I'm a Forward Deployed Engineer at CallKaro. I work with clients to take voice and chat AI bots from idea to working product: understanding their workflow, designing the conversation, testing real calls and chats, and fixing what breaks.
Before this I built backend systems in Python and FastAPI, including an ERP for a steel fabrication business and an open-source OT/ICS security platform. B.Tech student at KIIT, class of 2027.
I'm open to Forward Deployed Engineer and AI backend engineering roles.
- IndustryERP - ERP built for a steel fabrication plant and used by its staff: lot-level inventory, goods receipt and QA, drawing-based production tracking by stage, dispatch, scrap, accounting and print formats. About 190 API routes, six-role access control, versioned APIs, database migrations, and a CI pipeline (lint, tests with a coverage gate, dependency audit) with Azure deployment through GitHub Actions. FastAPI, PostgreSQL, Alembic.
- OneAlert - open-source security platform for OT/ICS networks. AI agents for detection, triage, threat hunting, response planning and purple-team validation, a compliance mapping engine (IEC 62443, NIST CSF), human approval gates for response actions, MITRE ATT&CK mapping, Suricata and Zeek ingestion, SBOM and CVE feeds (NVD, CISA KEV). FastAPI, React, PostgreSQL, Docker.
- Redline AI - team project: IVR platform for emergency dispatch (original repo: Ananya-Ghosh05/Redline-AI). Whisper speech-to-text, DistilBERT intent classification on ONNX, Celery workers, Prometheus metrics, Twilio telephony, WebSocket dashboard. My merged PRs there cover the backend, router and ML pipeline.
- SkyDash - open-source spatial intelligence dashboard: geospatial maps, entity graphs, missions, exports, simulated drone telemetry. CI and Azure Static Web Apps deployment. React, FastAPI, Leaflet, WebSockets.
- TaskForge - multi-agent crisis logistics coordinator built for Google Cloud Gen AI Academy APAC 2026, with CI and an Azure Container Apps demo. FastAPI, Gemini, MCP, PostgreSQL, pgvector.
- Cordis - open-source AI crisis coordination prototype: voice/text triage (Whisper, ONNX intent and emotion models) feeding a Gemini-based logistics planning pipeline, with a decision-support disclaimer. Includes adversarial and golden-scenario tests; CI is not currently green.
- VitalWatch - prototype for detecting patient events from live video: YOLOv8 detection, MediaPipe pose, a rule-based engine for falls, bed exit, immobility and abnormal movement, FastAPI/WebSocket alerts and a dashboard.
- SmartFlow Todo RN - React Native (TypeScript) task app with Firebase authentication, Firestore sync and a deadline-and-priority scoring sort.
- P2P on SAP BTP - procure-to-pay app on SAP CAP and UI5: requisitions, purchase orders, goods receipt, invoices and payments, with submit/approve/convert workflow actions.
- AI Ops Log Intelligence - Flask app that turns a plain-English question about ops logs into SQL (AlloyDB AI, with a local fallback), checks it is a safe SELECT, runs it with a row limit and shows the results.
- AutoGPT - two merged PRs: orchestrator fix forwarding credential input masks into tool-node execution (#13151) and a frontend test pinning the email rate-limit window (#15636)
- Flet - merged fix preserving effect cleanup callbacks across rerenders (#6983)
- Ruff - merged docs change on INP001 preview behavior (#29764)
- plone.scale - merged fix preserving the input image in contain mode (#169)
- PX4-Autopilot - merged docs fix for Docker networking in the simulation QuickStart (#29330)
- Biome - merged CLI fix for rdjson code-suggestion output (#10543)
- OpenTelemetry Collector Contrib - merged Datadog receiver feature-gate metadata migration (#48905)
- LosslessCut - co-author credit on the merged video preview scaling setting, which replaced my earlier proposal (#3134)
- AiSOC - proposed cross-tenant isolation tests and nightly RBAC regression CI (issue #159, closed); credited in the merged credits update (#223)
I start from the workflow, not the framework. For the ERP, that meant talking to the operators at a steel plant before writing the data model. I use agents, RAG or real-time streaming only when they make the system more useful than a normal API.
Python, FastAPI, SQLAlchemy, Pydantic, PostgreSQL, Redis, React, TypeScript, Docker, GitHub Actions, Azure, Whisper, ONNX, MCP.