Qalana · Architecture Review Tool

◆

Upload the architecture PDF. Claude Vision reads every page including diagrams, tables and component charts — not just text. Gap analysis runs against the full Blueprint v3.5. All versions are stored permanently for the team to download.

Upload Architecture PDF

PDF Only

📄

Drop your architecture PDF here

Claude Vision will read every page — text, diagrams, tables, component charts

PDF only · Max 20 pages recommended

Reading PDF with Claude Vision…

Uploading document…

Gap Analysis Results

Live Score

Chat with QArchReviewer

AI-Powered

🤖

QArchReviewer knows every Blueprint v3.5 constraint. Answer its questions, challenge its findings, and keep going until the architecture is 100% flawless.

QArchReviewer

Blueprint v3.5 · 360 Stories v6

Team Chat

Live · Real-Time

💬

Sharat, Architect and PM — discuss gaps, share decisions, align on the architecture in real-time. Messages sync across all browsers via Cloudflare KV.

Architecture Team

Not joined

Messages will appear here once you join and start chatting.

Connecting…

Document Versions

Version History

📁

Every uploaded PDF is stored permanently. Download any version, view its gap score, and track progress from v1 to GOAT level. New uploads auto-increment the version number.

No documents uploaded yet.

Locked Constraints

Non-Negotiable

Every constraint below is final. Design around these — they exist for compliance, patent or architectural integrity reasons.

Constraint	Decision	Reason Locked
Agent Runtime	LangGraph — Python TypedDict AgentState, conditional edges only	State machine integrity, checkpoint isolation, compliance traceability
Backend Language	Python + FastAPI exclusively. Node.js excluded.	Unified runtime — split stack breaks LangGraph state sharing
Redis Persistence	AOF + RDB disabled. Zero-disk mandate.	GDPR pre-consent — unconsented data must be ephemeral. Circuit breaker on TTL fail.
KMS Strategy	One CMK per candidate	Right to erasure = CMK deletion → all encrypted data permanently inaccessible
Cloud	AWS permanent. No multi-cloud.	Architectural decision — not cost
Multi-Region	Separate AWS accounts: EU / India / US / Central Compliance	Data residency at account level — VPC/IAM policies alone insufficient
Session Key Pattern	session:{tenant_id}:{job_id}:{candidate_hash}	Deterministic compound keys — cross-tenant leakage structurally impossible
LLM — Interview Agent	Claude Opus via direct Claude API (not Bedrock)	Compliance-critical reasoning + QRecruiter five-dimension intent labelling
LLM — All Other Agents	Claude Haiku via Amazon Bedrock	85–90% gross margin target at Phase 2 scale
LLM Optimisation	Prompt caching + Batch API mandatory from MVP	Pre-requisite to margin target — not optional
Sourcing Stack	PDL → Bright Data → RuneGrid (sequence locked)	SerpAPI rejected — legal risk for LinkedIn scraping
Divergence Engine™	AIF360, synchronous in Screening Agent — never async	Bias gate must block pipeline — async lets biased decisions pass through
QLDB	SHA-256 hash chain, append-only audit ledger	Tamper-proof compliance record. Smart Friction™ gate-pass issued here.

Proprietary Layer

Patent-Backed

◆

Qalana's moat. Every concept has a specific architectural footprint. Do not substitute with off-the-shelf equivalents.

Smart Friction™

Human Oversight Enforcement Gate

FastAPI Dependency Injection middleware. Every candidate-affecting agent action requires a QLDB gate-pass proof token. Without it: HTTP 403. Architectural enforcement — not a UI feature.

Requirement Clustering Engine · Patent P8

JD Intelligence Layer

Discrete, independently deployable service — NOT embedded in the Screening Lambda. Clusters JD requirements into weighted semantic groups. Output feeds Screening Agent and HPCL/FPCL learning engines.

QRecruiter · Patent P9

Native AI Interview Engine

Replaces Retell.ai at Phase 2. Five-dimension QRecruiter Intent Protocol. Runs on Interview Agent ECS Fargate via WebRTC DTLS-SRTP. Interview Agent must NOT couple to Retell.ai APIs from day one.

Divergence Engine™

Bias Detection + Fairness Gate

AIF360 integration running synchronously in the Screening Agent. Detects statistical divergence across protected groups. All output audit-logged to QLDB before any shortlist decision is emitted.

HPCL / FPCL Learning Engines

Continuous Outcome Learning

High-Performance Candidate Learning (hire outcomes) + Failure-Pattern Candidate Learning (rejection outcomes). Must persist to PostgreSQL — the only layer that grows continuously. Cannot be ephemeral.

Joiner Likelihood Score + Profile Twinning

Predictive Offer Acceptance Signal

Composite offer-acceptance predictor from Screening Agent. Profile Twinning matches against historical hire profiles. QLDB audit-logged — EU AI Act decision-influencing signal scope.

Phase Transitions

MVP → Phase 2

▲

Design every transition as a clean swap from day one. Adapter interfaces are mandatory at MVP for all three components below.

Component	MVP (Now)	Phase 2 (Target)	Design Requirement
ATS Integration	Merge.dev — free tier, field whitelist	Native ATS Connector	Abstract behind adapter interface from MVP. Connector must be swappable without agent changes.
Interview Engine	Retell.ai — interim only	QRecruiter — Patent P9	Interview Agent must NOT call Retell APIs directly. Interface layer required from MVP.
Background Check	Certn — primary provider	Certn + IDfy + NSC	RuneGrid abstracts all providers via adapter pattern. Adding providers must require zero agent-layer changes.