Sarniq Intelligence

From raw informations to clarity

Comment

Inspiration

In 2024–2025, we watched markets, supply chains and entire regions move on headlines that turned out to be wrong, exaggerated, or pure noise. Retail investors, journalists, SMBs and family offices were left with two bad options:

  • Free news feeds — fast, but unverified, biased, and impossible to act on.
  • Enterprise intelligence platforms (Stratfor, Recorded Future, Bloomberg Terminal, Janes) — accurate, but locked behind $25,000–$100,000+/year contracts.

We asked a simple question: what if a single person could access institutional-grade, AI-verified intelligence for the price of a Netflix subscription — and actually see the platform's track record before trusting it?

That's Sarniq.

What it does

Sarniq ingests global news 24/7 and runs every signal through a 4-agent AI pipeline:

  1. Pre-flight credibility check (Gemini Flash) — blocks satire and likely-fake content before it ever reaches analysis.
  2. Geo Agent — geographic, demographic and infrastructure context.
  3. Narrative Agent — media framing, bias detection, propaganda patterns.
  4. Risk Agent — escalation probability, market impact, affected sectors.
  5. Synthesis Agent — produces the final decision-ready brief.

Then — and this is the part nobody else does publicly — every signal is verified 48–72h later against ground truth and graded confirmed / partial / incorrect. The results are published live at sarniq.com/track-record.

Current public track record: 345 graded signals · 98.4% weighted accuracy · 33 self-learned lessons re-injected into the agents.

What we learned

  • Self-learning beats bigger models. Our 33 lessons (extracted from verification outcomes and re-injected as system prompts) improved accuracy more than swapping to a larger LLM ever did.
  • Transparency is a feature, not a liability. Publishing every wrong call publicly built more trust in 8 weeks than any marketing copy could.
  • Cost discipline is a moat. By routing 80% of work to Gemini Flash and only escalating hard cases to GPT-5/Gemini Pro, we run the full pipeline at ~$0.003 per signal — viable at $9.99/month pricing.
  • Prosumers were the real underserved market. Every "enterprise-only" assumption we challenged opened a new segment.

How we built it

  • Frontend: React 18 + Vite + TypeScript + Tailwind + shadcn/ui, deployed on Lovable.
  • Backend: Lovable Cloud (Supabase) — Postgres with RLS, Auth, Storage, Realtime, Edge Functions (Deno).
  • AI orchestration: Lovable AI Gateway routing across Gemini 2.5 Flash / Pro, Gemini 3.x preview, GPT-5 / 5.4 / 5.5 depending on task complexity and cost budget.
  • Payments: Stripe (5 tiers: Free, Starter $9.99, Pro $29, Elite $79, Enterprise).
  • Pipeline: scheduled Edge Functions for ingest-news, analyze, verify-signals, learn-from-verifications, weekly-digest.
  • Self-learning loop: verification outcomes → lesson extraction (GPT-5) → versioned agent_lessons table → injected into every future analysis prompt.
  • Observability: per-signal token & cost tracking, admin cost-control dashboard, daily/weekly budget caps.

Challenges we ran into

  • Hallucination on breaking news — solved with the pre-flight credibility gate + mandatory source citations + post-hoc verification.
  • Cost runaway at scale — solved with model-tier routing, aggressive caching, and a hard daily spend cap enforced at the Edge Function layer.
  • RLS policy complexity with a free/paid/admin role matrix — solved with a has_role() security-definer function and a dedicated user_roles table.
  • Proving accuracy publicly — required building an entire verification subsystem (a product inside the product) before we could claim a single number.

Track record (live, public)

  • 1,380+ signals processed
  • 345 verified against ground truth
  • 98.4% weighted accuracy
  • 33 lessons re-injected into agents
  • 8 weeks from prototype to paying customers

Try it

👉 sarniq.com — free tier, no credit card. 👉 sarniq.com/track-record — every prediction we got right and wrong.

Built With

  • deno
  • edge-functions
  • framer-motion
  • google-gemini
  • gpt-5
  • i18next
  • lovable
  • lovable-ai-gateway
  • lovable-cloud
  • openai
  • postgresql
  • react
  • react-query
  • recharts
  • resend
  • shadcn-ui
  • stripe
  • supabase
  • tailwindcss
  • typescript
  • vite
  • zod
Share this project:

Updates

Submission history