, truncation budgets — each with estimated savings. Output is a one-page pre-launch inference P&L plus an API/CI gate that flags when a feature's projected cost per user exceeds a budget. Monetization: free single-model estimate, then 20-40 EUR/month for multi-model simulation, trace ingestion, and CI integration; enterprise for procurement-level forecasting. Why now: AI features are table stakes, but inference is the COGS nobody models before launch, and margin surprises are a top reason AI products quietly die. Key constraints: usage patterns are genuinely hard to predict pre-launch (mitigated by prototype trace ingestion and conservative p95 defaults), model price volatility (track prices continuously and re-run estimates), and cross-provider tokenizer differences (use real tokenizers per provider).

Notes

Every LoL player knows the feeling of losing and not understanding why. Replays exist, but reviewing them yourself is slow and biased — you see what you did, not what you should have done. The idea: a personal AI coach that ingests your replays (via the Riot Games developer API or local .rofl files), reconstructs the game state, and produces a structured post-game debrief: what went wrong in key moments (a botched engage, a missed ward, a bad objective trade), why it went wrong, and what a better play would have looked like — with timestamps and minimap visualizations, not generic tips. How it works: parse the replay into events (combat, objectives, vision, gold/XP curves), then compare your decisions against (a) your own historical baseline (are you repeating the same mistakes?), (b) champion- and role-specific heuristics (level-6 power spikes, wave state, cooldown tracking), and (c) LLM analysis of key moments with the reconstructed game state as context. Output: a short per-game debrief (top 3 mistakes + 1 thing you did well), a weekly trend report (which mistake categories are improving), and a focus drill suggestion (this week: stop overextending before level 6). The data moat is longitudinal: the more games you feed it, the more it knows your specific tilt patterns, champion pool, and rank trajectory — a personalized coach, not a static guide site. Monetization: freemium (1 analyzed game/day), then €5-8/month for unlimited analysis, trend tracking, and champion-specific deep dives. Start with one role (mid or support, where decision density is highest) and one region's player base. Key constraints: Riot's API terms of service and rate limits for replay data, the complexity of game-state reconstruction (vision fog, minion waves), the risk of the coach giving confidently wrong advice (it must cite concrete events, not vibes), and competition from existing tools (OP.GG, Blitz, U.GG) — the differentiator is the personalized, longitudinal coaching loop, not static stats.

Notes

Everyone pays for 2-3 AI subscriptions but needs 10. Instead of each person buying everything, pool complementary access: a marketplace where you offer what you have and request what you lack. You have Figma Pro? You handle Figma requests for others. Someone else has Midjourney, Runway, or ChatGPT Pro and handles those for you. Exchange via credits, not credential sharing. Model: members list their active subscriptions (verified via OAuth / invoice proof, no password sharing). Requesters post a task ('generate 20 icons in Figma', 'run this prompt on GPT-4o with my files'). Providers claim, do the work inside their own authorized account, deliver the artifact, earn credits. Credits are spent to request tasks on tools you lack. Platform handles reputation, delivery SLAs, escrow, and notifies both sides. Optional pooled team plans (where ToS allows) as a second tier. Trust & compliance is core: no session/token sharing, work is done by the account owner who already pays — a service exchange, not account rental. Clear boundaries on IP (requester owns output), privacy (temporary file access, auto-expiry), and ToS (each provider must check their plan allows commercial work for others; many do for freelancers). Pricing: freemium (3 trades/month), then €9/mo or platform fee on paid credit packs. Complements previous ideas: like donating tokens to OSS, this turns sunk subscription spend into liquid value. Key constraints: Terms of Service variation by provider, quality/reliability of peer work, dispute resolution, and building enough liquidity (critical mass of complementary tools) in the first city/community.

Notes

Millions of developers and teams pay for LLM API credits and subscriptions where 20-40% of tokens go unused each cycle. Simultaneously, open-source maintainers burn their own credits building free tools that everyone depends on. The idea: a donation layer between inference providers and OSS. Users connect their OpenRouter / OpenAI / Anthropic / provider accounts (or donate via the platform's own pooled credits) and allocate unused tokens or a dollar amount to a curated registry of OSS projects. Maintainers register their repo, verify ownership (GitHub OAuth + package registry), set a monthly token need (for CI, issue triage agents, doc generation, code review bots), and receive credits they can spend on any supported model through the platform's proxy. Donors get a tax-deductible receipt, a leaderboard/badge, and attribution on the project page. The platform takes a small fee (5-8%) and handles the hard parts: model-agnostic credit accounting, anti-gaming (reputation + usage proofs, no cash-out), pooled procurement at wholesale rates, and compliance. Why now: AI-assisted OSS maintenance is exploding (agents for PRs, issues, security fixes) but funding is still stuck on GitHub Sponsors + grants. A token-native funding loop ties AI usage directly to the commons it benefits. Key constraints: provider ToS on transferability of credits, preventing wash trading / fake repos, and managing price volatility across models. Start with OpenRouter as the neutral router; expand to direct providers.

