Tool Orchestration
Custom tool-calling orchestration using regex-based parsing rather than native function-calling APIs.
Back to home
Android / Kotlin / Jetpack Compose
AGENTIC AI ASSISTANT
A custom AI framework built from scratch to decompose hard problems into autonomous tasks, execute multi-step workflows, and verify outputs before delivering a final answer.
DataVault Storage
Externalized storage that reduces context pollution by keeping intermediate data outside the main conversation window.
Dual-LLM Validation
Automatic verification of completed work against the original query for error detection and iteration.
JavaScript Execution
Integrated execution environment for generated code, calculations, and runtime testing.
Configurable Runtime
Support for Gemini 2.0 and 2.5 Flash and Pro variants with adjustable token and timeout controls.
Iterative Workflows
Autonomous retries and refinement until the solution reaches a verified state.
Runtime controls
Model selection and settings are part of the product, not an afterthought.
The app lets users switch between Gemini variants, tune output limits, and shape response behavior based on the task complexity.
Workflow breakdown
A hard problem, broken apart on purpose.
The screenshots below show the system solving a complex CAT 2025 logical reasoning problem through a staged workflow.
01
User Query
A complex multi-constraint ticketing problem enters the system with enough ambiguity and branching to require structured decomposition.
02
Agentic Processing
The framework breaks the problem into smaller tasks, tracks progress, verifies intermediate work, and keeps moving until the solution stabilizes.
03
Export Options
Users can export either the final answer or the entire working conversation as a formatted PDF.
04
Formatted Output
The exported document preserves structure, mathematical notation, and shareable formatting instead of dumping raw text.
05
Exported Document
The PDF below is the actual exported output, embedded directly on the page.
06
Answer Verification
These screens compare the generated answers against the CAT 2025 answer key and mark each one correct.
