Product leader bridging Big Tech PM rigor with zero-to-one AI engineering.

Live product — Ask Sukul AI

The chat on this site is the artifact: a Claude-powered assistant grounded in Sukul's own career notes and project write-ups, gated behind a lead-capture step after the first exchange, and built to say it doesn't know rather than hallucinate when a question falls outside what it's grounded in.

The highest-signal artifact on this page — you can test the actual product in the corner of this screen right now instead of taking a case study's word for it.

Eval suite — Ask Sukul AI

A documented eval board for Ask Sukul AI's answers — roughly 30 happy-path questions about Sukul's background, 10 edge cases (ambiguous or out-of-scope asks), and 10 adversarial prompts probing for hallucination or persona breaks, scored on factuality, refusal-appropriateness, and tone. Published as a spreadsheet/JSONL with a before/after comparison across system-prompt revisions.

Repeatedly called the single most under-prepared AI-PM skill — this proves it against a product you can go interrogate yourself, not a toy example.

AI-native PRD — the trading agent

A one-feature PRD for the trading agent's LLM level-setting step: model-choice rationale under real latency and cost constraints, the structured prompt that turns indicator state into entry/target/stop levels, eval criteria for a system where being wrong is expensive, a cost model for LLM calls per trading day, failure modes like hallucinated levels or bad fills with detection and response, and the risk-gate logic that keeps hard limits in the loop.

Most PRDs stop at the feature — this one shows the underlying system got designed, for a case where skipping that design costs real money.

Cost & latency model — KonvoLead

A spreadsheet modeling KonvoLead's real cost and latency envelope: tokens per WhatsApp turn across retrieval and generation, cost per qualified lead over a full multi-day conversation, p50/p95 response latency against WhatsApp's UX expectations, and monthly cost at 100/1k/10k active leads with caching impact on repeat-context RAG calls.

Most AI PMs can't do this confidently for a real multi-turn, tool-using system — it's a concrete, rare differentiator.

Model comparison — KonvoLead

A real multi-turn KonvoLead conversation — RAG lookups plus a scheduling tool call — run across Claude, GPT, and Gemini, compared on retrieval faithfulness, tool-calling reliability, multi-day state tracking, and cost per conversation, with a clear recommendation.

Shows the judgment to pick the right model for a production agent, not just familiarity with one.

Visuals & creatives — LitPartyShit

A small gallery of marketing visuals for LitPartyShit made with AI image and video-generation tools — real creative-pipeline output, shown next to the prompts and iterations behind them, not a mockup.

Generative creative that has to convert paid traffic is a different bar than a demo that only has to look plausible — this is that bar, with the receipts.