Anshika Goel

Multi-agent systems that plan, call tools, and hand work between each other — with typed contracts at every seam so a failure is visible instead of silent.

• —Supervisor and planner/executor orchestration
• —Agent-to-agent protocols with typed messages
• —Plan-reason-act-reflect loops
• —Tool calling and API orchestration
• —Agent memory: session, user, long-term
• —Resumable workflows via state checkpointing

Retrieval that gets measured rather than assumed: hybrid search, reranking, citations traced to a source, and an evaluation gate standing between a change and production.

• —Hybrid dense + BM25 retrieval and fusion
• —Chunking strategy chosen by ablation, not habit
• —Citation-grounded generation
• —Per-tenant isolation and filtering
• —RAGAS faithfulness scoring in CI
• —Drift detection and regression monitoring

Turning messy PDFs, scans, and photos into structured records you can post against — with an accuracy number attached, not a promise.

• —Multi-engine OCR with cloud fallback
• —Open-set document classification
• —Schema-driven field extraction
• —Deterministic validation and repair loops
• —Confidence scoring and review routing
• —Accuracy measured against a golden set

The layer that decides whether a demo survives contact with real users: what you trace, what you alert on, and what you refuse to ship.

• —Tracing and telemetry (LangSmith, Prometheus, Grafana, Sentry)
• —Latency SLAs and KPI dashboards
• —Multi-provider gateways with automatic failover
• —Cost control through routing and caching
• —Evaluation gates wired into CI
• —Golden sets that grow from production corrections

Making the model's errors land in a check or a review queue instead of in the record — and keeping tenant data on the right side of the boundary.

• —Deterministic validation gates
• —Human-in-the-loop escalation paths
• —Refusal and fallback behaviour
• —PII redaction before the model call
• —Tenant isolation, RBAC, and audit trails
• —Random-audit sampling for silent failures

The unglamorous half that decides whether any of it holds: services, schemas, queues, deploys, and the CI gate that stops a bad merge.

• —Python microservices (async FastAPI)
• —REST APIs and gateway design
• —PostgreSQL, Redis, and vector stores
• —Async workers and job queues
• —Docker, Kubernetes, GitHub Actions CI/CD
• —Architecture documentation and standards