Contact
Book a 30-minute technical diagnosis
Share the business context first. I will help assess whether the AI application is worth building, how to approach it, and where the main risks are.
Telegram @NieErAI Message me on Telegram
Articles
Practical notes on moving AI applications from demo to delivery.
English content starts with concise versions of the core topics and will expand over time.
RSSTurning ComfyUI Into SaaS Is Not About Calling /prompt
A practical engineering note on wrapping ComfyUI with contracts, permissions, credits, DAG workflows, and object-storage safety.
The hardest part of LLM data augmentation is not the model call
Long-running training-data pipelines often fail around pause, resume, and idempotent writes long before the cleaning algorithm becomes the real problem.
In Regulated Settings, A Right Answer Isn't Enough — The Model Has To Show Its Reasons
For securities-regulation content control, getting the verdict right is only the passing bar. The hard requirement is that every verdict carries an auditable reason — which means explainability has to be a training objective, not a bolt-on.
Contract Extraction Is Hardest After the JSON
A practical note on turning financial contract clauses into evaluated rules and transaction calendars, not just extracted JSON.
In Health Check AI, Generation Is The Easy Part. Boundary Control Is Harder.
In health check report pipelines, the real work is not making LLMs generate text but turning recommendation mapping, report style, false-conflict handling, and safety boundaries into system behavior.
Contract Extraction Needs Evaluation, Not Just Prompt Tweaks
Reliable contract extraction requires golden sets, field-level checks, evidence validation, error attribution, and version comparison.
In Investment Research RAG, Query Understanding Is Easier To Underestimate Than Vector Search
In real investment research Q&A, RAG quality depends less on vector search alone and more on coreference, time range, entity recognition, tag filters, and evidence boundaries.
Why Macro Report Generation Should Not Let LLMs Improvise
Financial and macro report generation needs engineered data cadence, traceable sources, explicit disagreement, structure, and refusal behavior.