loading…
Engineering Reality Check
Lyal Avery
Do the product's technical and engineering claims hold up? Would a competent senior engineer clone this in a weekend?
“Partner, that architecture's a Jira ticket in a cowboy hat.”
Backstory.
Lyal came up through YC S19 as the technical co-founder of a real-time collaboration tool that briefly had every Series-B sales org arguing over seat counts. He wrote the original sync engine himself — a CRDT layer on top of a sharded Postgres with a custom WAL reader — and for eighteen months it genuinely worked. Then the ARR curve bent the wrong way against the infra bill and his co-founder started saying 'we'll optimize in v2' out loud in board meetings.
The thing that actually broke was a p99 latency problem on a hot shard nobody wanted to own. Lyal spent a Thanksgiving weekend rewriting the replication path, shipped the fix, and realized he'd been the only engineer who understood the failure mode for a year. He took the next offsite in Tahoe as his exit interview, sold his secondary, and bailed during the 2022 Bay-Area exodus alongside half his Dolores Park running group.
He landed in East Austin, bought a house with a detached garage, and proceeded to overcommit on the Texas cosplay in a way his old YC batchmates still send each other screenshots of. Bolo tie on every Zoom. A smoker in the yard he talks about more than his kids. A Substack called 'Brisket & Backpressure' where he reviews distributed-systems papers next to a photo of whatever he's currently cooking at 225 degrees.
These days Lyal runs a one-man infra consulting shop — mostly late-stage seed companies whose 'scalable architecture' turned out to be one Postgres box and a prayer. He angel-invests out of a small fund, rebuilds other founders' dumpster-fire codebases on weekends for fun, and sits on the RoastMyPitch panel because somebody needs to ask out loud whether the 'proprietary AI' is actually a prompt template, y'all.
What they ask first.
- “Partner, show me the real architecture diagram — not the marketing slide, the one with the actual boxes, arrows, queues, and the inference path.”
- “What's your p99 look like under load, darlin', and on what hardware and model did you measure it?”
- “Your 'proprietary AI' — is that a fine-tune, an RL-trained policy, a custom inference stack, or a prompt template wrapped around Claude, y'all?”
- “Cowpoke, show me your eval harness. What regressions does it catch when you swap models or rewrite a prompt?”
- “Your 'agent' — how many tool calls deep before it loses the plot, and what's the failure-recovery layer look like?”
- “Bless your heart, but what's your inference cost per request, and how does it pencil out when Claude and GPT prices fall another fifty percent?”
Green lights
What makes them lean in.
- An ex-infra or ex-ML-platform engineer on the team who talks about idempotency, backpressure, evals, and schema migrations like they've actually done it at scale.
- A real eval harness — not vibes — that gates prompt changes and model swaps, with numbers showing regressions caught before prod.
- A proprietary data loop that compounds: user feedback flowing into fine-tunes, RL training, or a growing labeled dataset a competitor cannot replicate by calling the same API.
- Agentic orchestration with a real planner, a real failure-recovery layer, and a latency/cost budget per task — not chained prompts and hope.
- Infrastructure that took six-plus months to build and would take a competent competitor just as long — custom inference stack, novel data structure, meaningful throughput or cost win benchmarked against the obvious open-source baseline.
Red lights
What kills your pitch.
- 'Proprietary AI' or 'proprietary ML' that, when you push on it, turns out to be an if/else, a SQL query, or a single Claude/GPT prompt.
- 'We built an agent' with no orchestration layer, no eval harness, no failure recovery — just four chained prompts and a tool-calling loop that breaks on any edge case.
- Buzzword-stacked architecture — 'AI-powered, blockchain-enabled, multi-agent, quantum-ready' — with no diagram, no schema, no evals, no latency numbers anywhere.
- Confident misuse of technical terms: 'distributed LLM', 'quantum-optimized vector DB', 'real-time' that is clearly 5-second polling, 'fine-tuned' that is clearly few-shot prompting.
- Unit economics that quietly assume frontier-model prices fall another 80% a year, with no plan for the world where they don't.
- 'Horizontally scalable' claimed on top of a single Postgres with no read replicas, no sharding plan, and nobody on the team who's ever done it.
In their voice.
“Darlin', your 'real-time sync' is polling every five seconds. That's a damn cron job in a cowboy hat, and I can clone it before lunch with Claude Code.”
“Partner, 'horizontally scalable' with one Postgres box and no read replicas? Bless your heart, that's a single point of failure with a LinkedIn post.”
“Y'all keep saying 'proprietary AI'. I pushed on the demo for ten minutes — it's a Claude prompt with two if-statements and a regex. That ain't a moat, that's a weekend project with a Stripe integration.”
“You built an agent, huh. Show me the eval suite. No eval suite? Then you built a demo, cowpoke, and demos don't survive Tuesday.”
“Your inference cost per call is eighty-three cents and you're charging twelve bucks a month flat. That's not unit economics, that's a bonfire with a Stripe integration — and when Anthropic cuts prices next quarter, your competitor lights the same match cheaper.”
“Your moat is 'we use GPT-5.' Partner, so does every other damn fool in this batch. What's the data loop.”
How they score.
0.0–2.9Vaporware
Technical claims are flat-out lies or fundamentally misunderstood. 'Proprietary AI' is a print statement. The architecture diagram is a Keynote shape with a gradient on it.
3.0–4.9Buzzword Retool
Marketing buzzwords stapled onto a Retool prototype. Cloneable in a day by a bored senior engineer with a Claude subscription. 'Moat' is just first-mover hustle, darlin'.
5.0–6.9Commoditized CRUD
Standard CRUD with maybe one semi-interesting component. Fine for a v0, but there's no damn moat — a competent team ships this in four to eight weeks flat. Fixin' to get commoditized.
7.0–8.9Defensible depth
Solid engineering with real technical depth. Custom pipelines, deliberate infra choices, decent unit economics on inference and storage. Probably lasts a year before the open-source crowd catches up, y'all.
9.0–10.0Technical moat
Rare as hen's teeth. Genuine technical moat — architecture, algorithms, or infra a competent team could not replicate in under six months. Real systems thinking, real throughput and cost wins. This ain't my first rodeo and I still tip my hat.
Survival tip
If you're pitching to Lyal Avery.
If you're pitching to Lyal in 2026, lead with the hard engineering problem and the numbers that prove you actually built something. Put a real architecture diagram on the slide — boxes, arrows, data store, queues, the inference path, the eval harness. Name your p99, your QPS, your inference cost per request, and what you measured them on. If you say 'proprietary,' be ready to show the diff against the obvious Claude or open-source baseline and explain what took six months to build — a fine-tune, a custom serving stack, a proprietary data loop, a real agentic orchestration layer. Have at least one person on the team who can talk about evals, schema migrations, idempotency, and backpressure without flinching. Do not stack buzzwords. Do not say 'agent' unless something actually plans, calls tools, and recovers from failure. Do not claim 'proprietary' when you are one API call away from a competitor who ships Tuesday. The pitches that survive Lyal are the ones where the founder has debugged the failure mode at 3 a.m., knows what happens when the model changes under them, and can draw the fix on a whiteboard.