Problem. This is the one I’m most confident about — your careers page has an “AI Context & Harness Engineer” plus a matching PM role open, so I’m guessing it’s already on the list. From outside: a non-technical user chats an app into existence, closes the tab, comes back next week. The agent has forgotten the stack, the design decisions, the data model, the half-finished TODOs. They re-explain from scratch, and casual users churn.

Build. Per-user × per-workspace memory layer. Writes: stack preferences, design decisions, schema shape, open TODOs, last-deploy status. Reads: a compressed summary injected into every new turn’s system prompt. Scored by recall@5 on a held-out eval set so quality regresses whenever the agent prompt changes.

I've done this: this is the exact layer I lead at SalesMonk — memory layer on mem0 + Qdrant, agent evals via Braintrust.

Sketch. Rough timing if the codebase doesn’t surprise me: v0 in a week — mem0 instance per workspace + a Braintrust eval deck scoring retrieval on scripted multi-session builds. v1 in a month — auto-summarise old sessions into a compact memory graph so context windows stay short as projects age.

Problem. Less certain this is a live pain point — I don’t know your current failure rates — but if it is: “building your app” looks like a 2–5 minute pipeline (plan → scaffold → migrations → auth → routes → components → deploy). One provider blip or closed tab and the user sees a half-built app or a white screen. On a generator platform that’s a brutal failure mode.

Build. Wrap the pipeline in Temporal (or stay on Inngest if you want pure serverless). Twelve-ish idempotent activities, automatic retry with cost-aware backoff on LLM hiccups, resumable from the last successful step if the user closes the tab. Every build gets a run ID support can replay — which I’d actually find useful for diagnosing regressions regardless of whether retries become a user-visible feature.

I've done this: I architected SalesMonk's durable workflow engine, and the 100k+/mo Inngest scoring pipeline next to it — zero data loss across worker restarts.

Sketch. Activities roughly: plan, scaffold, db-gen, auth-gen, route-gen, component-gen, wire, smoke-test, deploy, domain, onboard, index. Compensating action per step for clean rollback. Target I’d aim for: <0.3% of builds hit a fatal retry ceiling, though that number is mostly a gut call until I see your actuals.

Problem. Speculative — this is a surface I’d want to explore, not something I’m sure is a priority. Most SaaS + ecom apps Hercules generates will end up needing support and sales chat; today the user wires Intercom, Crisp, or builds one themselves. That’s usage time leaking and a revenue surface Hercules doesn’t own. Could be worth picking up, could be a “later” — you’d know better.

Build. A native Chat component shipped with every generated app. Customisable colors + persona, routes to the app owner’s inbox (Slack / email / Hercules inbox), RAG’d against the app’s own docs + transaction history. Opt-in AI agent handles tier-1 and escalates on intent.

I've done this: I built SalesMonk's chat widget (2M+ visitors/day, +~50% meetings), and the Slack-escalation RAG layer behind it — two years in production.

Sketch. React embed + Pusher for realtime + pgvector on the app’s own DB (or a managed Hercules-hosted tier). Could monetise cleanly — included on paid plans, 2–5% take-rate on AI-resolved tickets — but that’s napkin math, not a pitch.