StellarumAtlas
AI Agentic Research Engine / Technical Specification
Drop a topic. Get a thesis-driven, citation-backed research essay —
written by an AI pit crew, not a chatbot.
-
7-Agent Pipeline — Planner, Researcher, Critic,
Thesis Generator, Decomposer, Writer, Titler — each with a dedicated role, working in
sequence like a turbocharged assembly line
-
Dual-Pass Research — first pass gathers raw
intelligence; an LLM-powered selector picks the 3 best-fit specialists from a roster
of 11 domain experts (economists, historians, geopolitical strategists, data scientists,
and more) to form a critic panel that identifies blind spots; second pass fills the gaps
-
Gap-Driven Research — the pipeline doesn't
just gather what's known; it actively identifies contradictions, unresolved debates,
and under-examined areas in existing scholarship. These gaps drive the thesis motivation
and frame the introduction — the same methodology a doctoral advisor would demand
-
Live Web Sources — Perplexity
sonar model pulls real-time, cited web data — no
stale training data, no hallucinated quotes
-
URL Validation — every source gets a HEAD request
before it touches the final draft; dead links get stripped automatically
-
Core Principle — AI handles thought
(argument structure, pattern recognition, synthesis); facts come exclusively
from verified external sources. The model never generates factual claims —
it organises, connects, and argues from a locked evidence base
-
Evidence Ledger — every research finding is
stored as an atomic
NoteChunk with
content, source URL, citation ID, and confidence tag — creating an auditable
chain of custody from source to final citation
-
4-Layer Verification — (1) Perplexity Sonar
retrieves live web data with source URLs, (2) HEAD-request validation strips
dead links, (3) dual-pass fact-checker audits numeric precision and named-entity
fabrication, (4) writer agent is constrained to citation IDs that exist in the
verified evidence pool — it cannot introduce unsourced claims
-
User-Verifiable Output — every bracketed
citation
[n] renders as a tooltip
showing the original source text and a clickable URL — the reader can verify
any claim in seconds. In manual spot-checking, cited figures, phrases, and
statistics trace back to their source with near-100% fidelity
-
Multi-Model Routing — Perplexity
sonar for research, Gemini 2.5 Flash Lite for
planning/critique, and a user-selectable writer model — choose from
Gemini 2.5 Flash Lite, Claude Haiku 4.5, Perplexity Sonar, Llama 4 Scout, or Codestral
2508 — each agent gets the right model for the job
-
Adjustable Concurrency — 1–10 parallel research
lanes; dial it up for speed or down for rate-limit safety
-
Exponential Backoff — automatic 429 retry with up
to 4 attempts; it doesn't stall, it paces itself
-
SSE Real-Time Telemetry — every pipeline stage
streams live to the client: planning → researching → critiquing → thesis → validating →
decomposing → writing → done
-
Academic — Intro → Body Arguments → Conclusion
(classic)
-
Narrative — Story arc with tension and resolution
-
Hybrid — Analytical rigor meets narrative flow
-
Magazine — Long-form feature: lede, nut graf,
thematic sections, kicker
-
Op-Ed — Sharp, declarative, conviction-driven
opinion
-
IB Literary (Comparative) — Paper 2
scaffold: comparative topic sentences, dual-text evidence with named devices,
authorial intention evaluation, anti-hallucination quote sourcing via Perplexity
-
IB Extended Essay — criterion-aligned
scaffold for the 4,000-word EE: intro maps to Criterion A (Focus & Method),
body sections enforce Criteria B–C (Knowledge & Understanding, Critical Thinking),
conclusion demands evaluation and limitations
-
Business Plan — investor-ready structure:
Executive Summary → Market Analysis → Business Model → Competitive Landscape →
Financial Projections → Investment Thesis, with founder intake and AI-generated
due-diligence questions
-
Rubric-Based Scoring — upload or paste an essay,
receive per-criterion scores grounded in official rubric descriptors with
evidence-cited justifications and actionable improvement suggestions
-
IB Extended Essay (Built-In) — Criteria A–D
(Focus & Method, Knowledge & Understanding, Critical Thinking, Presentation)
hardcoded from official IB assessment criteria, scored out of 28
-
4-Agent Parallel Evaluation — one dedicated
evaluator agent per criterion running concurrently, followed by a synthesis agent
producing overall feedback
-
9 Curation Pillars — Analytical lenses that bias
the entire pipeline: Failure Analysis, Cross-Domain Intelligence, Mechanism Mapping,
Strategic Foresight, Comparative Models, Legacy Thinking, Norm Disruption, Symbolic
Capital Lens, or Neutral
-
3 Citation Styles — Academic bracketed [1],
narrative inline, or minimal
-
3 Optional Sections — Bolt on an Analytical
Framework, Limitations & Counterarguments, or Case Studies
-
2–8 Body Sections — scale the argument depth
-
500–10,000 Word Target — with proportional budget
allocation (15% intro / 70% body / 15% conclusion)
-
AI Title Generation — post-hoc title with
intellectual tension, colon-separated punchline
-
20+ Authorial Voices — McKinsey, Economist,
Academic, Op-Ed, Narrative, and more — each voice profile steers tone, register,
and sentence rhythm across the entire essay
-
Quantitative Pressure
biases the entire pipeline toward data-rich output; injects a Quantitative Analyst
critic at ≥30%, forces numerical anchors into every body section, and steers
sub-question generation toward hard data
-
Fact Check Rigor — 0–100% slider driving a
dedicated fact-checker agent (not a prompt tweak). Pass 1 audits numeric claims —
classifying each as grounded, approximate (auto-hedged), or
unsupported (removed at high rigor). Pass 2 audits named entities —
legislation, regulatory programs, government initiatives — catching fabricated
names and overstated legal status
-
Firebase Firestore — every essay auto-saved with
full metadata: topic, thesis, outline, milestones, citations, voice, pillar, structure,
word count, generation duration
-
Per-User Auth —
x-user-id header gating; ownership-checked reads
-
Essay Archive API — list, retrieve, rename — full
CRUD for your research library
-
Citation Tooltips — each
[n] maps back to source content + URL for hover
previews
-
Per-Pipeline Usage Tracking — every run logs
token counts (input/output), model used, latency, and estimated cost per agent;
exposed in-app per run and stored in Firestore for aggregate analysis
-
Inline Paragraph Edit
the archive to edit its markdown directly;
⌘Enter to save, Esc to cancel — changes persist to Firestore
instantly
-
TypeScript (ES2022, strict mode) — fully typed
end-to-end
-
Express server — with static frontend serving
-
Smart Evidence Deduplication — notes used in one
section are excluded from subsequent sections to prevent repetition
-
Keyword Relevance Scoring — tokenized,
stop-word-filtered note retrieval per section
-
JSON Extraction with Retry — tolerates messy LLM
output; retries extraction up to 3 times before failing