Skip to content

🗺️ Yojaka AI: Product Roadmap

This document outlines high-level implementation milestones and strategic release targets for Yojaka AI (Agentic Profile Matching Engine).

📚 Detailed Specifications: For system architecture diagrams, mathematical formulations, and formal design decisions, see System Architecture, Architecture Decision Records (ADRs), and Changelog.

📍 Implementation Milestones (Phases 1–19)

Foundational Pipeline & Core State Machine (Phases 1–7 • v0.1.0) ✅

  • Agentic Core: 9-node LangGraph StateGraph workflow with deterministic state transitions and MemorySaver checkpointing.
  • Tooling & Gateway: Dual-mode tool execution gateway supporting local in-process and FastMCP JSON-RPC stdio servers (ADR-001).
  • Candidate Vetting Tools: Schema-validated requirement extraction, side-by-side comparison matrix, and tailored interview question generation.

Enterprise Abstractions & Background Workers (Phase 8 • v0.8.0) ✅

  • Service Decoupling: Dedicated IngestionService isolating core business logic from transport protocols.
  • Pluggable Storage: BaseVectorStore structural protocol with ChromaDB default and Qdrant stub (ADR-003).
  • Async Task Queue: Distributed Celery workers with Redis 7 message broker and Docker Compose orchestration (ADR-006).

Observability, Hybrid Retrieval & Multi-Provider Scale (Phases 9–10 • v1.0.0) ✅

  • Enterprise Observability: Structured JSON logging and latency tracing with Langfuse and OpenTelemetry backends (ADR-007).
  • Hybrid Scoring: Min-max normalized dense vector cosine + BM25Okapi lexical retrieval with cached sparse index (ADR-004, ADR-008).
  • Multi-Provider LLM Gateway: Unified provider abstraction supporting Groq (Llama 3.3 70B), Google Gemini, Sarvam AI (105B Indic), and OpenAI (ADR-010).

Security, State Invariance & Async Screening (Phases 11–12 • v1.1.0) ✅

  • Stateless Credential Isolation: Credentials passed via RunnableConfig, eliminating secret exposure in checkpoints (CWE-312) (ADR-011).
  • Functional State Immutability: Copy-on-write candidate dictionary updates ensuring safe LangGraph retries (ADR-012).
  • Concurrent Screening: Parallel candidate screening with ThreadPoolExecutor and rate-limit semaphores (ADR-014).

Quality Gates & Engine Resilience (Phases 13–14 • v1.1.5) ✅

  • Automated Quality Gates: Strict mypy static typing, pytest-cov ≥ 75% coverage gate, and gitleaks/pip-audit security scanning.
  • Deterministic Cache Correctness: Global BM25 corpus fingerprinting and async subprocess lifecycle management.

In-Memory Ingestion, Modern Routing & UI (Phase 15 • v1.2.0) ✅

  • Zero-Disk In-Memory Ingestion: Stream-based file parsing (io.BytesIO) without disk persistence, layered with ephemeral session stores (ADR-016).
  • Dynamic Semantic Skill Expansion: LLM requirement expansion with word-boundary matching, replacing brittle static taxonomies (ADR-013).
  • LLM Intent Routing & Caching: Structured zero-shot intent router with dynamic anchor synthesis and sub-millisecond query caching (ADR-009).
  • UI Ergonomics: Compact zero-scroll recruiter sidebar with instant repository path auto-discovery.

Modular UI Architecture & Headless API Sidecar (Phase 16 • v1.3.0) ✅

  • Modular Presentation Layer: Monolithic app.py decomposed into an executive <75-LOC conductor and modular ui/ components.
  • Streamlit-Native Streaming: Real-time typewriter output via st.write_stream() and dynamic intent-aware execution tracking.
  • Headless FastAPI Gateway: Additive ASGI sidecar exposing REST and Server-Sent Events (SSE) streaming endpoints.

Layout Parsing, Contextual Retrieval & Two-Stage Reranking (Phase 17 • v1.4.0) ✅

  • Layout-Aware Section Parsing: PyMuPDF visual bounding block extraction preserving work experience integrity without boundary truncation.
  • Anthropic Contextual Retrieval Prepending: Document-level metadata banners prepended to chunks before dense and BM25 embedding.
  • Two-Stage Hybrid Reranking: Fine-grained Cross-Encoder joint reranking (ms-marco-MiniLM-L-6-v2) with Sigmoid calibration and Reciprocal Rank Fusion.
  • Serverless Cloud Resilience: Zero-disk pure NumPy InMemoryVectorStore fallback ensuring 100% crash-free ephemeral execution.