Notes

Many bureaucracies (Germany: Behörden, Amtsgerichte, Krankenkassen; US: healthcare, legal, government) still legally require fax or physical Einschreiben/Einwurf-Einschreiben, even when the internal workflow is fully digital. Developers and ops teams waste weeks integrating fragmented fax gateways (T.38/SIP) and print-to-mail providers, with no unified tracking, templating, or compliance story. The idea: a single developer-first REST API that abstracts both channels behind one call: POST /v1/send { to, channel: 'fax'|'mail', templateId, data, options }. The service handles template rendering (HTML/Handlebars → PDF, with optional LaTeX for high-quality German formal letters), e-signature placement, fax transmission with retries and T.30/T.38 fallback, and postal fulfillment (print, envelope, frank, handover to Deutsche Post / USPS / La Poste with tracking). Delivery confirmations come via webhooks (fax transmission report, postal scan events) and a unified audit log for GDPR / GoBD compliance. Complements the Bürokraten-Pilot idea: that agent drafts and explains; this API actually sends. Monetization is per-page + postage/fax-minute, with enterprise tier for high-volume senders (insurances, law firms, HR, medical). Key constraints: per-country telecom/postal regulations, deliverability and sender-ID validation, cost vs. free E-Post alternatives, and strict PII handling (end-to-end encryption, automatic deletion, EU data residency).

Notes

Web development in AI coding harnesses enjoys instant visual feedback via embedded browser webviews, but building desktop applications (Electron, Tauri, Flutter, Qt) still forces developers and agents to manage external OS windows without in-harness visual debugging. The idea: embed a live, interactive desktop preview directly inside the coding harness workspace. This can be implemented through two complementary paths: 1. Modified runtime package: A tailored Electron/Tauri runner that renders off-screen and streams framebuffer updates over local IPC/WebSockets into the harness preview pane, while streaming synthetic pointer and keyboard events back to the runtime. 2. Harness-level virtual display: A harness extension running a lightweight virtual display server (e.g. headless X11/Wayland pipewire or macOS window capture) that isolates and embeds any native desktop application process directly into the editor UI. Beyond visual rendering, this gives coding agents direct access to the app runtime state, accessibility tree, and UI component hierarchy for autonomous layout verification, visual regression testing, and interactive clicking during automated self-verification loops. Key constraints: cross-platform windowing APIs (macOS WindowServer, Windows DWM, Linux Wayland), GPU rendering overhead with offscreen buffers, input latency, and security sandboxing.

Notes

Investigate whether life and evolution can be defined independently of organic chemistry, molecules, or any particular material substrate. Treat life as an open quantum-information process: a system that preserves, transforms, reproduces, and variationally selects quantum states while remaining far from equilibrium. Can self-maintenance, heredity, variation, and selection emerge entirely through quantum correlations, open-system dynamics, and measurement-feedback loops, without molecular replication? Key research directions: - Substrate-independent criteria for life formulated in terms of information theory, non-equilibrium thermodynamics, and quantum resource theories. - Quantum analogues of genotype, phenotype, mutation, replication, selection, and ecological niche competition. - Exploring whether quantum error-correcting codes, topological states, or structured environmental decoherence can support persistent hereditary information against thermal noise. - Open quantum systems in which information-bearing states compete for limited channel capacity, coherence budgets, or free energy. - Falsifiability target: Define an experimental or simulated system whose quantum state dynamics exhibit persistent replication with heritable variation and selection, then verify whether those evolutionary dynamics collapse when purely quantum resources (entanglement, coherence) are removed.

Notes

The idea: make agentic AI opt-in and auditable the way OAuth 2.0 makes API access opt-in, with two complementary modes. Interactive mode — the visitor sees a consent screen ("This portfolio wants to run a short research agent with these scopes: read-only access to X, time budget 30s, no writes. Allow / Deny") and grants a one-time, scoped credential, like an OAuth access token with a short TTL and fine-grained scopes. Upfront / non-interactive mode — the user defines the agent's roles, tasks, and permissions once; an agent token is minted and the agent authenticates itself per action without re-prompting the user, strictly bounded by that pre-granted scope. Either way it turns a black-box helper into a transparent, permissioned actor the user explicitly authorized, so the agent's actions stay within an auditable grant that can be read, reviewed, and revoked. Why this fits LLM agents: just as OAuth lets you grant a third-party app limited access to your Google Drive instead of your whole account, "OAuth 2.0 for agents" lets you give an agent bounded capability (which tools, which time budget, which data, which roles/tasks) instead of open-ended autonomy — and the upfront-permissions model is what makes multi-step, multi-tool agents usable without spamming the user with prompts and without the agent overstepping what it was told to do. Real constraints: agents that chain tools are hard to budget (token + tool-call + time costs can explode), mapping coarse "roles/tasks" onto fine-grained OAuth scopes is non-trivial, users rarely audit their own grants, and there is no standard "agent token" format or revocation UX yet — so the win is strongest for high-trust, visitor-facing demos and long-running assistant workflows where transparency and user control matter more than raw capability. No liability for damages.

