Hands-on AI engineering and reliability
Data Ceiling Audit
“We have a lot of proprietary data, but we do not know what it can actually support.”
We map coverage, history, grain, labels, outcomes, and joinability, then name the gaps that are actually holding the product back. You get the foundation you can ideate on top of: what is buildable today, and which gap to close first to raise the ceiling.
Data Representation Engineering
“Our data mostly lives in tables and SQL. We think there is more intelligence buried in it.”
Tables and SQL only surface what someone already thought to model. We build the representations that fit the problem, whether that is semantic embeddings, taxonomies, entity graphs, temporal journeys, or behavioural vectors, then map each one to the product or decision it unlocks.
Intelligence Ceiling Assessment
“We have a working model, but we do not know how close it is to the best we could do.”
We test your baseline against the modern approaches that genuinely fit the problem: representation learning, retrieval and ranking, sequence and graph models, foundation models, post-training. You get the techniques worth adopting before your competitors find them, each with achievable lift and what it costs to get there.
Agent Failure Analysis
“It works most of the time and we don't know why it fails.”
We take your production traces, cluster the failures, and hand back a labelled taxonomy with the rate of each. Fixes come ranked by how often it happens and what it costs you, so the backlog is ordered by revenue rather than by whoever complained last.
AI Reliability Audit
“We shipped quickly, and now every model, prompt, or data change feels risky.”
We trace the blast radius of a change through your agentic and LLM pipelines, then instrument what is missing: quality gates, slice coverage, regression detection, and a rollback you have actually rehearsed. You get the confidence to ship daily instead of the caution that costs you weeks.
RL Readiness Check → verifier-graded fine-tune
“We have plateaued. Better prompts and a bigger model stopped helping.”
Once prompting and a bigger model stop paying, the lever left is the signal only you have. We build the verifier from checks your team already runs, set a supervised baseline, then tune only where the evidence supports it. You keep the recipe, the weights, and control of your own know-how.
Founder led
You work directly with the person designing the workflow.
Founder note
I started Performance AI Lab because teams shipping AI deserve systems that hold up under real traffic, not another demo that looks good once and then quietly degrades.
The work draws on 25 years of engineering, including time at Meta Superintelligence Lab, Instagram Ads, PayPal, and Honey, plus projects with two of the largest Chinese tech companies and a high volume logistics carrier.
What that buys you is senior judgment and real infrastructure experience pointed at a business outcome: less manual work, safer data handling, predictable cost, and a return you can measure.
The wider goal is to give startups the kind of AI capability that large companies can already afford, so the gains are not concentrated in a handful of places.