Experience
Xyenta Solution
Software Developer (Founding Engineer), AI Platform & Systems
– 1 yr 5 mos
Founding engineer on the xflow.solutions AI platform: agent harness, compiler, backend services, cloud infrastructure and compliance.
Measured results
- about 3,000 times
- faster process builds: 15 min to under 100 ms
- 84 percent
- prompt-cache hit rate on OpenAI
- 40 percent lower
- agent failure rate with the unified tool layer
- from 2 hours to under 5 minutes
- Kubernetes cluster provisioning
- 3 times
- release velocity with self-hosted CI runners
- 99.9 percent
- production uptime
Headline work
Agent harness built from scratch
Built the agent harness for xflow.solutions from the ground up on the native Anthropic and OpenAI SDKs. Agent frameworks are excellent for prototyping; owning the requirements and the intent let us keep the agent loop, tool execution, and structured outputs fully under our own control and tuned for production.
Architecture note: Own the agent loopAI canvas execution engine and compiler
Architected the custom formula compiler and AST execution engine for xflow.solutions, translating business glossary definitions, SQL, and natural language into live DAG nodes, slashing process build time from 15 minutes to under 100 ms (~3,000× speedup).
Architecture note: Compile what has to be exactly right
AI platform
- Data semantics, data contracts and column-level lineage. Business data contracts mapped onto technical schemas, with end-to-end column lineage across multi-stage ETL.Mapped business-level data semantics and data contracts onto technical schemas, and designed an in-memory lineage state machine traversing the AST DAG, tracking end-to-end provenance and column mutations across multi-stage ETL operations with zero schema hallucination.
- MCP servers and unified agent tools. 10+ custom tools shared by the in-app agent and MCP clients; agent failure rates down 40%.Set up MCP servers with 10+ custom tools and built a unified tool layer, so the same agent tools work inside the app and over MCP; an abstract `ToolSpec` isolates UI side-effects from pure agent context and enforces sandboxed execution stages, reducing agent failure rates by 40%.Note: Define a tool once, run it anywhere
- SDK migration and determinism. DSPy replaced with the native SDKs: 35% fewer runtime dependencies and 99% execution accuracy.Migrated core agentic pipelines from DSPy to the native OpenAI and Anthropic SDKs (Claude 3.5 Sonnet/Haiku), debloating runtime dependencies by 35% and reaching 99% execution accuracy with deterministic Pydantic v2 structured schemas.Note: Own the agent loop
- Prompt caching economics and token metering. 84% prompt-cache hit rate on OpenAI, with token and cost metering per model, user and team.Structured static system-prompt prefixes and tool schemas for an 84% prompt-cache hit rate on OpenAI (68% across other providers); built multi-tenant token and cost metering per model, user, and team with automated executive cost reports.Note: Treat latency and cost as design inputs
Backend and infrastructure
- Observability: LLM tracing, metrics and dashboards. Phoenix, Prometheus and Grafana on Kubernetes; p90 tool-execution latency of 50 ms.Deployed Arize Phoenix, Prometheus, and Grafana via Helm onto Kubernetes, and added custom instrumentation across xflow services alongside OpenInference and OpenTelemetry spans, achieving p90 tool-execution latency of 50 ms and TTFT profiling across 100% of production LLM requests.Note: Treat latency and cost as design inputs
- Backend services and APIs in Python and Rust. Async FastAPI services with Rust on the hot paths; cached full-text search under 20 ms.Built async FastAPI services (SQLAlchemy, Alembic) with Rust (Axum, SQLx, Tokio) on the performance-critical paths, and designed the APIs to stay fast under load: async I/O end to end, connection pooling, query optimization, caching for hot reads, and cursor-based pagination with streamed responses; cached PostgreSQL full-text search serves at sub-20 ms retrieval latency.Note: Fast APIs are decided below the handler Note: Write the hot path in Rust
- Cloud infrastructure with Terraform, Kubernetes and VMs. Terraform-provisioned Kubernetes clusters and VMs, from 2 hours to under 5 minutes.Modularized provisioning for DigitalOcean Kubernetes clusters and Droplet VMs, cutting provisioning time from 2 hours to under 5 minutes; administered the VMs, authored Helm charts with Horizontal Pod Autoscaling, and managed clusters with Lens.
- CI/CD with self-hosted Kubernetes runners. Azure DevOps pipelines on self-hosted runners: release velocity tripled, zero-downtime deploys.Designed Azure DevOps pipelines with self-hosted agents inside private Kubernetes pods and Droplet VMs, tripling release velocity with zero-downtime rolling and canary deployments; optimized multi-stage Docker builds for faster builds and leaner images.
Security and leadership
- Security, RBAC and SOC 2 compliance. Automated SOC 2 compliance bench, 85%+ test coverage, Auth0 SSO and RBAC.Established an automated SOC 2 compliance bench with audit trails, vulnerability scanning (`pip-audit`, Trivy), Ruff, and strict Mypy, sustaining 85%+ test coverage; integrated Auth0 SSO/OAuth2 and administered DigitalOcean RBAC via doctl.
- Technical leadership and uptime. 99.9% production uptime behind Cloudflare; mentored 3+ junior engineers.Configured Cloudflare DNS, edge SSL, and routing to hold 99.9% production uptime; mentored 3+ junior engineers on system design, distributed data pipelines, and agentic workflows.
Backend Intern, Product Engineering
– 5 mos
Built the first backend services and AST formula translation for xflow.solutions.
- First backend services. Prototyped backend microservices for xflow.solutions with FastAPI and PostgreSQL, implementing the first AST formula translation routines that turn business definitions into live process nodes.
- Search and early LLM tools. Engineered cached full-text search indexing with sub-20 ms retrieval and built early LLM function-calling tools with Pydantic validation.
Data Analytics Intern
– 5 mos
Built and optimized enterprise ETL pipelines on Azure Data Factory and MSSQL.
- Enterprise ETL pipelines. Designed and maintained automated pipelines with Azure Data Factory and MSSQL, implementing SCD Type 1 and Type 2 loads from staging into enterprise fact and dimension tables.
- Transformation optimization. Streamlined complex SQL transformations, cutting staging processing latency by 35% and validating integrity across cross-tenant data models.
National Informatics Centre (NIC), MeitY
Software Engineering Intern, AI & Computer Vision
– 2 mos
Built and deployed a face-recognition web service and bridged it into legacy Java applications.
- Computer-vision web service. Built and deployed a face recognition service with Python, FaceNet, and MTCNN, processing 10,000+ verification queries at 98.5% recognition accuracy.
- Cross-platform integration. Architected a bridge connecting Python inference routines into legacy Java enterprise applications at under 150 ms latency.