Problem

Job applications are a long chain of small, stateful tasks that break the moment a single-prompt assistant loses context.

Approach

Graph-based multi-agent orchestration with tool calling and structured memory, persisting state between agent steps so a run can be resumed and inspected.

Trade-off

Chose an explicit state graph over a single autonomous agent loop: more wiring and more code per capability, in exchange for runs that can be replayed and debugged step by step.

What broke / what I'd change

Long runs drifted when a tool returned an unexpected shape. Next pass: schema-validate every tool result at the graph edge and fail the node instead of letting the model improvise around it.

Problem

Answers lived inside thousands of internal documents, and keyword search kept returning the wrong page.

Approach

Chunking, embedding, retrieval and grounded answer generation with source citations, exposed as a REST API other services can call.

Trade-off

Kept retrieval to dense vector search with citations rather than a reranking stack: slightly weaker on ambiguous phrasing, but predictable latency and a much smaller surface to operate.

What broke / what I'd change

Fixed-size chunking split tables down the middle and produced confident wrong answers. Structure-aware chunking is the first thing I would rebuild.

Problem

Analysts waited on engineers for every ad-hoc query.

Approach

Schema-aware SQL generation from plain English inside Apache Superset, with validation before execution so nothing unsafe reaches the warehouse.

Trade-off

Validation rejects anything it cannot prove safe, so some legitimate queries get refused. Blocking a valid question is cheaper than letting a generated statement touch production data.

What broke / what I'd change

Generation quality tracked schema documentation more than model choice. I would invest in column descriptions and query examples before touching prompts again.

Problem

High-volume device telemetry arriving faster than batch jobs could absorb it.

Approach

Fault-tolerant streaming from Kafka into Spark for transformation and aggregation, with checkpointing and recovery on failure.

Trade-off

Micro-batching over per-event streaming: seconds of added latency, in return for throughput and recovery behaviour that survives a node dying mid-window.

What broke / what I'd change

Checkpoint state grew until restarts got slow. Retention and compaction belong in the design from day one, not in a later fix.