Independent engineer for AI systems
that have to hold up in production.
I design agents, context engines, and product workflows with one rule: they must be dependable in the real world.
OPERATOR // J.KUMAR
DEPLOY / MEASURE / ITERATE
OpenAI, Anthropic, DeepSeek, Gemini with project-specific evaluation frameworks.
Advanced prompting, context management, RAG, and fine-tuning strategies.
500+ tool calls with 90% accuracy via rigorous context engineering.
Where I shipped
Software Engineer (AI / Full-stack)
Software Developer Intern
Full Stack Engineer (Remote)
R&D Intern
Selected builds · 2026
A TypeScript project exploring token receipts for agent workflows — making the cost and footprint of autonomous work easier to inspect.
A recent TypeScript build for live, interactive presentation experiences.
A TypeScript project exploring the connection layer between agents, tools, and product workflows.
Voice-first AI agent for learning Spanish through natural conversation — real-time STT, LLM orchestration, TTS, 5 teaching modes, and persistent sessions.
An AI critic that takes learning context, identifies weak points, and gives direct direction on what to improve next.
AI chatbot for a health insurance company with a custom knowledge-base, end-to-end ticket handling with real-time email notifications, and web-search grounded to admin-curated domains.
More systems, prototypes, and technical experiments live in the project archive.
Browse project archiveToolbox
My 5-tier GenAI build approach
Route across OpenAI, Anthropic, Gemini, and DeepSeek with unified interfaces, automatic retries, and graceful fallbacks — no glue code, no provider lock-in.
Shape what the model sees: structured payloads, trust tiers, freshness policies, and constraint schemas that make output predictable before touching a single prompt word.
Precision retrieval with intent-aware queries, metadata filtering, and source-grounded output — tools with typed contracts and bounded retries so failures stay explicit.
When prompting and retrieval hit their ceiling, fine-tune on domain-specific data to push accuracy, consistency, and latency beyond what off-the-shelf models can deliver.
Long-running agents with structured plans, 500+ tool call budgets, and eval frameworks that measure real outcomes — not just whether the output looks good.
Writing
How I design evals, fallbacks, and release guardrails for systems that survive real traffic.
02The practical method I follow to shape inputs, retrieval, and constraints so outputs stay stable.
Let's build
Let's explore ideas
Always up for interesting problems, early-stage ideas, and collaborations worth building.
Whether it's a half-formed idea or something almost ready to ship — I'm happy to think through it together.
Contact