Building AI agents is easy. Building one you can trust with a real business workflow is harder.

I built a Freelance Admin Agent to explore that problem using Strands Agents SDK and Claude Sonnet 4.5 on Amazon Bedrock.

The working MVP takes manually provided client emails and turns them into a structured payment workflow:

→ Classifies invoice and payment emails → Extracts client, invoice, amount, currency, dates and contextual details using structured output → Stores and matches records in SQLite → Uses deterministic logic to identify full/short and on-time/late payments → Uses a second agent to draft a contextual follow-up only when needed → Keeps the final decision with the human through Approve / Edit / Reject

One of the most valuable parts wasn't adding another AI feature. It was finding a real reliability issue during testing: reprocessing the same email could duplicate records and inflate payment totals.

I reproduced it through repeat testing, traced the database behavior, and fixed it.

That reinforced something I'm increasingly applying to my AI engineering work:

A useful AI product isn't just an LLM call. It needs clear boundaries between AI reasoning, structured data, deterministic business logic, persistence, testing, and human control.

The next iteration will focus on stronger production safeguards, including source-grounded financial extraction, precise currency handling, stricter trust boundaries for untrusted input, and eventually live inbox integrations.

I'm interested in building more systems like this for startups and teams with repetitive workflows that can genuinely benefit from AI agents and automation.

Stack: Python · Strands Agents SDK · Amazon Bedrock · Claude Sonnet 4.5 · Pydantic · SQLite · Streamlit

turning client emails into admin tasks is such a real freelancer pain, nice build 🔥

AI Code Performance Engineer is a developer tool I built to analyze codebases, identify performance bottlenecks, generate optimization candidates, verify correctness, and benchmark whether changes actually improve performance. It combines software performance engineering, automation, testing, CI/CD, and AI-assisted optimization into a single workflow focused on measurable results rather than speculative suggestions.

I'd run the baseline and AI-generated candidate on fixed inputs in the same benchmark job, then compare repeated timings and memory use after the correctness check. How are you filtering out CI runner noise before accepting a speedup?