Notes

Most RAG pipelines hard-code the chunking and store the result in a flat vector index, but the best retrieval structure depends on the kind of knowledge in the dataset. The idea: a representation-learning layer that, given a new corpus, discovers the structure that makes retrieval most accurate — often a knowledge graph where entities and relations are themselves learned, not hand-annotated. Why this fits LLMs: their hidden representations of tokens and documents are geometry-sensitive, so a graph that mirrors the manifold in which similar concepts cluster naturally (similar weights, shared attention neighborhoods) tends to vibe with the encoder and the generator at the same time. It learns, per dataset, whether dense co-occurrence, cross-encoder hard negatives, entity-relation triples, or a hybrid graph best improves exact and latent semantics — and re-ranks the index layout so that top-k retrieval aligns with the model's own learned similarity, not an arbitrary chunk size. Real constraints: learning the graph adds compute at ingestion, the optimal structure shifts as the base model changes, and evaluation is expensive; the win is strongest on dense, interlinked corpora (laws, specs, wikis) where a flat chunk index loses cross-ref signal. No liability for damages.

Notes

The idea: treat biology as a formal, mathematical language rather than a messy set of exceptions, and ask which structures from pure math actually carve nature at its joints. Is there a 'grammar' of life where number theory shows up not as metaphor but as mechanism? Concrete signs it is real: Fibonacci phyllotaxis — the Golden-angle spiral counts in sunflower heads, pinecones, and romanesco — is number theory literally shaping plant growth; reaction-diffusion (Turing patterns) turns partial differential equations into animal coats and digit formation; algebraic geometry describes the shape spaces of proteins and RNA secondary-structure folds; graph theory captures metabolism and neural wiring; group theory describes the symmetries of viral capsids and protein oligomers. The pull for AI: if biological organization respects deeper mathematical invariants, then a model 'speaking' that language — embeddings that respect the same symmetries/group actions, or a token space shaped like a phylogenetic tree/graph — could generalize from far fewer examples than a generic transformer, because it would be fitting the true generative structure instead of learning surface statistics. This fits the broader hunch that a lot of AI progress is just catching up to structures biology already uses. Real constraints: most of these correspondences are inspirational, not yet engineering-grade — phyllotaxis has alternative mechanistic explanations, Turing patterns need very specific parameter regimes, and translating a symmetry group into a useful neural loss is still hand-wavy; biology is also noisy, redundant, and historically contingent, so 'a mathematical language' may be a useful fiction rather than a literal truth. No liability for damages.

Notes

Built for ordinary German citizens facing the jungle of online bureaucracy portals: Anmeldung, Bürgeramt appointments, tax office forms, benefit applications, immigration paperwork, and more. The agent runs locally (Docker + Ollama), so user documents and personal data never leave the machine — GDPR-compliant by construction, and usable in a shared flat or on a train without uploading anything. It can read PDFs of official forms, extract the relevant questions, pre-fill them, walk the user step by step, and draft cover letters or explanations in plain German. It ships with curated, open datasets of known public-service procedures (mapped per Bundesland where possible) and template libraries for common requests. Funding model: free, open source (GPL-3.0), with optional paid support packs for municipalities and nonprofits. No liability for damages: the agent makes no decisions for the citizen and never signs electronic documents; it only drafts and explains. Real constraints: official PDFs are often scanned and OCR-resistant, forms change frequently, and Germanys public digital services are splintered across too many portals — so the value is in reducing paperwork friction and mistakes, not replacing the Behörde.

Notes

Starlink needs a clear view of the sky, but laptops usually sit under desks, in bags, or indoors where the dish can't see satellites. The idea: integrate a flat Starlink terminal into the laptop lid, paired with a small router and high-capacity battery, so the machine can do on-device AI over satellite wherever there is any view of the sky. Useful for field AI, disaster response, remote research, long-haul travel, and places with no mobile coverage. The real constraints: Starlink hardware draws serious power, the terminal must stay unobstructed and well-pointed, and fitting a phased-array antenna into a thin lid is hard; thermals and RF exposure matter too. Pairing it with local model serving keeps data usage and latency down over the satellite link.