Calibrated Margin Routing, Native Commands & Dual-Rubric Subgraphs (Phase 18 • v1.5.0) ✅

  • Calibrated Margin Routing: Mathematical embedding margin scoring (\(\Delta \ge 0.12\)) routing confident queries in < 2ms at $0 LLM cost while escalating ambiguity to structured LLMs (ADR-018).
  • LangGraph Native Commands: Refactored workflow nodes to native LangGraph Command(goto=..., update={...}) primitives with destination routing, eliminating conditional edge boilerplate.
  • Parallel Dual-Rubric Structured Evaluation: Concurrent structured scoring (Technical Architecture Competence 60% + Talent Sourcing Fit 40%) with deterministic mathematical aggregation, eliminating slow debate loops while delivering committee-grade scorecards.
  • Modular Typed Subgraphs: Pipeline decomposed into isolated, independently testable subgraphs (JDAnalyzerSubgraph, TalentRetrievalSubgraph, DeepScreeningSubgraph, SynthesisSubgraph).

Advanced RAG Architecture & Air-Gapped Local Inference (Phase 19 • v1.6.0) ✅

  • HyDE (Hypothetical Document Embeddings): Query expansion synthesizing realistic candidate profile summaries from job descriptions before vector search to bridge vocabulary mismatch (ADR-019).
  • Hierarchical Parent-Document Chunking: Fine-grained 150–250 token child vector retrieval linked to 1,000–1,500 character parent sections in ParentDocumentStore, eliminating context fragmentation during screening.
  • Air-Gapped Local Inference: Turnkey offline model execution via Ollama and vLLM daemons for GDPR sovereign cloud compliance and zero cloud API costs.
  • Pre-Retrieval Faceted Filtering: Direct vector and lexical metadata constraints for experience years and education criteria with dynamic skill expansion support.

🚀 Future Milestones (Phases 20–22)

Phase 20: Resume Threat Security, Zero-Trust PII Redaction Vault & Multi-Tenancy (v1.7.0)

  • Indirect Prompt Injection Defense (OWASP LLM01): Input sanitization heuristics, canary detection, and strict XML boundary tags neutralizing hidden resume injection payloads.
  • Zero-Trust PII Tokenization Vault: Pre-screening entity tokenization replacing personal candidate identifiers ([CANDIDATE_A]) for EEOC blind hiring meritocracy.
  • Reversible Recruiter Vault: AES-256 encrypted candidate identity mapping revealed only upon authorized recruiter outreach with immutable audit logging.
  • Hardware-Partitioned Multi-Tenancy: Isolated collection namespaces across vector stores (tenant_{org_id}_resumes) and session checkpoints.

Phase 21: Unit Economics, Semantic Caching & Budget Circuit Breakers (v1.8.0)

  • Pool-Aware Semantic Evaluation Cache: Sub-5ms response caching for semantically equivalent recruiter queries, guarded by pool-version hashes to prevent stale candidate rankings.
  • Granular Cost & Token Budgeting: Predictive tiktoken calculation, real-time per-query and per-tenant cost accounting, daily spend caps, and runaway loop circuit breakers.
  • Job Description Inclusivity Scanner: Automated open-source linguistic analysis flagging exclusionary or demographic-biased wording and generating auditable competency-grounded rationale trails.

Phase 22: Production Infrastructure, Containerization & Continuous Evals (v2.0.0)

  • Multi-Stage Hardened Production Dockerfile: Minimal non-root container image (<250MB) isolating build dependencies for ASGI sidecar and workers.
  • Docker Compose Multi-Service Topology: Single-command orchestration for FastAPI sidecar, Streamlit UI, Celery worker, Redis 7, vector services, and self-hosted Langfuse.
  • Continuous Ragas & DeepEval CI/CD Quality Gates: Automated RAG Triad benchmarks (Context Precision \(\ge 0.88\), Recall@K, Faithfulness \(\ge 0.90\)) blocking pull requests on score regression.
  • Deep Health Probes & Rate Limiting: /healthz and /readyz probing database, Redis, and workers with slowapi rate-limiting shield.