A growing collection of working demos that turn AI strategy into something you can touch. Built on a single open-weight model. Full stack visible.
Each card opens an interactive demo. Telemetry, request bodies, latency and token counts are visible on every page — these are working AI systems, not marketing mockups.
Same input, three system prompts side-by-side. See how prompt engineering changes output, structure, and tone — and inspect the exact messages sent.
Open demo LIVEAsk questions of a curated EA & AI corpus. Retrieved chunks light up, similarity scores are exposed, citations link back to source.
Open demo LIVEA semantic map of 77 EA and AI concepts. Type a query, watch it appear in vector space alongside its nearest neighbors.
Open demo LIVEThree agents tackle a business question. The Researcher iterates over the corpus (up to 3 rounds), the Analyst weighs trade-offs, the Writer synthesizes — every step streamed live.
Open demo LIVETry to break a corporate AI assistant. Five attack categories scored by a regex fast-path and an LLM guard, each mapped to EU AI Act and NIST AI RMF.
Open demo LIVEDescribe an AI use case. Get a complete reference architecture — model approach, data, integration, infrastructure, governance, KPIs — and a generated Mermaid diagram, all from one structured-output call.
Open demoEach demo extends a technique from the core six — caching strategy, document extraction, agent self-correction, graph-augmented retrieval, evaluation infrastructure, AI-readiness assessment. Same technical layers, more capability per layer.
Sliders for monthly volume, tokens, cache hit rate, and model tier mix produce live monthly cost, per-tier breakdown, caching savings, self-hosting break-even, plus an LLM-written executive summary on demand.
Open demo LIVEPaste a contract, claim, or proposal. One structured-output call returns doc type, summary, color-coded entities, and severity-ranked risks — rendered inline with hover-synchronised highlights.
Open demo LIVETwelve questions across five dimensions of AI readiness. The radar chart fills in as you answer, then one structured-output call generates a sequenced 3-phase roadmap with concrete actions tagged by role and effort.
Open demo LIVEThree sequential model passes — generate, critique, refine — then a word-level diff between v1 and v2 with insertions and deletions highlighted. Optional judge call honestly says when the revision made things worse.
Open demo LIVEPick a corpus source. One structured-output call extracts entities and relations; a custom Canvas force simulation lays them out as a draggable, hover-aware knowledge graph. The bridge from flat retrieval to graph-augmented retrieval.
Open demo LIVE15 tests across 7 categories run in parallel against the model. Seven scoring methods, accuracy by category, hallucination flags, and a localStorage-backed baseline so you can detect regressions. Production eval methodology made visible.
Open demoSix demos taking the technical foundations from the upper four tiers and applying them to specific business functions — CX/Product, Sales, Finance/IR, Risk/Compliance, Procurement, Executive. Each mirrors a real artifact a team produces today, generated in seconds against the same open-weight model. The point is not the technique — it's the business outcome.
Paste 30–80 customer comments. One structured-output call clusters them into 5–8 themes with sentiment, severity, illustrative verbatim quotes, and a recommended action per theme. The first 30 seconds of work that used to take a product team a week.
Open demo LIVEPaste a meeting transcript. One structured-output call extracts attendees, decisions, action items with owner and priority, open questions, risks, topics with time-share, and a deal signal. The CRM update writes itself; the rep just confirms it.
Open demo LIVEPaste an earnings call transcript across any sector. One structured-output call extracts the headline number grid, growth drivers with magnitude, headwinds with severity, capital allocation signals, Q&A directness grades, and forward guidance vs consensus. The afternoon's work an IR analyst does — generated in seconds.
Open demo LIVEPaste a regulation update on the left, your business-process inventory on the right. One call returns a severity-coded impact matrix (rows: processes, cols: requirements) plus action cards sorted by severity with gap, required action, and owning role. The triage that takes a CRO's team a week — done in seconds.
Open demo LIVEPaste a weighted RFP requirement list and 2-3 vendor proposals. One call scores every (requirement, vendor) cell on a 1-5 scale with rationale, computes weighted totals, ranks vendors with strengths and concerns, and produces a recommendation with explicit confidence and caveats. Procurement's first deck slide, populated.
Open demo LIVEDescribe a strategic decision in plain prose — build vs buy, market entry, vendor consolidation. The model designs its own criteria framework with weight rationale, scores each option, runs sensitivity analysis on which assumptions would flip the answer, and recommends with confidence bounds and decisive factor. The capstone of the AI Lab.
Open demo LIVEA Jira-shaped Agile simulator. Pick or write user stories, then walk each one through four AI-assisted SDLC stages — refinement, technical breakdown, test cases with realistic pass/fail, and an animated CI/CD deploy log. Close the sprint to get a velocity-backed retrospective with action items. AI across the full software lifecycle, in one screen.
Open demo LIVEPaste a wire-transfer log or load a realistic preset — structuring, business email compromise, layering through tax havens. One structured-output call surfaces flagged transactions with risk scores, named AML patterns (BSA, FATF, OFAC, AMLD6 citations), counterparty risk, geographic concentration, and recommended dispositions from approve to SAR filing. The 90-minute analyst review compressed to seconds.
Open demoThe next set of business applications under design. Each targets a real artifact a team builds repeatedly — account management, customer success, founder ops, strategy planning, sales enablement, talent — and each compresses it from hours to seconds against the same open-weight model. Same architectural pattern, more functions covered.
Paste a strategic account's profile, recent activity history, and known stakeholder context. One call returns a complete account plan: stakeholder map with role and influence, whitespace analysis showing untapped product fit, the next 90-day expansion play with talking points, and a renewal-risk read. The deck strategic-account managers spend a Friday building.
Paste a customer's signals — usage trend, support tickets, NPS, contract value, lifecycle stage. One call returns a composite health score, churn risk band with primary drivers, renewal probability, and a tailored save-play with talking points and discount thresholds. The CS team's morning queue triage, automated.
Paste your monthly KPI snapshot and qualitative notes — wins, losses, hires, asks. One call returns a structured investor update: metric framing with vs-plan call-outs, narrative arc tying it together, prioritized asks block, and a TL;DR for skim readers. Four hours of founder time compressed to fifteen minutes.
Paste your company-level OKRs and the org structure. One call cascades team-level OKRs that ladder cleanly to corporate goals, flags conflicts where two teams claim the same outcome, surfaces gaps where corporate OKRs have no team owner, and proposes resolutions. The Q1 planning ritual, fast.
Paste your product positioning and a competitor's public profile (website copy, marketing claims, pricing if known). One call returns a rip-and-replace battlecard: where you win, where they win, top 5 objections with proof-point responses, ICP gotchas, and three open-with lines for the discovery call. The asset every sales-enablement team rebuilds quarterly.
Paste a role context — team mission, must-haves, must-not-haves, team gaps. One call returns a structured job description tuned to candidate language, the four-stage interview loop with focus areas per round, role-specific scorecards, and the calibration questions for the hiring panel. The structure great hiring managers build by intuition, made repeatable.
The AI Portfolio describes what enterprise AI can do. The Lab lets you do it. Two reasons it's here: first, AI strategy is more credible when the architect can also build. Second, telling executives that "AI works in production" is easier when there's a working URL they can click.
Each demo runs against a single open-weight model (Kimi K2 via NVIDIA's free inference tier) so the cost stays at zero and the architecture stays honest. No model arbitrage, no hidden fine-tuning — same reasoning engine across all six tiers, with the lift coming from prompt design, retrieval, orchestration and evaluation.
Stack: FastAPI · Uvicorn · Nginx · Oracle Cloud · NVIDIA NIM · Kimi K2
Backend: checking…