Work with me

Build LLM systems and agents that hold up in production.

I take a small number of advisory engagements each quarter. Senior-IC depth on the technical questions, not deck-driven advisory.

Who I am

Hamidreza Saghir, Principal Applied Scientist at Microsoft. Previously Senior MLE Lead at X (Twitter), Applied Scientist at Amazon, ML Researcher at Borealis AI. PhD, U. Toronto. Publications at ACL, InterSpeech, WWW, IEEE TASLP.

I open-sourced Looplet, an iterator-first agent framework that exposes the loop instead of hiding it behind agent.run(task). I write opinion essays on agent design, evals, and the parts of LLM systems that don't show up in tutorials.

What I work on

• Agent architectures. Loop design, tool dispatch, recovery, observability. When to use a framework vs. own the loop. The territory I explore in Looplet and write about in The loop is the product and Your coding agent is under-specified.
• LLM evaluation systems. The pipeline most teams underbuild. Designing eval suites that actually predict production quality, judge-model scaffolding, regression tracking, calibration.
• Post-training stacks. SFT data curation, DPO vs. RLHF decisions, knowledge retention vs. instruction following. Why fine-tuning often degrades the base model and what to do about it.
• RAG and retrieval design. Two-tower vs. cross-encoder tradeoffs, chunking, hybrid retrieval, re-ranking. When the right answer is "smaller model + better retrieval," not more compute.
• Inference cost models. Reading actual serving traces, prefill vs. decode tradeoffs, KV-cache strategy, quantization. Identifying the 10x speedups hiding in your stack.

Engagement shapes

• Technical advisory (typical). 2–4 hour deep-dive sessions, async followup over a defined window. For teams with a specific decision in front of them: "should we fine-tune or RAG", "is our eval set actually measuring what we think", "why does our agent fail at scale".
• Architecture review. Read the code, the eval, the recent incidents. Produce a written assessment with prioritized actions.
• Hands-on prototype. For very specific problems where a prototype unblocks a decision. Limited availability.

Who I work with

• Engineering teams at Series B to growth-stage companies shipping LLM products.
• Established companies adding LLM features to existing products.
• Founders/CTOs who need a senior technical sounding board for specific decisions.

Who I don't work with

• Pre-product teams looking for general "AI strategy" advice.
• Anyone wanting full-time or fractional CTO arrangements.
• Projects that compete directly with Microsoft.

How to reach me

Email saghir.hr [at] gmail [dot] com with:

• One paragraph on what you're building.
• The specific decision or problem you're stuck on.
• Your team size and current stack (model, infra, eval setup if any).
• Timeline and rough engagement shape you have in mind.

I reply within a few business days with whether it's a good fit and how we'd start. First call is 30 minutes, no charge, no slide deck.