AI Solution Architect · Agentic AI · Reinforcement Learning · Production AI Systems
I architect AI systems that have to do more than return plausible text. I work on systems that retrieve evidence, reason through decisions, use tools, coordinate workflows, verify outcomes, learn from feedback, recover from failure, and remain observable and controllable in production.
I bring 16+ years of software engineering and architecture experience across Amdocs, Tech Mahindra, startups, and now Kognitiv Inc. My work has spanned Australia, Singapore, Europe, Brazil, and India, including systems serving more than 15 million users worldwide.
Agentic and adaptive systems
Agent runtimes, planning and tool use, memory, multi-step and long-running workflows, human approvals, multi-agent coordination, failure recovery, and reviewable learning from real outcomes.
Retrieval and data intelligence
Enterprise search and RAG, hybrid lexical and semantic retrieval, reranking, knowledge systems, natural-language data interfaces, time-series intelligence, anomaly detection, and recommendation systems.
Learning and model intelligence
Reinforcement learning, post-training, distillation, reward and rubric design, outcome-based evaluation, model comparison, and dynamic routing across models, providers, tools, retrieval strategies, and reasoning effort.
Secure production AI platforms
Sandboxed execution, least-privilege tool access, policy enforcement, secret and data protection, multi-tenancy, observability, regression testing, deployment infrastructure, auditability, and cost control.
Multimodal products
Audio, video, vision, content pipelines, native applications, web experiences, APIs, and the infrastructure required to turn an AI capability into a complete product.
These outcomes come from different delivery and evaluation scopes:
| Dimension | Outcome |
|---|---|
| Global reach | Systems serving more than 15 million users worldwide |
| Enterprise search scale | Approximately 3.4 million indexed documents and 7,000 peak concurrent users |
| Search performance | Approximately 986 ms p95 end-to-end latency and 99.8% Precision@1 |
| AI evaluation | 98.3% accuracy across defined benchmark and evaluation suites |
| Delivery reliability | Zero reported production defects within the delivered scope |
| Operational impact | Approximately 35% reduction in support average handling time through enterprise RAG |
- Treat evaluation, telemetry, regression testing, and failure analysis as product capabilities
- Measure whether the system completed the task, not whether its response sounded convincing
- Keep autonomous behavior bounded by permissions, policies, budgets, and human control
- Let models, tools, and retrieval methods form a dynamic compute fleet instead of hardcoded dependencies
- Design feedback loops that improve performance without allowing uncontrolled behavioral drift
- Build end to end, from models and runtimes to APIs, interfaces, operations, and recovery paths
| Area | Technologies and methods |
|---|---|
| Languages | Python, Rust, Java, TypeScript, Swift, SQL |
| Learning and post-training | PyTorch, Transformers, TRL, Unsloth, CUDA, LoRA, QLoRA, SFT, DPO, GRPO, RLVR |
| Agent systems | LangGraph, LangChain, LlamaIndex, MCP, tool calling, structured outputs |
| Retrieval | Qdrant, Milvus, FAISS, BM25, ColBERT, SPLADE, hybrid retrieval, reranking |
| Serving and infrastructure | FastAPI, Flask, Node.js, Docker, Linux, AWS, GCP, vLLM, Ollama |
| Models and providers | OpenAI, Anthropic, Qwen, Llama, OpenRouter |
- How can production agents learn from successes and failures without unsafe drift?
- How should a system choose its model, reasoning effort, tools, and retrieval strategy for each request?
- How do we evaluate an agent's full trajectory rather than only its final response?
- How can agents use memory and tools without receiving unnecessary privileges?
- When should a system simulate alternative actions before committing to one?
- How do we make increasingly autonomous systems reliable and secure enough for enterprise use?
I spend most of my time at the intersection of:
Agents × Reinforcement Learning × Routing × Retrieval × Evaluation × Security × Simulation
The operating loop I find most compelling is:
observe → reason → retrieve → simulate → act → verify → learn → adapt
I enjoy building AI systems that do not just generate answers, but can make decisions, use tools, evaluate outcomes, learn from experience, and operate reliably in production.
If my work helps your research or products, consider supporting independent research:
