I build production AI systems β not demos. My work centres on the infrastructure around models: inference gateways, retrieval, evaluation gates, and observable multi-agent systems.
| π’ Currently | π― Open to |
|---|---|
| AI Platform Engineer β SRE, Model Infrastructure Deloitte Consulting India Β· Oct 2024 β Present Β· Hyderabad |
AI Engineer Β· LLM Engineer Β· Applied AI Β· Agentic AI Β· AI Platform Remote-first Β· India Β· US Β· UK Β· Europe |
| 12+ GenAI microservices | ~50K requests/day | 99.9% uptime | P95 1.8s β 1.25s (~30% faster) | ~18% inference-spend cut |
|---|
I contribute to the libraries AI teams actually depend on β observability, memory, and MCP security. Only submitted PRs are listed, and merged vs. open is kept explicit.
| PR | What it fixed |
|---|---|
| mcp-drift-check #17 β merged 28 Sep 2026 | Zero-execution uvx package-reference classification (PEP 508): pinned β SAFE, ranges β MEDIUM, bare β HIGH; URLs / paths / wheels β REVIEW. 5 files, +239/β5, 58/58 tests. |
| mem0 #7506 β merged 7 Oct 2026 | Stopped a service-account private key reaching DEBUG logs in GoogleMatchingEngine β constructors log field names, not values. Reproducer-first regression test; 8/8 module tests. |
One coherent thread: unended OpenTelemetry spans on streaming and interrupted LLM calls β observability engineering inside the tracing layer.
Click to see all 9 open PRs
| PR | What it does |
|---|---|
| openllmetry #4534 | OpenAI: end chat span on ChatStream.close() (sync + async) |
| openllmetry #4539 | Anthropic: end streaming spans on early exit (close, with-block, GC) |
| openllmetry #4544 | OpenAI + Anthropic: materialise one-shot iterable messages / tools before recording |
| openllmetry #4559 | Bedrock: end streaming spans on early exit (break, close, dropped stream) |
| openllmetry #4567 | Google GenAI: end generate_content_stream span on early exit β maintainer-reviewed |
| openllmetry #4586 | OpenAI: record cache-read tokens when prompt_tokens_details is None β changes requested |
| openllmetry #4600 | LiteLLM: end span when a call is interrupted (KeyboardInterrupt / cancellation) |
| openllmetry #4639 | OpenAI Agents: scope handoff-parent lookup to the handoff's trace (cross-trace bug) |
| haystack #13497 | Preserve non-string dict keys through pipeline snapshot serialisation (9 new tests) β open draft |
I also work in mem0, instructor, agno, and the MCP Python SDK β fixes are submitted only when they are crisp, reproduced, and gate-clear.
| Project | What it is |
|---|---|
| Self-Healing LLM Gateway | Reliability layer for LLM calls |
| Knowledge Graph RAG | Hybrid KG + vector RAG, benchmarked |
| LLM Evaluation Platform | LLM-as-judge evals, human calibration, CI gating |
| Multi-Agent Research Assistant | LangGraph planner / researcher / writer / supervisor |
| AI Cost Doctor | Per-customer AI unit economics / tenant P&L |
Reliability layer for LLM calls: circuit breakers, idempotent retries, provider failover, per-tenant cost tracking.
Python FastAPI Redis LLMOps
Hybrid knowledge-graph + vector RAG over real company data: entity extraction, multi-hop reasoning, retrieval routing, cited answers.
828 chunks Β· 982 entities Β· 36-question benchmark Β· hybrid vs. vector: parity, not a win β reported honestly.
Python LangChain LangGraph RAG Vector Search
LLM-as-judge evaluation with human calibration, agreement metrics, bias tests, regression detection, and CI gating.
30-item calibration Β· deterministic ΞΊ = 0.907 Β· Groq 120B ΞΊ = 0.984 Β· citations-stripped variant: relative mean citation-criterion score drops of β68.8% / β62.0% in the stub demo.
Python Streamlit LLM Evaluation CI/CD
LangGraph planner / researcher / writer / supervisor with durable Redis state, resumable runs, budgets, full execution traces, and cited reports.
Deterministic benchmark 5/5 Β· p50 0.06s Β· 100% citations. Real-provider benchmark blocked by provider rate limits β not claimed as completed.
Python LangGraph Redis Multi-Agent Systems
Per-customer AI unit economics / tenant P&L for LLM spend β cost diagnosis, margin-killer detection, and savings recommendations. Demo uses clearly labelled synthetic data; same calculations power real data.
TypeScript Next.js FastAPI PostgreSQL
AI / LLM: RAG Β· Agentic AI Β· Tool-calling agents Β· Prompt engineering Β· LLM evaluation Β· Guardrails Β· Structured outputs
Infra / platform: OpenSearch Β· Vector databases Β· SRE / observability
Certifications: AWS ML Engineer β Associate Β· NVIDIA Agentic AI Professional Β· Databricks GenAI Associate Β· Claude Architect β Foundations
Hyderabad, India β open to remote worldwide
If you're building serious AI products β especially around LLMs, RAG, agents, or AI infrastructure β I'd love to connect.