Work Experience

Technologies & Tools

What I've done

• Building three internal AI systems on the AI Engineering team at Lenskart's Gurugram HQ: a hiring intelligence platform, a CEO intelligence brief, and a competitor creative-intelligence platform.
• Hiring intelligence platform. Built the candidate-discovery and durable-execution layers: multi-provider LLM sourcing across 4 providers with live web-search grounding and per-provider failure classification driving mid-search model fallback; a lease and heartbeat state machine over Postgres that reclaims a crashed worker's in-flight work; identity dedup that routes ambiguous matches to human adjudication instead of auto-merging; and a deterministic 4-key comparator for ranking in place of one opaque LLM score. 40 tables, 199 test files.
• CEO intelligence brief. Built it end to end: 40 source connectors across 16 signal categories writing through a single compliance-gated writer into a Postgres system of record spanning 149 migrations; a 9-tool MCP server backing 18 curation skills with per-stage, per-run, and per-month spend caps; and an evidence layer that admits a claim only on 1 primary or 3 or more independent origins above 0.75 match confidence, with database triggers freezing finalized issues.
• Competitor creative intelligence. Built it end to end: a 7-layer Python/FastAPI and Next.js pipeline tagging assets on 24 structured dimensions through the Gemini Batch API and rolling them into per-brand and cross-brand views; an agent layer running 6 switchable LLM providers behind one adapter with SSE-streamed traces; deduplication resolving one asset to one identity across 3 source platforms; and a deterministic numeric guardrail grounding every generated figure in tool evidence instead of an LLM judge.

Technologies & Tools

What I've done

• Built and shipped production AI agent systems on a multi-tenant agent platform: a FastAPI agent engine, a Django API layer, and a Next.js workflow editor.
• Cut cost and latency in the multi-agent orchestrator: rewrote the classifier and planner prompts (19.1K to 5.4K and 22.8K to 9.5K characters, roughly 3,300 tokens saved per request each), fixed a prompt-cache invalidation bug by moving query-dependent context below the cache boundary (23.6K tokens cached per request, a 35 to 42% cost saving), and parallelized context assembly that had been running as sequential awaits. Verified by 125 passing tests.
• Audited and fixed the agentic workflow engine across 3 repositories: 21 defects including roughly 50 endpoints returning success on 4xx and 5xx errors, collapsed 3 tool-router handshakes to 1 on linear runs by detecting linear-versus-DAG topology, removed a hard tool allowlist that had caused a 15-turn failure loop, and moved a 900-second blocking test-run call to a background thread with WebSocket delivery.
• Built a bot-free meeting-ingestion path that polls the Google Meet API on a schedule instead of running a paid recording bot, plus a vector re-key migration that re-scopes orphaned embeddings to their meeting without re-embedding, covered by 10 tests.
• Migrated the production memory layer across a breaking Mem0 v1 to v2 upgrade, fixing silently broken call sites where the API had changed sync behavior, prompt contracts, filter requirements, and method signatures.
• Ran code-grounded reliability and token-efficiency audits of the agent backend, surfacing a permissive CORS configuration on production hosts, 17 fire-and-forget async tasks with no error handling, a cross-user memory leak between agents, and verbatim tool output re-sent every turn as the largest remaining cost lever.

Technologies & Tools

What I've done

• Building across the ArmorIQ platform as a full-stack engineer, spanning backend APIs, the React console, AI-agent features, performance, and security.
• Built the observability system end to end: instrumented the SDK's core operations as nested trace spans, then built the ingest API, Postgres store, and query endpoints that persist and serve those traces.
• Hardened an LLM-powered policy-authoring feature with guardrails that reject fabricated tool references and sanitize untrusted input before it reaches the model.
• Shipped security and performance work on the backend, including tenant-isolation fixes on the policy engine and a redesign of API rate limiting into isolated per-route buckets.
• Built a documentation MCP server from scratch in Python (FastMCP) with API-key auth, per-IP rate limiting, and a CI-gated search-quality eval, deployed on Google Cloud Run.

Technologies & Tools

What I've done

• Co-authored a research paper on parameter-efficient LLM fine-tuning - arXiv:2601.00231
• Co-developed GRIT, a geometry-aware PEFT method using K-FAC preconditioning, Fisher-guided reprojection, and dynamic rank adaptation, that fine-tunes under 1% of a model's parameters while outperforming full fine-tuning and LoRA on standard benchmarks.
• Ran ablation studies comparing GRIT against LoRA, QLoRA, and AdaLoRA across multiple LLMs and standard NLP benchmarks.

Technologies & Tools

What I've done

• Autonomous Market Research Agent: Developed an agent that automates market and competitor research using real-time data collection, web search, and report generation for business insights.
• AI Voice Interviewer: Built and deployed a voice-based AI interviewer to conduct inbound calls, capture candidate responses, and generate structured interview summaries.
• Personalized Front Desk Agent: Created a multimodal front desk AI agent capable of handling text, voice, and video inquiries for customer support and client onboarding processes.
• Whatsapp and Social Media Automation Agent: Engineered agents that automate group administration, content posting, and engagement analysis on WhatsApp and other social platforms.
• Credit Risk Analysis Agent: Designed a credit risk agent using custom ML models to evaluate applicant default risk and streamline lending decisions with automated scoring and documentation.
• Outbound Calling Platform Agent: Deployed outbound calling agents capable of scheduling, executing, and logging calls for lead generation and customer feedback with seamless voice pipeline integration.