Jared Lemler

Founder & Lead Engineer

Technologies: TypeScript, Fastify, Next.js, Prisma Postgres, AT Protocol, Cloudflare R2, Fly.io, Vercel, Axiom

• 1 Built and operate a production platform end to end, solo: creator studio, public listener pages, Bluesky OAuth identity, R2 media pipeline, browser-native recording booth (off-main-thread MP3 encoding, IndexedDB crash recovery), and Fly.io/Vercel deploys. Serves 13 creators and 489 published episodes, with ~12K RSS feed pulls/month from Spotify, Apple, and Pocket Casts.
• 2 Used an LLM agent to investigate production incidents: ran it over Axiom logs and Postgres stats (pg_stat_statements, pg_stat_user_tables) to produce a root-cause audit of runaway database spend. It traced the cost to a firehose reconnect storm (~820 reconnects/day), one cursor row taking ~5k writes/day, and scheduled jobs re-running on every restart.
• 3 Cut production DB operations 96% (~209k to ~3-8k/day) based on that audit by adding an in-memory creator-ID cache, lengthening the cursor flush interval, and batching writes into transactions, which removed ~60k table scans/day.
• 4 Built an event-driven ingestion service on a 2,000+ events/sec Bluesky firehose. Server-side filtering across 7 collection types drops ~96% of events before they reach application code or the database.
• 5 Resolved an OAuth sign-in outage: isolated Vercel IP routing as the root cause with cross-network probes, then routed client metadata through a Cloudflare proxy, restoring PDS sign-in from complete failure.
• 6 Automated incident response with an AI Slack bot: a TypeScript service on GitHub Actions checks Axiom health metrics hourly and posts a Claude-written incident summary to Slack. A stall check catches silent failures that never log an error.
• 7 Automated sprint planning, issue tracking, and test generation by connecting agents to Linear and Notion over MCP, using open-weights models.
• 8 Hardened untrusted input paths: protected RSS import against SSRF and malicious uploads with RFC-1918 blocking, magic-number validation, and ffmpeg metadata stripping.

Backend Engineer, Contract

Technologies: TypeScript, React, Vite, Tailwind CSS, Python, FastAPI, Cloudflare, Plausible Analytics

• 1 Built internal analytics dashboards and automated reporting pipelines that turn raw event metrics into business insights, cutting manual reporting effort ~90%.
• 2 Shipped a privacy-compliant web platform (React/Vite/Tailwind, FastAPI backend) serving 200+ active university students.
• 3 Moved fast with the founder: turned Lovable prototypes into production features, like a QR-code trigger linking live events to onboarding.

• 1 Led crews of 2-4 electricians on commercial jobsites, building and inspecting critical infrastructure to blueprint, schedule, and strict code where failure is not an option. Journeyman through ABC Indiana; IBEW member.

NFL Contract Value Pipeline & Ask Tool

Technologies: Python, Dagster+ Serverless, Snowflake, dbt, Postgres, FastAPI, HTMX, Claude API, GitHub Actions, Railway

• 1 Built a production ELT pipeline: Dagster orchestrates nflverse play-by-play, schedules, and OverTheCap contract data (2015 to present) into Snowflake RAW, STAGING, and MARTS layers via dbt; 20 assets with season partitions, refreshed on schedule by Dagster+ Serverless with no dependency on a local machine.
• 2 Stopped bad data before publish: each load deletes, copies, and audits inside one Snowflake transaction and rolls back on failure. Wrote a public postmortem on a race between blocking checks and downstream automation, and fixed it by validating before commit.
• 3 Shipped a serving layer: an audited Dagster asset publishes finished marts to Postgres in a single-transaction swap, verified against Snowflake by row counts and primary-key fingerprints, so the website never queries the warehouse.
• 4 Modeled spend vs. performance: QB cap share against EPA per dropback, with "worth it" defined as EPA above a fitted value line; documented edge cases (mid-season trades, missing cap rows) as tested decisions, 43/43 dbt tests passing in prod.
• 5 Grounded LLM question answering: a Claude Haiku tool loop over select-only, parameterized, read-only queries with timeouts and row caps; code rejects any number not returned by a query, and out-of-scope questions get a "no data" reply. 81-question eval suite scores 78/81 (30/30 refusals) at ~$0.0005 and 4s per question; chose the model over GLM-5.3 Flash on eval accuracy, cost, and latency.
• 6 Ran it like production: live public deploy with rate limits, spend caps, and no stored user data; required CI checks (pytest with mutation-checked tests, dbt parse, gitleaks full-history secret scan); Snowflake key-pair auth, least-privilege role, and a 10-credit monthly resource monitor.

• 1 Agents kept spending tokens re-reading the same remote docs, so I wrote a zero-dependency TypeScript library and CLI that condenses Markdown links into compact markers and uses a two-phase HTTP HEAD/ETag cache to skip redundant downloads.

• 1 A published AT Protocol skill that gives AI coding agents working knowledge of the protocol. I also write about building on ATProto and RSS on pckt.blog.