Notes

Laptops mostly ship without a SIM slot, so mobile internet today means tethering to a phone, which drains the phone battery and heat-soaks it. A small USB-C dongle that accepts a nano-SIM and acts as a USB modem (RNDIS over LTE/5G) gives the laptop its own connection and its own power budget. Keep it cheap, open-source on firmware, with eSIM support for a second profile. Good for travel developers, field work, backup connectivity, and places with no wifi. The real constraints are power and heat on such a small board, antenna tuning in a metal chassis, and carrier/eSIM certification. With eSIM provisioning built in, it could switch profiles and carriers without swapping the card at all.

Notes

Manufacturers of products with digital elements now face the EU Cyber Resilience Act (CRA), which requires a Software Bill of Materials (SBOM), a coordinated vulnerability disclosure process, and a declaration of conformity with a cybersecurity summary. This idea: a service that generates and continuously manages SBOMs (SPDX and CycloneDX) straight from the build and release pipeline, then maps every component to CRA obligations, such as the essential and important cybersecurity requirements, the five-year security-support commitment, and the pre-market and post-market reporting duties. It ingests CVE and VEX feeds, enforces a component allowlist or denylist, and emits the artefacts manufacturers must retain or submit: the SBOM, the CVD plan, the cybersecurity summary, and the EC declaration of conformity. The payoff is strongest for AI products: an SBOM that also covers models, datasets, and inference dependencies lets CRA obligations and the EU AI Act overlap be satisfied together on model risk and incident reporting.

Notes

The product is a conversational assistant powered by Grok that consumers can use at no direct cost. Revenue comes from sponsored content and advertising rather than subscriptions, which removes the payment barrier that keeps many users away from paid AI assistants. The main challenge is balance: ad placement must stay unobtrusive, targeting must respect privacy, and the level of ad load has to cover model costs without degrading the conversation experience. The model has the strongest fit for audiences where monthly plans are not viable, such as students and price-sensitive markets.

Notes

Most CRM software is built for sales pipelines, but the people in your life are not leads. A personal CRM would track the relationships that actually matter: birthdays, interests, what someone is going through, and when you last talked, then quietly nudge you to reach out before too much time passes. It is a memory aid plus a gentle reminder system, designed for a few hundred people rather than a million records. Two hard problems: making entry so effortless it actually happens, and keeping the data private, since it is deeply personal.

Notes

After every sales call is recorded and transcribed (natively in Meet/Zoom, or via a dedicated recorder), an agent analyzes the transcript for the one or two most useful things to improve, such as a missed objection, a weak next step, or a phrase that cost the deal, and drafts a short, specific, actionable notification delivered straight into the Gmail inbox. It lives where the seller already works, so there is no new dashboard to check. The feedback is grounded in the actual call, not generic playbook advice, and it arrives right after the meeting while the details are fresh.

Notes

Aggregate funding data such as deals, rounds, ticket sizes, valuations, and active investors, and surface the patterns behind what VCs are actually paying attention to: hot sectors, repeated investor themes, and the signal-to-noise of round characteristics. Instead of gut feeling or anecdote, you get a ranked, sourced view of what is sparking interest right now. An early-warning radar for founders deciding where to build next.

Notes

A single MCP endpoint that answers 'which model should I use for X?' backed by real benchmark data such as reasoning, code, latency, price, context window, and EU availability. Instead of chasing leaderboards and Reddit threads, humans and agents query one index that ranks models against their actual constraints: task type, budget, latency budget, and GDPR/EU AI Act fit. Curated, versioned, and open.

Notes

The Casimir effect is an attractive force between closely spaced surfaces caused by vacuum fluctuations (zero-point energy). The idea: engineer nanoscale structures where that force is asymmetric, producing directed motion, a vacuum-driven actuator with no moving parts in the classical sense. Scale it across an array and you get propulsion. Two honest caveats: the force is minuscule at practical scales, and extracting usable energy from the vacuum is a famously contentious topic, so the realistic target is actuation and precision positioning (e.g. MEMS, micro-robots) rather than free energy.

Notes

An open device you drop on the table that records who said what. On-device transcription, no cloud involved. The software side syncs and manages transcripts locally and runs a local model (e.g. Ollama) for summaries and action items. Built for GDPR: nothing leaves the room. Hardware: off-the-shelf mic array + a small board.

Notes

A cheap, open-source robot that takes care of a home veggie garden: watering, light, soil sensing, and a camera to spot pests. Everything on standard, low-cost parts like a printable frame, stepper, and pump, with firmware and app fully open source. The point is a <$100 starter build and a friendly community.

Notes

A programming book written for how people actually read now: with AI agents as co-readers. Every chapter ships as source + prompts, so an agent can build the examples, answer questions, and keep the book current. Not AI-washing a normal book: the structure itself is agent-native. I am actively pursuing this idea, and it is live at https://agentic-book.org.