Peter Vijeh

I make AI coding agents reliable inside real production codebases.

I build the infrastructure that makes AI coding agents usable in production: repo-scoped agent skills, review and test gates for agent-written code, and the adoption metrics that show whether any of it worked. Day job is platform tech lead at a fintech in New York, where I also own identity and CI.

Same problem I've worked on for ten years, new shape: making the fast path the safe path in systems where being wrong is expensive.

start here

I scanned 1,074 companies for a SaaS worth cloning. The best answer was “don’t.”

The premise everyone repeats right now is that AI made legacy vertical SaaS clonable by one person. I tried to name the company to clone, out of the largest pile of aging vertical software anyone owns, and the thesis died at the pricing page: rebuilding was never the hard part.

1,074

companies scanned

≤3%

publish a price

clear the screen

Read the write-up → 11 min

write-ups

📉

AI is supposed to be deflationary. Here's what the macro data suggests.10 min read — Is AI obviously deflationary? Zillow, FAO, EIA, 32 software 10-Ks and five status pages, no CPI. Rent, food, power and streaming cost more than in 2022, software margins rose from 29% to 34% and cloud incidents doubled. Token prices fell from $60 to $10 per million.

🕵️

Does Reddit have an astroturfing problem? What the data suggests11 min read — Are Reddit knife recommendations astroturfed? In 51,129 comments from six knife subreddits, the 5% most brand-heavy accounts wrote 11.3% of buying-thread brand mentions against 7.9% expected, and 31% against 8% for one brand. Their histories match a comparison group: 4.5 years old, 66 subreddits.

🧠

I had Gemini train its own replacement for $96 min read — A 459M-parameter model that tags brands, models and materials in Reddit comments about knives. ~0.65 (est.) to 0.83 F1 for $9 of Gemini labels and 24 minutes on a T4. Covers the labeling pipeline and the mask bug that wasted five training runs.

🏘️

Gentrification signals are a lagging indicator of home prices, not a leading one4 min read — Can public data pick the New York ZIPs that will outrun the metro? Fort Greene's coffee shops arrived a decade after its prices moved, the 2005-2015 winners lost 2015-2025, and a model on 26 years of Zillow data beats sort-by-last-year at one year only, +0.44 to +0.39.

🗽

Is Manhattan pricing out the young? What 14 years of data show11 min read — Do fewer young people move to New York than in 2011? Census estimates, IRS tax-return flows and 14 years of ACS microdata, for people aged 20 to 29, by income and neighborhood. Total arrivals are flat; arrivals earning under $50k are down 29% in Manhattan.

tools & demos

🏚️

NYC building-compliance market screen — Every permit-expediting and violation-resolution firm in NYC, recovered from public data: 711,643 attributed filings resolved to 3,742 firms, with repeat-client and revenue-concentration metrics. Top five firms hold ~11% of the market.

🤖

Can LLMs predict steel properties? — A benchmark that gives 7 LLMs nothing but a steel's chemistry and asks them to rank it. Graded against CATRA and Charpy lab data. Leaderboard and methods.

🔬

Steel property predictor — ML models that score edge retention, toughness, corrosion and sharpening for 134 steels from composition alone. Trained on CATRA wear tests, with an interactive tool.

history

2023 -

Platform tech lead at a fintech in New York. Lead AI enablement for the engineering org: reusable agent skills, quality gates for agent-written code, automated cleanup of engineering debt, and the adoption metrics behind it. Also own authentication and SSO, CI and build infrastructure, and developer experience — including a company-wide move to standards-based auth (OIDC + PKCE, passkeys) across every login surface.

2022 - 2023

Lead frontend engineer at Coast. Owned the architecture and roadmap of the cardholder portal as it grew from fuel cards into a full fleet and expense platform.

2019 - 2022

Tech lead at Better.com. Designed and shipped the real-time platform connecting borrowers with loan officers — zero to production with a team of four, and a measurable lift in loan conversion.

2018 - 2019

Senior engineer at Capsule, building the pharmacy's customer-facing product.

2016 - 2018

Senior engineer on web platforms at R/GA. Nike.com's ecommerce stack and its migration to React; a new customer-onboarding SPA for Verizon.

earlier

B.A. in econometrics and applied math at Boston University. Optimization, time series, causal inference — which is why most of my side projects end up being measurement problems.

misc

Based in New York.
Interested in identity, CI, agent infrastructure, and small businesses that run on public data.
Best way to reach me is the form at the bottom, or the email icon up top.

contact