Nithya Shree
Kannan

The problem: Multi-step tasks collapse when one agent fails and there's no shared memory or recovery path.

Built graph-based workflows with tool integration, persistent memory, and ReAct reasoning loops — designed so agents can fail, recover, and hand off state without losing context.

Thinking: I started with failure modes, not features. What does the system do when an agent times out mid-task? That question shaped the whole architecture.

The problem: Swapping models without a structured eval means you're guessing whether quality improved or regressed.

Built a concurrent benchmarking framework across model providers with configurable scoring, regression testing, and CI/CD-compatible outputs — so model changes become data-driven decisions.

Thinking: If you can't measure it, you can't improve it. I built the eval infra before optimizing anything else.

The problem: Users don't tolerate slow or inconsistent AI responses in consumer apps — one bad output loses trust.

Designed a multi-model strategy across 3 endpoints with intelligent fallback and retry handling, keeping extraction under 15 seconds and output quality consistent across edge cases.

Thinking: I treated latency and reliability as UX constraints, not engineering afterthoughts. The model routing logic was designed around user tolerance, not API cost alone.

The problem: Clinical reasoning needs to be explainable and auditable — a black-box model isn't acceptable in diagnostic contexts.

Built a rule-based clinical agent that formalizes diagnostic logic into auditable workflows, with an evaluation framework to catch misclassifications and a front-end non-technical users could operate.

Thinking: Interpretability wasn't a constraint — it was the design goal. I chose Prolog precisely because every inference step can be traced and explained.

The problem: Small grocery retailers lose sales to poor UX and inventory mismatches that erode customer trust.

Built a full-stack e-commerce platform with login, categorized browsing, and cart — backed by fault-tolerant storage and SQL tracking that delivered 40% reliability improvement and real-time inventory accuracy.

Thinking: I designed the data layer before the UI. Reliability at the storage level meant the front-end could be simple and fast without defensive coding everywhere.

The problem: Stock prediction pipelines fail at scale when batch processing can't keep pace with live market volatility.

Built a Spark Streaming pipeline for real-time data ingestion feeding RNN/LSTM training — reducing RMSE to 0.05 and boosting prediction accuracy by 35% on live market data.

Thinking: The model architecture was the easy part. The hard part was making the data pipeline fast and reliable enough that the model was always training on reality, not stale history.