Skip to content

Terminal

david@ddiscua.dev — zsh

Real input. Try `help`. Nothing is logged, including your typos.

Systems that shipped

Client work. I describe what changed, not how their systems are wired inside, since that part is theirs. Open any row for details. The systems I can draw in full are in the next section.

Houston, we have a problem.

A made-up Friday with a real failure: the autoscaler eats every database connection. It replays on a loop. Nobody got fired.

Rather have this talk before the pager goes off?

Book an architecture review
oncall@prod — SEV1 · payments
  1. oncall@prod: kubectl get pods -n payments
  2. NAME READY STATUS RESTARTS
  3. api-7c9f4d8b6-x2kq9 0/1 CrashLoopBackOff 14
  4. api-7c9f4d8b6-p8mz1 0/1 CrashLoopBackOff 13
  5. … 38 more, all of them unhappy
  6. oncall@prod: kubectl logs deploy/api -n payments --tail=2
  7. FATAL: too many connections for role "api"
  8. FATAL: remaining connection slots are reserved
  9. oncall@prod: kubectl rollout restart deploy/api -n payments
  10. deployment.apps/api restarted
  11. oncall@prod: # 40 pods restarted. 40 pods. Still on fire.
  12. oncall@prod: echo "Houston, we have a problem."
  13. Houston, we have a problem.
  14. oncall@prod: escalate --to architect --sev 1 --reason "it's Friday"
  15. → paging ddiscua … acknowledged in 00:42
  16. → handoff: payments · postgres · autoscaler · 17:58
  17. ddiscua@advisory: kubectl get hpa,pdb -n payments
  18. hpa/api cpu 91% / 70% replicas 40 / 40 (max 40)
  19. pdb No resources found in payments namespace.
  20. ddiscua@advisory: psql -c "show max_connections" -c "select count(*) from pg_stat_activity"
  21. max_connections 500
  22. count 500
  23. ddiscua@advisory: # 40 pods × 25 pool = 1000 wanted, 500 exist. The autoscaler was the DDoS.
  24. ddiscua@advisory: helm upgrade api ./chart --set pgbouncer.enabled=true,pool.size=8,hpa.max=24,pdb.minAvailable=2
  25. Release "api" has been upgraded. Happy Helming!
  26. ddiscua@advisory: watch -n1 'psql -tc "select count(*) from pg_stat_activity"'
  27. 500 → 312 → 188 → 96 → 94
  28. ddiscua@advisory: curl -s https://api.example.com/health | jq -c '{status, p99_ms, errors}'
  29. {"status":"ok","p99_ms":88,"errors":0}
  30. ddiscua@advisory: git commit -am "fix(payments): pgbouncer, bounded HPA, PDB. No mention of Friday."
  31. [main 4f1c2a9] fix(payments): pgbouncer, bounded HPA, PDB. No mention of Friday.
  32. ddiscua@advisory: echo "Houston, the problem was between the autoscaler and the connection pool."
  33. Houston, the problem was between the autoscaler and the connection pool.
  34. ddiscua@advisory: # next time: /advisory, before the pager.

Lab

Stuff I build for myself. No client, no deadline. One of these runs. The other is still a design, mostly because designing agents is more fun than babysitting them.

lab / 01 · status: running

Prediction-market research worker

A Bun and Node worker that records Polymarket BTC markets and Binance prices in the order they arrive, paper-trades complete-set arbitrage, and can replay any past day through the same engine.

Live and replay run the same code. Fills account for fees. If the order book goes stale, a write fails, or there's a file called KILL_SWITCH on disk, it stops. Real trading is blocked in the config on purpose.

