AVRS
Adaptive Voice Rendering System. Three-tier audio routing (corpus → cache → live TTS) takes voice-agent latency from seconds down to single-digit milliseconds on warm turns.
Voice AITTSReal-time
View on GitHubResearch · Open source · Essays
Research and working systems focused on governable agents, reliable retrieval, and intelligent data infrastructure.
Flagship open-source work spanning voice AI, agentic infrastructure, and data-native runtimes.
Adaptive Voice Rendering System. Three-tier audio routing (corpus → cache → live TTS) takes voice-agent latency from seconds down to single-digit milliseconds on warm turns.
AI call screener for Android, built on AVRS. The AI picks up, asks who's calling and why, transcribes live, and gives you Join / Block / Reply actions before the call interrupts you.
Multi-agent debate reliability layer. Persona-driven reasoning with concurrent scoring across multiple models for quality, security, SOP compliance, and outcome.
Personal AI ecosystem with agent-to-agent (A2A) commerce: your personal agents collaborating, transacting, and acting on your behalf.
Declarative AI runtime for PostgreSQL. Define agents in SQL and bind them to tables, intelligence flows through your data automatically and transaction-safely.
Change-stream AI mutation runtime for MongoDB. Auto-classify tickets, summarize, extract entities, analyse sentiment, all driven by document changes. MIT licensed.
Production implementation of Anthropic's open Agent Skills standard. Progressive disclosure (3-level skill loading), self-evolving agents, and the WIF (Workflow Interchange Format) standard for portable deterministic state-machine workflows.
Sovereign hybrid-RAG appliance. Single-tenant Docker deployment with dual-strategy retrieval (BGE-M3 + BGE-reranker + PageIndex tree-walk), admin UI, multi-channel chat, auto-sync ingestion, and a navigable knowledge graph.
Privacy-first PDF & PPT chat for Android (संवाद, "dialogue"). Three modes: fully on-device (MiniLM ONNX + MediaPipe LLM), Claude with your key, or OpenAI with your key. Hybrid PDF parser with OCR fallback.
System of Record for AI agents. Stage-gated completion, multi-model LLM Council scoring, outbox pattern, async non-blocking verification, wallet metering with hard-stop, and post-run trace reconciliation. PyPI: agent-sor.
Living AI organism orchestration. Transforms agents into long-lived organisms with developmental stages, immune-system guardrails, and evolutionary fitness, defined declaratively via SOUL.md genomes.
India's first agentic payment platform. Mandate-based architecture for autonomous agent-driven commerce: auditable, revocable, bounded mandates. Launched at Global Fintech Fest 2025, Mumbai with CCAvenue.
Camera-first, voice-driven on-device AI companion built on Gemma 4: perception and action on the edge.
A living inventory of 40+ agentic-AI failure modes, the analysis underpinning the Determinism in Agentic Systems whitepaper.
Enterprise AI control plane. Universal MCP gateway with OPA/Rego policy engine, DLP/PII redaction, risk scoring, HITL approval workflows, tool-visibility filtering, audit trail, and a connected-account vault.
A thesis on pushing generative AI from probabilistic black boxes to auditable, bounded, repeatable automation enterprises can run mission-critical work on. Embodied in WIF, agent-sor, AVRS deterministic templating, and PayCentral's mandate-based payments.
Long-form essays on AI agents, retrieval, infrastructure, and the future of intelligent systems.
An in-depth analysis of how Anthropic's Agent Skills and OpenAI's adoption is reshaping the AI agent ecosystem. Exploring progressive disclosure architecture, the implementation gap, and what it means for developers building custom agents.
Why standard RAG fails 25 to 40% of the time, and how PolyRAG addresses one-size-fits-all retrieval, context collapse, multi-hop blindness, embedding redundancy, and memory fragmentation, benchmarking 99% retrieval success on 2,391 documents.
A declarative AI runtime inside PostgreSQL. Bind agents to tables in SQL and let intelligence flow through your database transactionally, every insert or update fans out to an agent and writes back safely.
A change-stream AI mutation runtime for MongoDB. A Clawbot army for every collection: watches changes, queues documents, runs enrichment via configurable providers, and writes results back. Open source under MIT.
Subscribe to my Substack to get notified when new research is published.
Subscribe to Newsletter