Building reliable AI systems for real users, tools, and production constraints.
I work across AI product architecture, backend systems, tool integrations, evaluations, deployment, and product iteration.
- Co-founder of Verly AI, building voice and chatbot engines with MCP integrations.
- Founder of Saient, a multi-channel AI sales platform across voice, WhatsApp, email, and web.
- Previously an AI Engineering Intern at Nester Labs, working on voice-agent evaluation, AI security, and production conversational systems.
- Electronics & Communication Engineering at IIIT Nagpur.
| Product | Focus | Explore |
|---|---|---|
| Verly AI | Voice agents, chatbots, product architecture, and MCP integrations. | Website ↗ |
| Saient | Multi-channel AI sales agents across voice, messaging, email, and web. | Website ↗ |
| nForge | Built at Nester Labs to test voice agents with simulated callers, adversarial probes, and regression evaluations. | Live ↗ |
| nGuard | Built at Nester Labs for AI security, agent red-teaming, and adversarial testing. | Live ↗ |
| Production Voice Agent | Built at Nester Labs for natural, real-time customer conversations. | Live demo ↗ |
| Intake & Booking Agent | Built at Nester Labs to automate customer intake and appointment booking through voice. | Live demo ↗ |
Contributions merged across observability, AI infrastructure, developer tooling, and model-training projects.
| Project | Contribution | Pull requests |
|---|---|---|
| OpenTelemetry JS | Fixed optimized XPath propagation in browser trace instrumentation. | #6335 |
| Soup | 8 merged PRs across live canary routing and rollback, efficient outcome persistence, notebook migration, config diagnostics, dataset validation, and training recipes. | #1791 · #1712 · #1711 · #1104 · #1015 · #933 · #860 · #858 |
| Orval | Added JSON5 package-manifest support and pinned its parser dependency. | #4080 · #4081 |
| Repowise | Kept generation progress within the valid 0–100% range. | #2224 |
| Agent Swarm | Added tests that keep provider credentials, pricing, and Docker registration synchronized. | #1428 |
| Draw.io Skill | Added an AsyncAPI importer for editable architecture diagrams. | #127 |
Reproduce the failure through the real execution path, fix the responsible layer, and add a regression that fails when the fix is removed. Keep the change focused, explain its limits, and follow it through review and CI.
| Problem | Approach & result | Merged work |
|---|---|---|
| A rollout policy did not control live traffic. | Wired request routing and rollback into the serving path, then cached policy reads and batched outcome writes while preserving outage and rollback safeguards. | #1104 · #1712 |
| Notebook migration picked the wrong training config. | Followed the final trainer's actual arguments instead of the last config call. Task, output and hyperparameters now stay consistent even with warm-up stages or unused configs. | #1711 |
| Config diagnostics warned about harmless defaults. | Read defaults from the Pydantic schema, preserved valid zero-valued seeds, and removed unreachable warnings. Full YAML dumps now agree with minimal configs; a mutation-tested regression pins the rule. | #1791 |
| Invalid data and recipe assumptions reached users. | Made validation fail when no rows are usable, fixed terminal rendering, and added DPO/GRPO recipes with public model IDs checked through the library's resolver. | #858 · #860 · #933 · #1015 |
 OpenTelemetry |
 Orval Labs |
 Repowise |
 Desplega AI |
 Agents365 AI |
| Project | What it does | Links |
|---|---|---|
| Clawspan | Voice-first personal AI chief of staff coordinating specialized agents across developer and productivity tools. | Code · Demo |
| Lookalike Search | Reasoning and reranking system for difficult entity-search cases, backed by reproducible evaluation runs. | Code · Breakdown |
| Manus Decoded | Technical analysis of agent architecture, context isolation, parallel execution, memory, caching, and failure handling. | Code · Breakdown |
Technical stack
| Area | Technologies |
|---|---|
| AI agents | LLM orchestration, MCP, tool calling, RAG, agent handoffs, safety boundaries |
| Evaluation | LLM-as-judge, adversarial testing, regression suites, reproducible benchmarks |
| Voice AI | Pipecat, Deepgram, Cartesia, OpenAI Realtime, Twilio Media Streams, Silero VAD |
| Backend | Python, TypeScript, FastAPI, Django, Node.js, PostgreSQL, MongoDB, WebSockets, SSE |
| Infrastructure | AWS, Docker, GitHub Actions, OpenTelemetry, Langfuse, Prometheus, Qdrant, Neo4j |
| Highlight | Details | Explore |
|---|---|---|
| Multi-Agent Orchestration | Built an eight-agent system connected to GitHub, Gmail, Slack, PostgreSQL, and other tools. | Watch demo ↗ |
| Featured in Inc42 | OpenClaw work featured in The New Garage: OpenClaw And India's DIY AI Agent Boom. | View feature ↗ |
| Voice AI Engineering | Wrote about engineering a more natural production voice stack. | Read article ↗ |
| Competitive Programming | Ranked 8th in the Unstop Coding Challenge and 1861 globally in LeetCode Biweekly Contest 153. | LeetCode profile ↗ |
Building reliable AI systems that work with real tools, users, and latency constraints.
Start a conversation on LinkedIn · Send an email
February 1–12, 2026