TypeScript · Bun · Node 24 · WebSockets · Zod · decimal.js · MongoDB · GCS · Cloud Run

CLOUD RUN WORKER · 1 INSTANCEbookstickseventsrecordre-runauditlatchPolymarket SDKGamma · market WSBinance WSspot · perp ticksWorkersequencer · timersEnginestate · risk · paperTick archivegzip JSONL → GCSReplay CLIsame engine, offlineMongoDBorders · fills · auditKill switchlatch · stale feeds
runtime: Bun for dev · Node 24 in the container · Cloud Run worker pool, one instance · GCS for ticks · Mongo for audit · Pino → Cloud Logging

lab / 02 · status: design · not yet trusted with kubectl

Agent mesh for the homelab

A VoltAgent supervisor that takes tasks from Telegram and hands them to a research agent or an ops agent. Everything goes through MCP, and it asks me before it touches the cluster.

Local models on Ollama first, cloud only when a task needs it. Memory in Mem0 on the homelab Postgres, so it remembers last week's incident. Every run is traced in VoltOps so I can replay it. It's still a design. The ops agent doesn't have kubectl yet.

VoltAgent · MCP · Ollama · Mem0 · pgvector · Telegram Bot API · Home Assistant · VoltOps

HOMELABtasksdelegatetoolsinferencerecalltracesInboxTelegram · approvalsSupervisorVoltAgent · plannerResearch agentweb · RAG · summarizeOps agentkubectl · tf planMCP serversGitHub · HASS · filesModel routerOllama first · cloudMem0 · pgvectorepisodic · factsVoltOpstraces · evals
runtime: VoltAgent supervisor + sub-agents · Ollama on a homelab GPU, cloud fallback · Mem0 over Postgres/pgvector · MCP for every tool · VoltOps traces

What I work in

Six areas I work in, and the tools I actually use in each. Type cat stack.json in the terminal for the full list.

  • Software architecture

    Clear module boundaries and contracts, and monorepos that don't turn into a mess.

    Nx · Zod · GraphQL · TypeORM · Temporal · Storybook

  • Full stack engineering

    TypeScript front to back: React, Angular, NestJS, GraphQL.

    TypeScript · React · Next.js · Angular · NestJS · Node.js

  • Cloud & DevOps

    Azure, AWS and GCP, Terraform and Pulumi, Kubernetes, and CI/CD that people actually use.

    AWS · Azure · GCP · Terraform · Pulumi · Kubernetes · GitHub Actions

  • AI engineering

    LLM products, agents and RAG, with evals so you know when it gets worse.

    OpenAI · Azure OpenAI · RAG · pgvector · Agents · Mem0

  • FinTech & payments

    Wallets and payment APIs next to old banking cores.

    Node.js · Java · SAP · Jenkins · PostgreSQL · Pandas

  • Security & compliance

    RBAC, secrets, scanning and audit trails, mostly in healthcare and finance.

    Vault · RBAC · SSO/OTP · Tenable · HL7 · Datadog

Off the clock, still on a terminal

I run a homelab. On any given day it's either working fine or half torn apart because something seemed like a good idea at 1 a.m. Linux on the desk, terminals everywhere, and I don't feel like I understand a tool until I've self-hosted it.

I open up hardware. I build agents for fun, and a few of them helped write this page. I read more open source than I star. The epilepsy research years also left me with a soft spot for a good dataset.

now.md1 section
building
ddiscua.dev, with agents doing the typing and a human doing the blaming

Research

The genetic epilepsy platform helped doctors and researchers work together, and the work got published. I still have a soft spot for a good dataset.

  • IEEE CONCAPAN

    A paper from the genetic epilepsy research platform.

  • ILAE / AAN abstracts

    Conference abstracts from the clinical and genetic research.

If the system has to ship, talk.

Message me on LinkedIn and tell me what you're building.

Connect on LinkedInGitHub

How I work

  • Architecture first

    I agree on contracts and boundaries before anyone writes a feature.

  • Ship in slices

    Vague tickets get cut into small pieces, each with tests, a migration and a way back.

  • Automate the boring

    Pipelines, infrastructure as code, agents. People should spend their time on decisions.

  • Lead without theater

    Code reviews, standards, and unblocking people. The client's goals come first.