The chatbot runs inside Slack, so people get answers where they already work.

RAG over structured and unstructured data with FAISS and pgvector, Langfuse tracing, and an evaluation regression suite that gates every deployment on accuracy and latency. When retrieval finds nothing relevant, the assistant declines instead of guessing.

• Retrieval: RAG over structured and unstructured data (FAISS, pgvector)
• Quality: Evaluation regression suite gates deployments on accuracy and latency
• Observability: Langfuse tracing in production

A supervisor agent plans the work and directs worker agents, each isolated in its own git worktree, and ships through git and the GitHub CLI.

Each task has a token and cost budget that hard-pauses the worker at 100%. JSONL audit logs act as durable agent state, so every run can be inspected after the fact.

• Structure: One supervisor directs many workers through a single point of coordination
• Budgets: Per-task token and cost budgets hard-pause at 100%
• Isolation: Each worker runs in its own git worktree and branch

No runtime dependencies, and images never leave the machine. Output follows a semver-versioned JSON contract, and the CLI installs as an Agent Skill, so agents can call it as a local vision step.

• Cost: Local inference replaces billed multimodal tokens, cutting cost per request
• Privacy: Runs fully offline, so images never leave the machine
• Integration: Semver-versioned JSON contract; installs as an Agent Skill

Built approval workflows and the Google Calendar and Airtable integrations, and owned the product's data pipelines, including their monitoring.

• Workflows: Approval workflows for bookings and media assets
• Integrations: Google Calendar and Airtable
• Ownership: Data pipelines and their monitoring

Cloud-native automation used by 5+ teams: event-driven services on AWS, a workflow scheduler on Step Functions, and the analytics path from Kafka into Snowflake.

• Team: Led a 12-engineer Cloud Apps team; platform adopted by 5+ teams
• Scale: Workflow scheduler running 1,000+ workflows daily
• Architecture: Event-driven services on AWS, with analytics running Kafka into Snowflake

Built resumable uploads for large files captured over unreliable field connections, duplicate detection across a case, case workflow tasks, and API integrations with partner evidence platforms.

• Reliability: Resumable uploads survive dropped connections mid-transfer
• Storage: Duplicate detection avoids storing the same file twice
• Interoperability: Partner API integrations for exchanging evidence

Event photographers upload a shoot, and a facial-recognition pipeline sends each attendee only the photos they appear in.

Designed the AWS architecture behind it and owned the infrastructure cost alongside it.

• Scale: AWS architecture processing 1M+ images/day
• Business: Grew to ₹1 Cr (~$120K) revenue in two years

Uses on-device pose estimation with TensorFlow.js MoveNet, so the camera feed never leaves the device.

Built confidence-weighted pose matching that ignores low-certainty joints, an offline pipeline that generates reference poses from instructor video, and a compact binary pose format for fast loads on mobile.

• Privacy: Fully client-side, so the camera feed never leaves the device
• Accuracy: Confidence-weighted scoring that discounts low-certainty joints
• Performance: Custom binary pose format for a fast first load on mobile