Aadesh Ingle, founding engineer building production AI systems (agents, RAG, voice, evals). Writing on shipping reliable AI, deep learning, and the craft of engineering. https://aadesh-ingle.github.io/ 2026-06-10T00:00:00.000Z Aadesh Ingle https://aadesh-ingle.github.io/blog/stop-being-the-thing-inside-the-loop 2026-06-10T00:00:00.000Z 2026-06-10T00:00:00.000Z

A production agent loop needs external checks, durable state, and hard limits. Without those, it is repeated prompting with a budget.

https://aadesh-ingle.github.io/blog/golden-cases-come-from-funerals 2026-06-07T00:00:00.000Z 2026-06-07T00:00:00.000Z

Our golden cases are anonymized production failures with assertions attached. Keeping the original thread, trace, and expected behavior has made them more useful than cases invented before launch.

https://aadesh-ingle.github.io/blog/the-model-is-the-easy-part 2026-06-02T00:00:00.000Z 2026-06-02T00:00:00.000Z

In production, the model call sits inside permissions, policy, retries, evals, and traces. Those surrounding systems decide whether a promising demo survives contact with a customer workflow.

https://aadesh-ingle.github.io/blog/the-first-two-weeks-are-for-listening 2026-05-17T00:00:00.000Z 2026-05-17T00:00:00.000Z

Before I build for a new client, I watch the people doing the work: their tabs, workarounds, retyping, and judgment calls. Those first two weeks often change what we build first.

https://aadesh-ingle.github.io/blog/evals-are-the-operating-system 2026-04-12T00:00:00.000Z 2026-04-12T00:00:00.000Z

A freight-negotiation agent had become too eager to close, and we could not measure the cost cleanly. An 87-test suite with 19 production cases gave us a release gate for the next model migration.

https://aadesh-ingle.github.io/blog/the-bug-was-in-the-transcript 2026-03-14T00:00:00.000Z 2026-03-14T00:00:00.000Z

A voice-agent complaint can originate in speech recognition, model decisions, tool execution, or turn timing. I read the transcript and event trace in that order before touching the prompt.

https://aadesh-ingle.github.io/blog/the-voice-agent-tradeoff 2026-02-08T00:00:00.000Z 2026-02-08T00:00:00.000Z

A recruitment voice agent sounded natural but kept missing tool calls. We replaced one fused speech model with a slower, inspectable pipeline and measured whether reliability justified the extra latency.

https://aadesh-ingle.github.io/blog/structured-demand-from-chaos 2026-01-17T00:00:00.000Z 2026-01-17T00:00:00.000Z

A logistics customer receives about 3,000 WhatsApp messages and 500 emails a day. We turn them into CRM demand records with a queue, an LLM extractor, entity resolution, deduplication, and a review lane.

https://aadesh-ingle.github.io/blog/where-the-human-belongs-in-the-loop 2025-11-08T00:00:00.000Z 2025-11-08T00:00:00.000Z

Reviewing every AI output wastes the attention human oversight is meant to provide. Route uncertain and irreversible cases to people, then keep their corrections as evaluation data.

https://aadesh-ingle.github.io/blog/writing-down-what-you-didnt-measure 2025-10-11T00:00:00.000Z 2025-10-11T00:00:00.000Z

I keep a small file of project claims beside the evidence for each one. It preserves the difference between a measured result, an estimate, and something we believe but have not tested.

https://aadesh-ingle.github.io/blog/demos-close-deals-boring-systems-keep-them 2025-09-13T00:00:00.000Z 2025-09-13T00:00:00.000Z

Ninety days after a successful AI demo, clients ask about failed batches, monthly cost, and recovery paths. Renewals depend on the answers.

https://aadesh-ingle.github.io/blog/what-actually-happens-when-you-call-an-llm-api 2025-08-09T00:00:00.000Z 2025-08-09T00:00:00.000Z

An LLM request crosses rate limits, tokenization, scheduling, inference, streaming, and billing before your application sees a response. That fuller model makes 429s, latency drift, and surprising bills easier to debug.

https://aadesh-ingle.github.io/blog/running-two-model-providers-in-production 2025-06-21T00:00:00.000Z 2025-06-21T00:00:00.000Z

We run OpenAI and Gemini across client deployments. Prompts need provider-specific work; the eval cases, contracts, and traces are what let us move safely between them.

https://aadesh-ingle.github.io/blog/own-your-got-damn-failures 2025-03-21T00:00:00.000Z 2025-03-21T00:00:00.000Z

After watching a friend write carefully about a breakup, I started treating personal failures more like postmortems: record what happened before memory makes my role smaller.

https://aadesh-ingle.github.io/blog/why-output-tokens-cost-more 2025-02-15T00:00:00.000Z 2025-02-15T00:00:00.000Z

Output generation is sequential and keeps cached state in GPU memory while a response runs. That mechanism helps explain the price gap and changes how I set output limits, prompt order, and streaming.

https://aadesh-ingle.github.io/blog/principles-i-reread-before-starting-anything-new 2024-08-18T00:00:00.000Z 2024-08-18T00:00:00.000Z

Six notes I reread before a new project or commitment. Each comes from a decision that cost more than writing the rule down would have.

https://aadesh-ingle.github.io/blog/the-unreasonable-effectiveness-of-cleaning-data 2024-04-14T00:00:00.000Z 2024-04-14T00:00:00.000Z

A logistics customer's transporter table had missing identifiers, aliases, and duplicate companies. Cleaning it improved the downstream systems that depended on those records.

https://aadesh-ingle.github.io/blog/bleu-lies 2022-11-12T00:00:00.000Z 2022-11-12T00:00:00.000Z

My image-captioning model scored ROUGE-L 39 and BLEU 10.3 on the same outputs. Reading the captions showed where both numbers were misleading.

https://aadesh-ingle.github.io/blog/search-for-languages-the-internet-forgot 2022-07-09T00:00:00.000Z 2022-07-09T00:00:00.000Z

My Master's work tested dense retrieval where training pairs, benchmarks, and web text were scarce. The hardest part was building evidence that retrieval worked in the target languages.